Workflow Library
- Distinct
- 15
- Found
- 231
- Runs on APIs
- 118
- Sources
- 7
Video11
Real-frame person video21A real-looking person on camera: UGC, podcast clip, street interview, testimonial, founder talk (format presets)
Seedance 2.5Sora 2Veo 3.1The operator gives: a reference video (or a screenshot), a face photo, a product photo, a voice sample, the script or product facts
- 1Pick a real video whose look you want and take a screenshot of the first frame
- 2Swap only the person (and product) in that screenshot with GPT-Image; keep framing, light and background
- 3Write the script in the format preset (UGC, podcast, street interview, testimonial)
- 4Generate the clip with the start frame, face photo, product photo and voice sample as references
- 5Longer ads: extend from the last frame and join
- 6Finish: captions, upscale
Folded in 21
- Seedance 2.0/2.5 UGC ad video (structured prompt system)BaseThe AI Content SystemThe course's flagship 'one prompt engineer, five format templates' system: a single 6-block prompt structure reused across all 5 video-type templates, each with its own trigger words.
- Research-to-render UGC ad factoryBaseAIOpsThe only full research-to-publish pipeline found: 'byte-identical' reference files are reused unchanged across every generation call, a rule that 'faces anchor, products drift', and a machine QA gate that is deliberately cut-blind and only flags 'broken' rather than judging 'good'.
- Seedance 2.0 — Cockpit field-template ('Michael's method')BaseOursField-by-field structure ('Michael's method') with up to 9 tagged @image references and a literal QC-negatives block; this is the doctrine the live creative-pipeline cockpit's consumer/*.mjs Seedance code actually follows.
- Realistic home-video film scene with a talking personBaseOutside postsClaude Opus 5.5 writes every downstream prompt from one idea plus style screenshots, explicitly told WHO holds the camera, WHERE they sit and HOW they move. A 'camera lock' block forces one continuous take with zero cuts inside each 10s clip. Scene 2 is a single Seedance call producing 6 cuts timed to an invented surf-rock track's drum hits — one prompt making 6 music-synced cuts is new in this range.
- UGC video (Claude Skill + Seedance)AI Ads (Ad Creators Lab)Claude Skill auto-segments the script and tags reference images with @image labels; no native multi-clip continuity so the last frame is extracted between segments to keep the character consistent.
- AI podcast ad (two-speaker)AI Ads (Ad Creators Lab)Same Pinterest+Max Fusion+Seedance/Kling method as index 0, run manually without the Claude Skill front-end; flips/edits the base image to generate the second speaker.
- Research-doc-to-script Arcads hookAI Ads (Ad Creators Lab)Names 4 specific scripting frameworks (Industry Contrarian, Bandwagon Effect, Listicle, How-You-Connect(s)-Without-Why) taught to ChatGPT before scripting.
- Seedance 2.0 facial-recognition-filter bypass (black-circle method)The AI Content SystemMasks the body image's face with a black circle in Canva and attaches a separate face-cutout PNG so Seedance's facial-recognition filter doesn't block the generation; a bolt-on technique to index 25, not a standalone recipe. Ethical concern flagged: uses a real person's likeness without confirmed rights.
- Sora 2 podcast-style ad clip (purpose-built GPT prompter)The AI Content SystemShortcuts index 30's two-model Gemini-then-Claude pipeline with one purpose-built GPT trained on the same underlying prompt structure.
- Gemini Omni AI UGC video (start-frame + voice-reference method)The AI Content SystemSmuggles a real voice into Omni as a trimmed ~14s black-screen mp4 (Omni can't import MP3 directly and a visible reference face would get copied), and writes the prompt as a screenshot of a Google Doc to route around Omni's 'unable to edit speech' refusal bug. Simon's own 2026-09-24 guidance: Omni doesn't take downloaded voice samples the way Seedance does and is reserved for animation; a 2026-07-03 internal test found the API dubs supplied audio but won't make a recorded voice say new words.
- Automated end-to-end AI UGC ad factory (Claude Code skill + MCP)The AI Content SystemFully unattended Claude Code skill that installs/runs a local captions server (nicolaigaina/ai-video-captions, no Docker) and an FFmpeg pipeline with literal captured commands; a QC prompt checks warped hands/garbled text/lipsync drift/wrong words and auto-regenerates failed takes up to 3x per clip.
- Multi-reference video clip with character consistencyAIOpsReal human photos are banned as character references since Feb 2026 (illustrated/3D only); re-injects the character reference at the start of every scene specifically to counter drift; native audio includes phoneme-level lip sync in 8+ languages.
- Full UGC ad pipeline: character image to Seedance 2.0 videoFree onlineRuns as full Playwright browser automation against the Higgsfield UI (not an API), state-handed between image and video stages via an @image1 reference tag; 15 independent large style sub-skills each carry a '2-second hook framework' with 5 worked examples (75 total) plus a deliberate anti-uncanny-valley word list.
- UGC talking/product video ad from a single prompt chainFree onlineGives verbatim fill-in-the-blank prompt templates at all three stages (image, motion, script) and deliberately engineers imperfection at each ('slight camera shake', 'unposed', 'casual and authentic') as the core authenticity mechanism rather than a byproduct.
- Start-frame person swap (i2i, for i2v handoff)OursPrompt is a strict CHANGE/PRESERVE split ('CHANGE — only this: ${change} / PRESERVE — everything else exactly as in the reference: framing, crop, composition, camera angle...') so only the on-camera person swaps while framing/lighting/background are locked; feeds the (currently dead-code) video step (entry 101) and the manual Flow hand-off (entry 100).
- Single-shot text-to-video (Kie.ai Seedance) — DEAD CODEOursDead code with a real routing bug flagged explicitly to Simon in the source notes: it hardcodes Kie.ai regardless of which engine the operator's picker selects, and even if it were wired to a UI button that bug would still fire.
- Seedance 2.0 — Prompt Formula LibraryOursOne universal formula ([Subject]+[Action]+[Scene]+[Lighting]+[Camera]+[Style]+[Quality]+[Constraints]) reused with per-content-type fill-ins across 9 named formats, plus a 00-05s/05-10s/10-15s multi-beat template. Flagged as the preferred model for restricted Icarus clients (cannabis/peptide/vape) since it is less censored than RunwayML/Veo/OpenAI.
- Seedance 2.0 — Standalone prompting skillOurs6-layer core formula plus a 7-element cinematic Image/Video Prompt split, a ban-word list with replacements, a mandatory literal quality suffix ('sharp clarity, natural colors, stable picture, no blur, no ghosting, no flickering'), and a 3-angle character-consistency sheet protocol with a literal preservation-keywords block.
- Seedance longform UGC production pipeline (30s+)OursMulti-Angle Character Reference System (9+ persona images stored under 02-Brands/vero/creative/avatars/<name>/); a mandatory imperfection_spec field per shot; a model-selection flowchart routes each shot to Seedance or the far cheaper Gemini Omni ($0.12/10s vs $0.90-$10/10s) by cost.
- Meta 'Native & Stealth' Format Library (4 sub-formats)OursOne persona setup switched across 4 named sub-formats purely by changing the tool chain, scene and camera feel — the exact format-preset principle described in WORKFLOW-ANATOMY.md; no filled worked example is given per format though.
- Medical/Professional Persona UGC (Pharmacist/Doctor authority)OursGenerated manually per script line via Veo 3 Frames-to-Video with the literal formula 'Make him say: [line]. No background music.'; a physical-action variant of the same formula is also given. Full automation is flagged 'Coming Soon' in the source — current state is manual, line-by-line.
Clone a winning ad14A new ad that copies the structure of an ad that already works, for our brand
Seedance 2.5KlingVeo 3The operator gives: a winning ad (link or file), product photos, brand site
- 1Find winning ads (ad library, TikTok, Instagram) and score them
- 2Break the chosen ad down shot by shot: subject, setting, action, light, camera, words
- 3Rewrite every shot for our product, never borrowing its product or category
- 4Make a start frame per shot with our product and person
- 5Generate each shot, then stitch with music
Folded in 14
- GPT-6 Astra ad studio (brand research + reference breakdown + Seedance 2.5 generation)BaseThe AI Content SystemSame breakdown-then-generate shape as 40/41 (same Drive-SOP-family skill, 'Miko Video Ad Stack'), run through GPT-6 Astra/Codex, which itself queues and assembles the Higgsfield/Seedance jobs; combines FFmpeg frame-change inspection with Higgsfield's video analyzer for the breakdown.
- Reference clip breakdown + brand-adapted scriptBaseOursExplicit 'brand context is source of truth, never borrow the reference's product category' rule stops the analysis step from leaking the wrong product category into the adapted script; live, Analyze & Script screen.
- Antigravity creative cloner (agentic, scripted)BaseRoboNuggetsPackaged as a standalone scripted Python-skill agent (no n8n) with explicit cost-approval gates before every paid image/video call, and a global plain-English/cost-transparency agent rule captured verbatim; the most fully runnable entry in the whole range (exact endpoints/payloads given).
- Winning-ad concept script adaptationAI Ads (Ad Creators Lab)Text-only script/hook output feeding other UGC-ad workflows' script step; no family in the fixed enum covers a script-only asset, family assigned here as the video type it typically feeds.
- Sora 2 AI UGC ad video (deconstruct-then-generate)The AI Content SystemDeliberate dual-LLM pipeline: Gemini analyzes the reference video, Claude refines the resulting prompt, rather than one model doing both.
- Grok Bot research-generate-launch UGC ad systemThe AI Content SystemUses the same Google-Drive-hosted 'Seedance 2.5 video breakdown SOP file' as indices 41/42; orchestrated entirely inside a Grok Bot with a Higgsfield connector, ending in an automated Meta launch step.
- Claude + Higgsfield MCP reference-breakdown-then-generate (Seedance 2.5)The AI Content SystemNear-identical breakdown-then-generate shape to index 40 (same Drive SOP link), run through Claude/Higgsfield MCP instead of Grok Bot; deliberately generates the first pass with NO character reference image, claiming this gives the most realistic characters, then extracts a frame from that result as the reusable reference.
- Viral marketing ad video via VEO3 (JSON prompt template + GPT-4.1 rewrite)Free onlineMaintains a library of 12 JSON prompt templates each reverse-engineered from, and traceable to, a specific real viral X/Twitter ad post; an LLM is instructed to 'edit don't rewrite' — adapt only the fields needed for a new idea while preserving the proven structure.
- Reference clip sourcing (Source & Swipe)OursFacebook Ad Library sourcing is deliberately left unbuilt (a scoped-out gap, not a bug) — sourcing is scraper-plus-score-plus-human-curation only, no generation step; live, Source & Swipe screen.
- Sora 2 Two-Step Prompt Pipeline (Gemini deconstruct → Claude refine)OursTwo-step prompt pipeline: a Gemini Gem deconstructs a real reference video into a bracketed 5000-char-capped UGC prompt template, then a separate Claude Project (rules-configured) refines it into scene/cinematography/action-beats/dialogue/sound.
- UGC Video Analyzer Gem (video → reusable Sora 2 prompt)OursTrigger prompt: 'Analyze this video in depth and make it into a Sora 2 prompt using all the knowledge you have.' The Gem's core instructions (Master Prompt + Knowledge File) are missing from the harvested files, so this is not independently runnable.
- Reverse-Engineering Ads with Gemini (script extraction/rewrite)OursTwo distinct literal Gemini prompts — one for a strategic breakdown, one for script extraction and rewrite — both operating directly on an uploaded reference ad video.
- Veo3 branded ad video (viral-prompt remix)RoboNuggetsRewrites a viral '$100k ad' Veo3 prompt (sourced from a Notion library, written in YAML/JSON for higher output quality) for a new brand rather than writing from scratch; the agent has a think tool and access to prior sheet entries specifically to avoid duplicate ideas.
- Winning-ad repurposer / creative cloner (n8n)RoboNuggetsReverse-engineers a winning/viral reference ad shot-by-shot via the SEALCaM framework (Subject/Environment/Action/Lighting/Camera/Metatokens) before generating a swapped-brand clone -- copying an ad's exact structure, not just its vibe.
Quick UGC batch from one product photo9Many short UGC-style ads with no reference video
Veo 3.1SeedanceSora 2WanThe operator gives: a product photo, product facts, how many ads
- 1Give one product photo and the product facts
- 2An agent writes N different UGC scenes (who, where, what they say)
- 3One generated start frame per scene with the product in hand
- 4A clip per scene with sound made in the same call
- 5Keep the best, optionally stitch into longer ads
Folded in 9
- Infinite UGC video agent (Sheet-driven, multi-model)BaseRoboNuggetsFlagship/most complete UGC workflow in the RoboNuggets batch — the agent has a think tool, a structured-output parser, AND a sheet-dedup tool to avoid repeating prior scene ideas; a separate bulk-run trigger sub-workflow exists alongside the single-row flow.
- Infinite UGC ad generatorBaseRoboNuggetsSystem guidance explicitly instructs an 'unremarkable amateur iPhone photo' realism style and natural dialogue punctuation rules (no em-dashes/quotes, ellipses for pauses) to avoid AI-sounding speech.
- Sora 2 Pro UGC hookAI Ads (Ad Creators Lab)Prompt explicitly names the native format ('shot on iPhone as if for TikTok') to force platform realism; once a Sora 2 Pro face is generated it can't be regenerated identically, which is why index 4 exists.
- Veo 3.1 Frame-to-Video UGC Pipeline (Snubroot + Google Flow)OursLoads the external Snubroot Veo-3 Meta Framework into a Claude Project (timestamp-prompting section manually deleted) so Claude dynamically writes the Veo prompt each time; chains by screenshotting the last frame as the next clip's first frame. Text-to-Video banned for ads; Ingredients-to-Video reserved for unboxing only.
- UGC Ads Full Production (Higgsfield-centered)Ours'Keep the same influencer... Add a realistic [PRODUCT]... Do not change the face' — the same composite reference image is reused across several 4s scenes via Higgsfield Frames mode; a full worked Coca-Cola example is given.
- Infinite UGC video (custom character + product)RoboNuggetsCharacter 'consistency' here is fake continuity: every scene reuses the SAME single generated reference image as its Veo3 start frame rather than chaining last-frame-to-next-frame — the source explicitly flags this as a known limitation, distinct from entry 161's true last-frame chaining.
- UGC ad video autopilot (Nano Banana + Veo3)RoboNuggetsDocuments migrating this same pipeline from ChatGPT-4o-image to Nano Banana on fal.ai — different URL, credential, body params, and an added if-node to match fal.ai's response format.
- Unrestricted-content video generator (Seedream + Wan2.5 vs Veo3)RoboNuggetsAn A/B comparison workflow specifically testing content-restriction boundaries: Veo3 rejected 3 of 7 test cases (fictional character/Joker, foreign-dialogue animated character, viral faceless anime vlog) that Wan2.5 generated without issue.
- Product photo to video adRoboNuggetsTelegram yes/no+comment human-approval gate before any paid image/video spend, with a revision loop on 'no'.
Talking head from a face and a voice11A face speaking a script, short or long form (explainer, founder, authority)
InfiniteTalkKling AvatarOmniHumanThe operator gives: a face photo or clip, a voice sample, the script
- 1Get the voice: clone from a sample or design one
- 2Write the script, split long ones into chunks
- 3Lipsync the face photo or clip to each chunk
- 4Join the chunks, add captions and B-roll
Folded in 11
- Talking-head-over-B-roll UGC recreation (Sora 2 B-roll + ElevenLabs + Infinite Talk + CapCut composite)BaseThe AI Content SystemGenerates B-roll and talking-head as two fully separate assets (Sora2 B-roll + Infinite Talk lipsync clip) then manually composites them in CapCut with background removal and a playback-speed nudge to match dialogue timing to the B-roll.
- InfiniteTalk Long-Form UGC (WAN2GP on RunPod)BaseOursSelf-hosted WAN2GP on a RunPod A40 GPU (~$0.40/hr), login admin/gpuPoor2025 overridable via env vars, port 7862; the pod must always be manually stopped after generation. The most fully specified self-hosted recipe in the vault.
- Longform AI talking-head explainer + B-roll systemBaseRoboNuggetsSplits long scripts into <=5-minute segments specifically to avoid generation timeouts, chaining segments for arbitrarily long output (up to 10 min per generation click).
- RIS talking avatar from a locked Sora faceAI Ads (Ad Creators Lab)Reuses a locked Sora 2 face image (which can't be regenerated identically) and lipsyncs new dialogue onto it via RIS instead of regenerating the face.
- Long-form AI talking-head video (script + voice + photo-to-video)The AI Content SystemTwo interchangeable routes (HeyGen vs VEED Fabric 1.0) merged into one entry since both run the identical script-then-voice-then-face pipeline, differing only by final tool; HeyGen's 'more expressive motion' toggle takes an audio upload, not text.
- Daily AI-persona talking-head social video (n8n queue)OursThe entire generation chain (script/voice/avatar/publish) is stated only in the dashboard's own UI copy, not implemented in this repo — the repo side is a Supabase queue table; live as a dispatch screen, not as generation code.
- Authority thought-leadership video (Story-Sell framework)OursSame 'described in UI copy, generation lives in an external n8n webhook, not this repo' pattern as entry 103, here paired with a research-first step (Perplexity) before script writing.
- Personal-brand AI avatar video pipelineOursOpusClip auto-clips the finished avatar+B-roll video into 10-15 vertical shorts automatically; runs on a FastAPI/Hetzner endpoint (no Zapier/Make) with an explicit banned-tools list (Blotato, HubSpot, HeyReach, Make.com, MindStudio, Sora/Veo3, WordPress, Repurpose.io, AuthorityMax, Dumpling AI, Scalenut).
- Longform AI talking-character videoRoboNuggetsInfiniteTalk's call takes an optional motion-guidance prompt (e.g. 'minimal movement') alongside the audio+image; SeaArt offers an 'edit' feature to place the character into a new scene, unlike free Gemini which has no aspect-ratio control.
- Higgsfield Supercomputer UGC talking-head review videoRoboNuggetsCourse flags visible AI tells (product duplicating in hand, off-brand scream) — quality self-assessed as 'not fully there yet'.
- Baby podcasts AI machine (Hedra lip-sync)RoboNuggetsReuses the ORIGINAL creator's real scraped TikTok audio as the voice track and lipsyncs a generated 'baby' portrait to it -- the audio is captured, not generated.
Copy a real clip's motion7Our person or product doing exactly what happens in a real clip
Kling motion controlWan AnimateSeedance smart editThe operator gives: the real clip, a face or product photo
- 1Give the real clip to copy
- 2Make its first frame with our person or product in it
- 3Copy the clip's movement onto that frame (or swap the person inside the clip)
- 4Add our voice if it speaks
Folded in 7
- Clone any video with a new presenter (Nano Banana Pro + Kling Motion Control)BaseThe AI Content SystemKling Motion Control transfers the exact motion of the original reference video onto the new swapped-person still; used deliberately to test different presenter angles on the same source video.
- Kling Character Swap / Motion CloningBaseOursConfirmed superset of entry 113: identical Kling Motion Control (Higgsfield) generation step, with a Nano Banana Pro face-swap-anchor pre-step spliced in front of it so the cloned motion carries a new identity instead of the original video's person; not wired.
- Seedance 2.5 video edit (smart edit / Omni Reference person, language and object swap)The AI Content SystemEdits an existing video directly in Dreamina's smart-edit mode ('replace the person... keep everything else the same') rather than regenerating from scratch; also supports 'edit with marks' for localized fixes like logo removal, and an Omni Reference mode for combined language+person swap.
- Higgsfield Genjutsu motion-transfer / object-swap video recreationThe AI Content SystemA video-to-video technique (not generate-from-scratch like index 43): Genjutsu directly swaps the character or held object inside an existing reference video while keeping every cut/camera-move/background identical; the automated variant produces 3 split-tested variations in one MCP call.
- Video restyle (video-to-video, 20 style presets)OursTakes only a source video URL and a style-preset name — no text prompt at all; Omni Flash as the executing model is named only in a code comment and is unverified against any real call in the repo.
- Kling Motion Control (motion transfer onto existing character)OursExplicit rule baked into the method: 'motion=video, identity/scene=image' — which reference type is supplied decides whether the run is motion-driven or framing-driven; not wired.
- AI character/motion transfer (motion capture replacement)RoboNuggets'replace' keeps the original background and swaps only the character; 'move' transfers the character into the reference image's full new scene -- two distinct modes of the same model.
Cinematic ad from a locked board11A polished, filmed-looking ad or B-roll with one consistent look and cast
Kling 3Seedance 2.5Veo 3.1The operator gives: product photos, a character or face photo, a setting or mood reference, the brief
- 1Lock the board: character, setting, product, one style paragraph
- 2Make several keyframes per shot from the board (edit, never regenerate)
- 3Pick the best keyframe per shot
- 4Animate each 3 to 5 second shot
- 5Assemble with music and voice-over, upscale the final only
Folded in 11
- Style-locked cinematic short film pipelineBaseAIOpsLocks a single 100-130 word 'style contract' paragraph from reference stills (lens feel/lighting/palette/texture/blocking/era markers) appended verbatim to every keyframe prompt; shots capped at 3-5s because identity drift appears past ~30s of continuous scene; resolution ladder is 720p explore, 1080p keepers, upscale master once.
- Full static + video ad pipelineBaseFree onlineEvery edit step uses masked, surgical 'keep X exactly, change nothing else' instructions with named lighting/lens/materials rather than free regeneration — an edit-not-regenerate discipline chained across 8 separate models.
- Cinematic product advert agent systemBaseRoboNuggetsBuilds one 'Core Elements Board' composite image (character+setting+product multi-angle grid via a wireframe template) as the single reference source for every subsequent scene prompt; stays human-in-the-loop at every gate rather than autopiloting, and final edit is deliberately manual.
- Character-consistent B-roll setAI Ads (Ad Creators Lab)Locks one hero reference then generates many independent scene stills before animating each separately, rather than one continuous narrative; Kling preferred over Veo 3.1 for cost, 5s preferred over 10s for quality.
- Prompt-optimized Kling video with TTS voiceover, captions and 9-platform publishFree onlineThe prompt-rewrite agent is internally labeled 'GPT-4' in the template UI but actually runs gpt-4o-mini — a labeling mismatch worth noting for de-dup against other 'GPT-4' entries.
- Product photography stills to short product video (ComfyUI)Free onlineUses a mask-editor circle drawn directly on the image as prompt shorthand for 'edit inside this region', and a camera-motion line ('Camera slowly orbits left to right...') drives first/last-frame video interpolation; built-in ImageCompare QA node checks output against source.
- Ads dashboard creative pipeline (n8n hand-off)OursTopic-based engine routing lives in the brief hand-off itself: Seedance is forced for restricted topics (cannabis/peptides), Sora 2 is reserved for hero shots only — generation actually happens outside this repo, in the external n8n webhook.
- Kling 3.0 Omni Multi-Shot Product AdOursUses Nano Banana Pro to construct a reusable Kling 'Element' object from 3-4 product references before generation, then drives Kling 3.0's native Multi-Shot mode for 4 fixed shot types in one call rather than 4 separate generations; not wired.
- Veo3.1 multi-format video generator (Sheet-driven)RoboNuggetsOne single Kie.ai request node handles all 4 Veo3.1 modes; the source lists 21 illustrative use cases (UGC try-on, real estate, food, CCTV-style, anime-vlog, logo animation, etc.) but it is built as ONE generic pipeline switched by preset — exactly the format-preset principle in WORKFLOW-ANATOMY.md, catalogued here as one workflow not 21.
- Claude Code + Seedance 2.0 cinematic promo/motion-graphics videoRoboNuggetsClaude Code skill proposes 3 creative directions before generating; the screenshot itself becomes the reference-to-video start frame.
- Cinematic ads agent (storyboard-to-video)RoboNuggetsFace-blurring trick: blurs any person's face in the storyboard panel before sending to the video model, since the model's safety filter blocks recognizable faces in image-to-video -- the blurred storyboard passes the filter and the video model fills in a face itself.
Story in scenes23A narrated video built scene by scene: faceless shorts, POV, quote videos, explainers, long YouTube, sleep stories
Kling or HailuoVeo 3SeedanceThe operator gives: a topic or niche, a voice, length
- 1Give the topic or niche
- 2An agent writes the story as timed scenes
- 3A picture per scene
- 4A clip per scene, or chain each from the last frame
- 5Voice-over, sound, captions, stitch
Folded in 23
- Longform YouTube creator AI agent (JSON2Video)BaseRoboNuggetsWraps the entire generation (voice+images+music+intro clip+captions) into a SINGLE JSON2Video 'create movie' API call via one template ID, rather than assembling each piece separately.
- Monetizable faceless TikToks (fal.ai + ffmpeg)BaseRoboNuggetsGPT-4o vision re-reads the generated image itself to write the motion prompt (not the original text idea), and MMAudio generates video-conditioned ambient sound rather than stock music; the most fully verbatim-captured entry in the whole range.
- Multi-scene narrated story videoBaseRoboNuggetsCharacter-consistency trick: for a 'transform' request, do the transformation once manually then feed that result back in as the reference image for all subsequent scenes.
- Coding-agent explainer/documentary video projectAIOpsA coding agent runs the whole production line from a one-line brief with a free-vs-paid asset fork behind a spend guard (estimates cost first, can hard-stop); 12 recipes total, and the licence may force publishing changes if sold as a service.
- Extend and merge UGC video, auto-publishFree online
- Unhinged POV job-hunting UGC video (GPT prompt chain + Flux + Kling + ElevenLabs + Creatomate)Free onlineExplicit DUPLICATE PIPELINE FLAG in the source data against entry 69 — byte-identical core generation nodes (image prompt / script / video caption generation), 69 adds transcription+caption-gen+multiplatform publish on top of this exact pipeline.
- Unhinged POV job-hunting UGC video, auto-published to 5 platformsFree onlineByte-identical generation nodes to entry 68 (confirmed via direct dict-equality check on raw node JSON) — this entry is 68's exact pipeline plus a transcription/caption-generation step and 5-platform publish (TikTok/IG/YouTube/Facebook/LinkedIn via Upload-Post).
- Cinematic nature quote video (Flux stills + Kling motion + ElevenLabs ambient soundscape + burned-in quote text)Free onlineNo LLM node anywhere in the pipeline — prompts are fixed templates with Google Sheet-column variables only; overlay is a Code-node FFmpeg drawtext with a specific Thai font (Kanit-Italic), not a captioning model.
- Faceless YouTube Short from a topic (ElevenLabs voice + Hailuo AI visuals)Free onlinePaid ($10) n8n template with every node parameter blanked out in the source — pipeline shape is fully described but no verbatim prompt text is recoverable from any step.
- End-to-end video production inside Claude Code (script to render)Free onlineSlash-command-driven, self-hosted, named-model pipeline where every stage (voice/image/music/video/talking-head) is separately swappable with its own per-unit cost, and brand identity is stored as versioned JSON per-brand so templates reuse across brands; each per-model skill file bundles its own worked prompt examples (e.g. Qwen-Image-Edit's guidance-scale 2.0 sweet spot for identity-preserving composites).
- Scripted vertical short (Short studio)OursCost (~$2.80/short) is computed but deliberately never shown to the operator; validates the script structurally (beat count, word count, scene-hint count) before allowing spend. Shares the identical produce_short_scripted verb and validation rules with entry 111, which wraps this exact render step in a full remix/publish lifecycle.
- Shorts full lifecycle (script→render→voice/caption/remix→crosspost)OursWraps the exact same produce_short_scripted render step and validation rules as entry 107 in a full post-production/publish lifecycle (remix, voice-swap, caption-fix, crosspost, harvest-viral) — same core render, entry 107 is the bare render with no lifecycle around it; live, creator studio.
- Documentary/history-style Short — per-shot styled animationOurs'Continues' shots reuse the previous shot's last frame for continuity; a headless-Claude vision judge automatically picks the best of several frame_candidates rather than a human; Omni is deliberately kept over Seedance 2.5 specifically because Omni does not require a downloaded voice sample (per Simon's 2026-09-24 note); style is rotated per brand/slug as a deliberate A/B variable; ~$1.20-2.80/short.
- Veo3 faceless character video clipRoboNuggets
- Narrative-chained video extension (Veo3)RoboNuggetsCore reusable technique named explicitly by the source: true last-frame-to-first-frame chaining (distinct from entry 160's fake single-reference-image 'consistency') — any narrative-length video is buildable by repeating extract-last-frame → generate-next → merge.
- Live-action character-swap selfie videoRoboNuggetsUses Kling's first/last-frame mode to animate transitions between successive AI selfie stills, building a multi-scene story purely from still images.
- Quote YouTube Shorts / IG Reels with cloned voiceRoboNuggetsPulls viral quotes straight from a real transcript (tactiq.io) rather than from an idea/topic brief.
- Zero-human faceless video generator (Kling direct)RoboNuggetsDirect text-to-video with no image/start-frame stage at all, the simplest of the faceless-video family in this range.
- Faceless video generator with multi-clip image-to-video stitchingRoboNuggetsv2.0 successor to r16: swaps direct text-to-video for a 3-image-per-idea start-frame + image-to-video approach, explicitly flagged by the source as materially different, not a duplicate of r16.
- Faceless POV AI video machine (n8n)RoboNuggetsThree-agent n8n structure (ideas/production/publishing); uses a fixed Creatomate template (Audio-1..5/Video-1..5/Text-1..5 slots) to render the final stitched video in one call.
- VeoRobo: Veo3-fast ASMR factory for any nicheRoboNuggetsRandomly assembles ASMR cutting-scene elements (instrument/colour/material/object/surface pools) via structured JSON before building the video prompt -- a randomized-recipe generator rather than a topic-driven one.
- Long-form sleep story videoRoboNuggetsChatterbox chosen over ElevenLabs specifically for cost and custom-voice access; a 30-second hard output cap forces splitting narration into ~70-75 word chunks per chapter.
- Multi-model batch video generatorRoboNuggetsGenerates the SAME idea in parallel across three different video models (Sora2/Wan2.5/Veo3) for side-by-side comparison, not for final output selection.
Start and end frame transition5A short move between two pictures: product reveal, 3D spin, exploded view, food shot, logo animation
KlingVeo 3.1SeedanceThe operator gives: a product or logo image, what the move should be
- 1Give the product or logo picture
- 2Make the first frame and the last frame (reveal, spin, exploded view, new angle)
- 3Generate the move between them
- 4Optionally join several moves
Folded in 5
- 3D-style product animation (Nano Banana Pro + Veo3.1, Airtable-driven)BaseRoboNuggetsDeliberately staged (prompts→images→video) so a human can review/regenerate the start/end images before committing to the pricier video step; the ending-image prompt is intentionally simpler since it reuses the start image as its own input to keep both frames in the same scene.
- Food-photo-to-video (Kling image-to-video, start/end frame)BaseRoboNuggetsPositioned as an upsell extension bolted onto entry 172's food-photo service — the two Nano Banana Pro images generated there (hero + angle variant) become the fixed start and end frames fed straight into one Kling image-to-video call; the presenter explicitly calls this 'outside the scope of this specific tutorial' and skips node-level setup.
- Start/end-frame scene transitionsAI Ads (Ad Creators Lab)Generates two still keyframes (start+end) independently, then asks the video model to generate the transition motion between them rather than animating forward from one frame.
- Logo animationRoboNuggetsPrompt encodes a fixed 3-part timeline (0-2s intro / 2-60s main / final scene) with per-scene motion intensity.
- Product motion-graphics transition videoRoboNuggetsFrame-generation is a manual/interactive playground step (not wired into n8n) that edits/combines two reference images into start and end frames before the video model bridges them.
Animated ad in a house style10Animation: claymation, 3D, anime, talking objects, brand storyboards
Google OmniKlingThe operator gives: the script or idea, a style pick, characters or product
- 1Pick the house style (claymation, 3D, anime, talking object)
- 2An agent writes the storyboard in that style
- 3A character sheet keeps the cast the same
- 4Picture per scene, then animate each
- 5Voice, music, stitch
Folded in 10
- Crochet/claymation storyboard-to-video adBaseAI Ads (Ad Creators Lab)Claude Skill splits the script phrase-by-phrase into a 5-part image structure + 4-part image-to-video structure with Type A/B/C director's-note motion classes (self-contained / state-change-with-end-frame / held-subtle-motion).
- Multi-style AI animation ad (Omni Ad Engine SOP via Claude)BaseThe AI Content SystemClaude hands the user paste-ready prompts (a 'Locked Product Block' replacing a reference image, since Omni is prompted text-only) rather than generating end-to-end itself, unlike index 35.
- Talking product/object avatarAI Ads (Ad Creators Lab)Same still animated across 4 competing engines and compared on cost/quality (Veo best/priciest, Seedance cheapest but hands break) rather than fixed on one engine.
- Claymation frame prompts (ChatGPT method)AI Ads (Ad Creators Lab)Same 4-stage shape (script>scenes>image prompt>generate>reference-based consistency) as index 9, via a bare ChatGPT custom GPT instead of Poppy's project system.
- Claymation frame prompts (Poppy method)AI Ads (Ad Creators Lab)Same 4-stage shape as index 8; uses a fixed consistency-prefix phrase 'Keeping the character the same, execute this prompt: [x]' and Poppy's project-level house-style knowledge base instead of a bare custom GPT.
- 3D animated explainer ad generator (packaged Replit + KIE.AI app)The AI Content SystemA fully self-contained packaged Replit+KIE.AI app that generates end-to-end from one launch prompt, distinct from index 36 which instead hands the user paste-ready prompts rather than generating end to end itself.
- Stylized anime video with consistent character/styleFree onlineTreats image-to-video as a style-preservation problem: deliberately LOW motion scale (30-40%) paired with HIGH guidance scale (8-12) to stop drift off the anime look, plus explicit anti-morph phrasing ('no morphing', 'high fidelity') in the motion prompt.
- 3D/Stylized Character RecreationOursPing-pong loop per new scene: back to Nano Banana for the next scene still, back to Veo with a new literal action prompt appending 'No background sounds.', with matching negative prompts on both sides to remove unwanted elements.
- Medical/Biological 3D AnimationOurs'A hyper-realistic 3D visualization of [specific biological process or product mechanism]' with 4 full worked examples for the base image and 4 more for the motion prompt — 8 worked examples total across both formulas.
- Higgsfield MCP brand book to animated logoRoboNuggetsCompares two brand-book generation methods (Nano Banana 2 vs GPT Image 2) side by side before picking one to proceed with.
Consistent AI persona11One made-up person who posts again and again and always looks and sounds the same
Nano Banana Pro + SeedanceSora 2 cameoThe operator gives: a persona idea (or reference photos), a voice, then each post's topic
- 1Build the persona once: character sheet, face-lock phrases, look
- 2Give it a voice and save both
- 3Each post: give only the topic or trend
- 4The post starts from the saved face and voice
- 5Talking, selfie or product posts from the same persona
Folded in 11
- AI Persona Ad Pipeline (ICP → Face-Lock → Start Frame → Video)BaseOursMandatory JSON face-preservation block {"face_preservation": {"preserve_original": true, "accuracy_level": "100% identical to reference"}} injected into every subsequent generation of the persona. Notes flag an unresolved policy divergence between two copies of the CREATIVE-PRODUCTION-FRAMEWORK.md compliance section.
- AI-Influencer network — Pipeline B character setup (one-time)BaseOurs7-Layer Persona Blueprint (Bio/Lore/Topics/Adjectives/Style/Examples/Voice), then a dedicated LoRA trained on fal.ai from the locked face seed (~$2-5), plus a recurring environment anchor for cross-shoot consistency. One of the most fully specified recipes in the whole vault.
- AI influencer content generatorBaseRoboNuggetsBuilds a reusable multi-angle character reference set (front/side/3-quarter/back + expressions) plus a 360-rotation video once, then reuses it as the consistency anchor for every subsequent post.
- Consistent character across many images/scenes (Midjourney + Nano Banana)Free onlineOnce the reference image carries identity, every downstream prompt is deliberately short and generic ('woman, full body shot') rather than re-described — consistency comes entirely from the image reference, with Midjourney used once for a high-quality base and Nano Banana for every derivative; scene/prop assets are built independently of the character and reused across multiple characters.
- Sora 2 Character Cameo (reusable @tag)OursDescribe a character via 6 micro-detail categories, generate a plain base video, then use Sora's 'Create Character' menu to trim to the clearest section and save it as a permanent @tag reusable in future prompts.
- Nano Banana Pro/2.0 — Ads & Persona Formula LibraryOurs5 literal face-lock phrases with matching negative prompts, a 5-rule Persona Consistency Protocol, and a vertical-aesthetic lookup table mapping brand type to look (VERO→editorial luxury, cannabis→relaxed lifestyle).
- Ultra-Realism Character GenerationOursLiteral imperfection keyword fragment ('natural skin texture, small imperfections, realistic pores, slight blemishes') deliberately fights AI over-smoothing before a 4K upscale pass.
- AI-Influencer network — Pipeline A (repurpose + trend avatar)OursTwo distinct sub-workflows bundled under one skill: real-video repurposing into captioned shorts, and a trend-driven HeyGen avatar of Simon on a scheduled cadence table — both approval-gated via Telegram.
- AI-Influencer network — Pipeline B trend-driven video (fully executable)OursEvaluator-Optimizer loop: Claude Haiku scores the script (threshold ≥7 avg, no criterion ≤4, max 2 retries) before voice generation proceeds; explicit rule 'do NOT generate an AI image for the persona — use Face Seed Direct'. Per the pointer inventory, the most fully executable recipe in the entire estate (literal API bodies throughout).
- Sora2/Sora2 Pro autopilot with Character Cameos (Sheet-driven)RoboNuggetsCharacter Cameos: create a reusable Sora character profile (video upload + username + description) once in the Sora app, set visibility to public, then reference it via '@username' inline in any future Kie.ai prompt text. This lesson explicitly supersedes an earlier n23 version, folded in here.
- Sora2 Character Cameo video (client-facing web form)RoboNuggetsClient-facing web-form trigger (one submission per job) versus entry 166's bulk Google-Sheet run; uniquely documents the manual cameo-creation workaround for photorealistic real people (generate an AI character video first, then 'create a character' from it), and found cameo+product-reference combos give poorer product fidelity than product-only prompts without a cameo.
Song with visuals cut to the beat4Music videos, lyric edits, lofi loops, AI covers
Suno + KlingHailuoThe operator gives: a song or a music brief, characters or a theme
- 1Lock the song (make it or bring it)
- 2Split it into sections, find the beats
- 3Clips per section, the best engine on the chorus
- 4Cut the clips to the beat, add lyrics
Folded in 4
- K-pop AI cover music video (voice clone + character bible + segmented video)BaseFree onlineCost-tiers the video engine by which song section matters most for retention — expensive 4-reference Kling 2 only on the chorus 'killing part', cheap 1-reference Hailuo 02 on verses, free SunoMV visualizer on transitions, turning a $10+ full-Kling video into $2-3; beat-alignment extracts word-level timestamps from the vocal track itself rather than manual scrubbing.
- Music-video lyric-synced editAI Ads (Ad Creators Lab)
- Lo-fi ambient loop video (manual HaiLuo + Suno + DaVinci)RoboNuggetsFully manual/UI pipeline explicitly NOT automated; DaVinci template auto-crops the frame to hide the HaiLuo watermark.
- Auto music album video creatorRoboNuggetsBuilds the FFmpeg concat command by asking ChatGPT to write it, then loops one static cover image behind all stitched songs; optional custom-GPT step auto-generates YouTube chapter timestamps from song titles/durations.
Pictures3
Static ad with copy17A finished picture ad: headline, copy and image that agree
GPT-Image 2.5Nano Banana ProThe operator gives: a product photo, brand site, the offer
- 1Give the product photo, brand site and offer
- 2Pick the angle and buyer stage
- 3Write the exact words on the ad first
- 4Make the picture around the words with the real product
- 5Check picture, headline and copy agree
- 6Variants change one thing at a time
Folded in 17
- Static ad generated from a brand's website + logo + product photoBaseFree onlineLanguage is detected purely from the domain TLD (not content analysis), and the brand-insight extraction step is explicitly instructed to 'never invent facts' — strictly grounded in scraped site text; a conditional rule keeps any UI screenshot in the product photo legible rather than letting the model restyle it away.
- Static ad concept + image (money-gated)BaseOurs5-step 'craft bible' concepting order enforced inside the Claude prompt itself (strategy first, then exact on-image copy in voice-of-customer BEFORE the visual, then pick one named concept type, then write the model-specific prompt, then a congruence check that visual/headline/copy sit in the same awareness stage) — the only cockpit in the estate with real, checked-in provider-call code; live and wired end to end.
- Static ad creation (research to Canva)AI Ads (Ad Creators Lab)Sources concepts at a defined 70/20/10 ratio across AdSpy tools (4Print/Atria/adnova.ai) before writing headlines matched to awareness levels.
- Winning-ad pattern to new static ad variationFree onlineSafety-gated to paused ads only, and only regenerates once a pattern shows a real lift (min 3 samples / 1.3x rule) rather than always spinning variants.
- Ad creative variant generator (image analysis + GPT-5 recipe-based rewriting + Nano Banana edit)Free onlineUses a named 'creative recipe' menu (metaphor, pointer-platter, comparison-combo) as the variant-generation parameter rather than free-text creative direction — a reusable structured-variation trick.
- Fact-grounded infographic imageFree onlinePrompt explicitly instructs 'use only the provided facts' with exact label text required, plus a manual QA checklist step to catch hallucinated numbers/labels before publishing.
- Static-ad structural guides (Recraft v3 / gpt-image-2 / Nano Banana)Oursgpt-image-2 carries a literal PRESERVE-clause i2i pattern that keeps a real product's shape/material/label/typography/proportions pixel-identical while only the surrounding scene changes.
- Scroll-stop prompt trio (product photography/video templates)OursPer-object component lists given verbatim for laptop/phone/shoe/food for the Deconstructed/Exploded Shot; also includes a companion HTML delivery-page mechanism.
- AI news infographicRoboNuggetsA daily-scheduled n8n pipeline where a Tavily-backed research agent finds same-day news, then a dedicated image-prompt agent merges research+caption+a fixed infographic style guide so the 4o image model renders accurate on-image text and a source line.
- Fantastic Posters bulk client poster/brochure generatorRoboNuggetsBatch-adapts one poster template across a whole folder of listing photos while keeping consistent font/logo/palette (real-estate use case).
- Fable 5 ad content factory (reverse-engineer top ads, regenerate)RoboNuggets
- Higgsfield Supercomputer batch product ad imagesRoboNuggetsSingle command 'Make 10 image ads for this product' triggers autonomous research + skill load + ad-creative-pack reference; shows a cost checkpoint (model/aspect/resolution/duration/cost) before spend.
- Blog/news summary to social posts with an AI-generated Instagram imageRoboNuggetsPre-merged 3x-taught lesson: shortened link -> article summary -> single photorealistic image -> caption -> auto-post, no manual step.
- Batch image ad splitterRoboNuggetsUses the free Nanobanana tier via OpenRouter (cost $0) rather than a paid image API, explicitly flagged as possibly not permanent.
- Brand-grounded ChatGPT image agent (via n8n MCP)RoboNuggetsChatGPT itself calls an n8n workflow as an MCP tool to run the actual generation, bridging the chat agent and the image pipeline live via MCP, with a fixed append-rule forcing brand-book style without copying its literal contents.
- 30-day branded image-ad batch (making step inside a posting-focused lesson)RoboNuggetsThin/partial entry leaning entirely on a separate lesson's (R56) engine setup; posting step deliberately excluded per the making-vs-posting rule.
- Render Studio brand-to-carousel/merch generatorRoboNuggetsBuilt and framed primarily as a sellable white-label SaaS app (one-prompt reskin for merch mockups instead of carousels for different clients), rather than a one-off content recipe.
Product photo shoot21Product pictures: on a person, in a scene, new angles, whole sets for a campaign
Nano Banana ProGPT-Image 2.5SeedreamThe operator gives: product photos, optional face or scene photos, how many
- 1Give the real product photos (and a face or scene if wanted)
- 2Place the real product on a person or in a scene; never repaint it
- 3Change one thing at a time: background, light, angle
- 4Pick from a grid
- 5Crop and resize to every ad size
Folded in 21
- Influencer-holding-product composite image (Gemini 2.5 Flash Image edit)BaseFree onlineSingle-call composite whose prompt explicitly instructs 'not a professional photo... like a friend is taking a picture' plus a varied-angle instruction to defeat the polished-studio look.
- One base image expanded into a full ad campaignBaseFree online'One prompt to campaign system': variants change exactly one variable at a time using a named reference-image role system (product ref / background ref / lighting ref), then later formats are produced by cropping/resizing the same asset rather than regenerating.
- Nano Banana Pro — Action TransferBaseOurs'make the model in the first image perform the action in the second image' — the action is taken from a stock/Getty reference photo, with the product fed in as a third 'ingredient' image; regenerate 2-3x if imperfect.
- 5-slot static image, then motion handoffAIOpsA named '5-slot' prompt structure (scene, subject, important details, use case, constraints) plus explicit anti-slop editing rules (kill mood words like 'cinematic, 8k, masterpiece', quote in-image text literally, one revision per turn); Thinking Mode self-verifies against the brief and holds 8-frame batch consistency.
- Mood-board to product-on-model shotAIOpsNode-canvas pipeline where a text model only writes the brief and the image model only renders it ('put this on the model' / 'different angle'); explicitly composites the REAL product photo onto the model rather than a text description of the product.
- Monthly content calendar to approved social postsFree onlineThree hard guardrails baked into the workflow itself: real product photos are never repainted, price/approve gate before any generation spend, and a human approves every post before it goes out.
- Auto-written product photography prompt + Gemini Nano Banana compositeFree onlineCheap-LLM-writes/expensive-model-executes cost pattern — a cheap model (flash-lite) drafts the photography prompt (with verbatim source typos 'protography'/'photogaric' preserved) before the pricier image model runs it.
- 50 UGC product images from one photoFree onlineOne LLM call writes all 50 UGC prompts as a single JSON array up front (not per-image), then loops the image model over them; ~$0.039/image, ~$2 for 50.
- Product creative image via a form + Defapi Nano Banana APIFree onlineThin/generic recipe whose only distinctive feature is routing Nano Banana through an alternate paid API host (Defapi.org) instead of the usual fal.ai/Kie.ai/muAPI routes.
- New picture (Studio/Make chat-driven image edit)OursProvider/model names only surface via a reply-text-stripping regex in words.ts rather than being explicitly called out anywhere else in the code — the regenerate_image verb is the actual dispatch mechanism; live, Studio/Make screen.
- Static/social-post image (t2i and i2i, model-picker)OursThe prompt shown to the operator is pre-seeded from draft.image_prompt and is editable before send; the actual provider call is not in this repo — it is an insert-only row, with muAPI named only in a type comment, not called code.
- Freepik Spaces Fashion/Ecommerce Photography & VideoOursA Pinterest pose reference is converted into a stick-figure pose reference via Nano Banana before being used to pose the model — a distinct 'pose-hack' method not seen elsewhere in this range.
- MUAPI Gateway — 6 executable API recipesOursOne gateway (muAPI) with verbatim bash/python/CLI call paths, async polling, and a mandatory single smoke-test before mass regeneration across 6 distinct recipe types; this is the same API gateway the live creative-pipeline/cockpit consumer/providers/image.mjs calls, but these skill files are reference doctrine, not the running code itself.
- Flora AI — 8 node-based canvas recipesOursRecipe 6 (UGC-style video) is the only single-generation recipe in this entire range that fuses actor + product + script + auto-lipsync audio in ONE Veo 3.1 Ingredients-mode call, rather than stitching separate voice/lipsync steps.
- Digital product image generator (Nano Banana Pro, Telegram agent)RoboNuggetsNano Banana Pro chosen specifically because it correctly spells rendered text and hits 4K, unlike base Nano Banana (misspellings, lower res) or the native Gemini app (returns framed/combined images unsuitable for print delivery); no native transparent-background support — the 'transparency' squares are baked into the image, not a real alpha channel.
- Restaurant food-photo enhancement (Nano Banana Pro)RoboNuggetsTwo parallel paths taught: a fully manual, free Google Gemini browser path for one-off client pitching (watermarked) versus a scaled n8n+Airtable automation sourcing leads via an Apify Google Maps scraper for whole-menu batch production; priced in the course as a $10-20/image service against near-zero AI generation cost.
- GPT Image 2 brand asset batch (Claude + GPT Image 2 skill)RoboNuggetsGenerates a numbered 5x5 grid in one call so the operator picks a cell by number, then only upscales that cell.
- Claude /generate skill for product image adsRoboNuggetsRoutes automatically to the cheapest provider (Kie.ai > fal.ai > Wavespeed) and keeps a reusable 'styles' shelf, positioned as a pay-as-you-go Higgsfield-subscription replacement.
- AI character art factory (TikTok image carousel)RoboNuggets3 chained GPT-4.1 agents each randomly pick from a 50-style and 50-character list before a Think-tool agent builds the final prompt -- fully autonomous ideation with no operator input required.
- shirtRobo: A.G.E.N.T.-framework shirt design generatorRoboNuggetsThe agent's own system prompt is generated once, offline, by a separate interactive 'A.G.E.N.T. Framework GPT' -- a meta-step not seen elsewhere in this range.
- Multi-model creative engine (image + video ad generator agent)RoboNuggetsGeneral-purpose multi-model 'connect-anything' engine: maintains a persistent prompt-best-practices file (BOPA framework for influencer consistency, SEALCaM for cinematic) and can be extended to new models by pasting a provider's API docs page for the agent to write its own connector.
Copy a picture's look6A new picture in the exact look of one you like (Pinterest, a competitor, a film still)
Nano Banana ProGPT-Image 2.5The operator gives: a reference picture, what to put in it
- 1Give the picture whose look you want
- 2Turn it into a detailed description (light, colour, lens, layout)
- 3Change only the subject or product in that description
- 4Make the new picture
Folded in 6
- Realistic reference-JSON images with Nano Banana ProBaseThe AI Content SystemThe core technique nearly every other workflow in this course reuses as its starting-frame method: extract a detailed JSON description of a reference photo (lighting/coloring/composition) then regenerate with a new subject or product, iterating via the same JSON.
- Reference-DNA style extraction and transferBaseAIOpsTwo extraction modes: 'Full Profile' (forensic JSON incl. hex colours, camera/lens estimate, for cloning-with-modifications) vs 'Style-Only' (strips content, keeps just a reusable 'style_dna' paragraph) - a deliberate content-vs-style extraction fork.
- Fill-in-the-blank brand image templatesAIOpsThe tool itself only outputs the PROMPT text, not an image; drawing happens in a separate tool. Working prompts get saved as a permanent house template file (brand/image-templates.md).
- Nano Banana Pro — end-to-end persona/product imageOursClaude converts a Pinterest reference photo directly into a JSON prompt ('Give me a detailed JSON prompt for this image'), the operator then feeds Claude a product/clothing photo plus a free-text instruction to update that same JSON before generation.
- Nano Banana Pro — 20 Curated PromptsOursA pure prompt library, not a pipeline — 20 fully-written, distinct style prompts to paste directly into Nano Banana Pro.
- Universal Prompt-Writing / JSON Prompt SystemOursThe JSON-round-trip technique (image→JSON via Claude→flatten to plain text) used as the base technique throughout the Nano Banana/medical/3D workflows in this range.
Edit1
Re-edit real footage8A new edit of footage we already have: shorts from a long video, a reference-led recut, a realistic UGC edit, reframe to 9:16
ChatCutFFmpegRemotionThe operator gives: the footage, optionally a reference edit
- 1Give the footage and, if wanted, an edit to copy
- 2Transcribe it and read the edit to copy
- 3Plan the cut in words
- 4Cut, caption, reframe to each size
- 5Check for jumps and sound pops
Folded in 8
- Transcript-driven chat edit of raw footageBaseAIOpsEdits by describing the desired cut in plain English against transcripts rather than reviewing frames, keeping cost low versus frame-by-frame analysis; saves style preferences to a memory file for future runs.
- Video Reference Recreation (ChatCut)BaseOursNot generative — a real-footage re-edit. ChatCut tables the reference's full structure (shots/rhythm/layout/captions/motion-graphics/B-roll/transitions/music) into reference.md before any rebuild starts. Cut/caption/motion-graphics free on the Claude plan; AI video/voice/music/images credit-gated.
- Realistic AI UGC video editAI Ads (Ad Creators Lab)Blends small green-screen AI inserts into real B-roll and applies a 'Wiggle' effect for handheld realism; explicit philosophy is 'no flashy transitions/VFX'.
- CapCut UGC reaction-video editAI Ads (Ad Creators Lab)Extracts and resyncs audio from a different take than the chosen visual take ('remix audio to best visual') to combine the best of each.
- Recreate a Video (ChatCut sister preset)OursSister preset to entry 148, offering 3 selectable modes; the multi-reference generation mode is the only one of the three that costs Seedance credits, i.e. can go generative unlike 148's pure re-edit.
- Hyperframes video edit + motion-graphics finish (BIT framework)RoboNuggetsSelf-improving skill capture (BIT framework: Build-Integrate-Tune) banks learnings from each edit session back into the skill file for next time.
- Infinite short clips from longform video (Klap)RoboNuggetsRepurposing only -- takes an existing finished video and splits it into shorts; no generation model involved at all.
- AI vertical reframe system (Lumalabs)RoboNuggetsUses an AI reframe model (not a plain crop) so the reframed 9:16 version can pan/recompose intelligently within each 29s chunk (Lumalabs API's 30s cap).
Building blocks7
Finishing pass5Upscale, smoother motion, watermark removal: added to the end of any video button
TopazRIFEFolded in 5
- General AI-video upscale (Topaz + CapCut)BaseThe AI Content System
- 4K 60fps video finishing passBaseRoboNuggetsDeliberately runs interpolation before audio re-attachment because RIFE strips audio as a side effect, so a separate FFmpeg compose step re-layers the original audio onto the silent higher-fps output. Functions as a finishing pass bolted onto other lessons' raw Veo3 outputs.
- Sora 2 upscale and realism grade (Higgsfield + Topaz + CapCut)The AI Content SystemExtends index 33 with an extra Higgsfield Enhancer pass specific to Sora 2 Regular (to reach Pro-comparable quality), plus named CapCut grade values (Contrast -7, Exposure +3, Motion Blur 20%, Particles 10%).
- Upscaling (image + video finishing)Ours
- Sora watermark removerRoboNuggets
Make a voice11A designed or cloned voice other buttons use
ElevenLabsMiniMaxOmniVoiceThe operator gives: a voice sample or a description
Design from a description or a face photo (3 candidates, pick one), or clone from a clean sample; face never uploaded (black-screen trick)
Folded in 11
- Custom AI UGC voice design (ElevenLabs + Gemini Gem)BaseThe AI Content SystemA dedicated Gemini Gem, pre-loaded with ElevenLabs' own best-practices doc, writes the voice-design prompt from a character photo; explicit do-not-use list bans the ElevenLabs Voice Changer and premade catalogue voices as 'sound too AI'.
- Voice reference asset (face-hidden)BaseOursDeliberate face-hiding trick: pure deterministic ffmpeg (no AI/API call at all) renders the extracted voice track against a black frame so a real person's voice can be reused for cloning (into Google Omni) without their face ever being uploaded to a model.
- ElevenLabs voice design from a text descriptorAI Ads (Ad Creators Lab)10 literal voice-descriptor examples captured (specific accent/age/tone combinations) as the reusable descriptor recipe.
- Coherent voice clone across clipsAI Ads (Ad Creators Lab)Explicitly removes background music from the extracted clip before cloning, to keep the clone clean for reuse.
- Real-creator voice clone with matched backing trackAI Ads (Ad Creators Lab)Deliberately does NOT remove background noise before cloning a real creator's YouTube audio, contrary to usual advice, to preserve the creator's ambience.
- Local voice clone, dub, and transcriptAIOpsRuns entirely locally with no cloud upload; a licence trap is flagged: the app licence (AGPL, commercial OK) and the default OmniVoice model weights (CC-BY-NC, non-commercial only) are two separate documents that must be checked separately before selling any output.
- Agent-driven ElevenLabs audio productionAIOpsOne MCP server chained into a fully unattended watched-folder podcast-to-clip routine: isolate audio, diarize/transcribe, an LLM step identifies the 3 highest-emotion 60-90s segments as JSON, then renders a 5s cloned-voice hook intro per segment; explicit agent-safety flags (trust_context low, response_timeout_secs 180) for unattended use.
- One-agent image/video/voice/music generation (MiniMax)AIOpsA multi-modal MCP tool (image+video+voice+music), family set to 'voice' since the course rates its image/video capability as duplicating Icarus's existing gpt-image-2/Higgsfield stack, leaving voice cloning + 40-language TTS + music as the only additive slice. API key region must match the API host or auth fails.
- AI Voice Generation for UGC (MiniMax primary / ElevenLabs custom-only)OursGemini writes the ElevenLabs voice-design prompt itself, explicitly instructed to make the voice 'sound like it's in an actual room'; ElevenLabs restricted to custom-only voices, never pre-made/speech-to-speech.
- AI podcast episode generator (script + voice)RoboNuggets
- Cloned voice via ElevenLabs voice sharingRoboNuggetsMakes a monetizable voice ASSET (shareable to ElevenLabs Voice Library with Stripe payouts), not a piece of content.
Make a music track3A jingle or backing track other buttons use
SunoThe operator gives: a music brief
Folded in 3
- Suno ad jingle/songBaseAI Ads (Ad Creators Lab)Style prompt written to avoid naming real artists, since Suno rejects those names.
- Suno V5 song remixerRoboNuggetsRemixes an existing uploaded track rather than generating a new one from scratch.
- Lo-fi music album generator (Suno via PiAPI, Make.com automated)RoboNuggetsv2.0 upgrade adds 2 takes per title and an instrumental-only flag; uses PiAPI as a workaround since Suno's direct API was deprecated April 2025.
Script frameworks7Ad script shapes (hook, story, listicle, transformation) used as the script step of any video button
ClaudeThe operator gives: product facts
Folded in 7
- UGC brief template with 8-beat script arcBaseAI Ads (Ad Creators Lab)A fill-in TEMPLATE (concept+hooks table, 8-beat arc from 'The Realization' to 'The Invitation', 10-slot B-roll shot list, 3-hook editor brief table), not a single worked example.
- Ad Script Writing FrameworksBaseOurs3 named frameworks plus one literal 'Start Prompt' quoted in full and reused verbatim across multiple files — feeds every script-writing step in the video workflows above.
- Triple G scriptwriting framework (Goal, Gap, Gains)AI Ads (Ad Creators Lab)Reusable Goal-Gap-Gains structure with no tool prompts; text-only output, family assigned by nearest downstream use since the enum has no script-only category.
- Transformation trifecta scriptwriting frameworkAI Ads (Ad Creators Lab)Stages the same offer across 3 different levels of perceived viewer effort (Value Equation), not just one hook variant.
- Curiosity loop scriptwriting frameworkAI Ads (Ad Creators Lab)5-step stacked-curiosity-loop structure (clarify emotion > 3 open loops > payoff); a Valentine's seasonal variant is treated as the same method, not a separate workflow.
- CURE listicle scriptwriting frameworkAI Ads (Ad Creators Lab)Curiosity/Uncover/Reveal/Evidence+CTA (CURE) listicle structure.
- Video ad creative frameworks (hooks, scripts, formats, A/B testing) — Claude skillFree onlineA strategy/framework skill rather than a generation pipeline — packages direct-response copywriting structures (5 hook formulas, 5 script-body frameworks with per-beat timing) as an installable Claude Code skill runnable purely as planning guidance with no API key, delegating actual video rendering entirely to Creatify's API.
Made by code, no AI model15Motion graphics, carousels, slide decks and sites built from templates; a layout layer on top of AI pictures
RemotionHyperframesHTMLFolded in 15
- Video-as-code motion graphics renderBaseFree onlineNo generative model at all: animation is built as paused GSAP timelines the agent writes and previews before a final deterministic render (window.__timelines pattern) — pure code, not a diffusion/video model.
- Batch ad-variant video generator (1 template x CSV = N ads)BaseFree onlinePure code-template finishing layer (no generative model) over hero shots made elsewhere; enforces a named 5-type hook taxonomy, a one-variable-at-a-time testing order (hook > offer > CTA > format), a 60/30/10 budget guide, and a message-match rule tying each hook to its CTA.
- Code-defined repeatable clips with local captionsFree onlineLoad-bearing agent instruction 'do not render until told' — the skill is built to iterate cheaply on structure before paying the render cost.
- HTML-rendered social carousel and cover cardsFree onlineCards are literal HTML/CSS rendered and screenshotted via Playwright, not an image model — 28 layouts x 10 themes x 2 design systems as a swappable preset library.
- Beat-synced product launch / sizzle filmFree online'Sound design leads picture' — the music track and its exact drop timecode are locked before any visual timing is built; verbatim GSAP code hard-codes a 5-beat arc (hook/tease/reveal/montage/end-card) to a specific DROP=9.0s constant.
- Animated testimonial / review quote videoFree onlineFixed 5-element reveal order (quote-mark, quote body line-by-line, emphasis highlight, star rating fill only AFTER the quote finishes reading, then author block) enforced by a verbatim TypeScript validation function that blocks render if the 'highlight' prop isn't an exact substring of the quote.
- Carousel slide deck (Studio/Make)OursNo engine/provider is named anywhere in the code for this verb — deck.ts renders the carousel deck itself; live, Studio/Make screen.
- Scroll-stop scrolly-telling site buildOursRepurposes an already-produced video into a code-generated scroll-scrubbing site rather than generating any new creative content; code-complete with an error-recovery table.
- Brand-styled animated HTML slide deck (frontend-slides skill)RoboNuggetsPure code-template deck, no generative image/video model; Claude writes animated HTML/CSS/JS components directly.
- Claude Design slide deck with MCP-generated imagesRoboNuggets
- Claude Design automated deck with live data and video clipsRoboNuggetsWeekly-refresh routine scheduled by Claude Code re-pulls live spreadsheet data into the deck automatically.
- Marp markdown slide decksRoboNuggets
- Claude Design brand asset generationRoboNuggets
- Claude Code /design skill assetsRoboNuggetsDeploys directly to a live Vercel URL from the design skill itself, unlike Claude Design which needs manual deploy.
- Claude Code + Remotion promo-video generatorRoboNuggetsThe only entry where video is produced entirely by an LLM writing render code (Remotion/React) against real scraped site assets -- no generative image or video model is called at all.
Not a creative workflow6Checklists, model testers, hand-off notes, brand identity setup, diagrams
Folded in 6
- Shorts QC and iteration desk: thumbnail + hookFree onlineOverview/menu lesson for one of six QC 'desks' — no model or generation step at all, purely a manual review checklist.
- Google Flow hand-off assembly (manual step)OursDeliberately ends at a manual hand-off to a human operator working Google Flow's Frames-to-Video UI by hand — the return path (bringing finished Flow clips back into the pipeline) is explicitly flagged in the file's own comment as 'not yet built — next slice'.
- Higgsfield access/orchestration playbookOursExplicitly not a content recipe itself — documents 3 access surfaces and a manual brand-kit workaround since Higgsfield has no native brand kit. Simon (2026-09-24): 'We do not use Higgsfield ... We use Kie.ai, ApiMart, MuAPI and fal.ai', even though several other reference files in this range route through it.
- Brand identity build (upstream of all creative work)OursMandatory rule: never keep AI-generated logo lettering — always replace AI letterforms with licensed typography after vectorising. Literal Midjourney v6 prompt template with --no/--style raw/--sref/--sw/--cref flags; a 360° 9-angle character sheet; a 5-rule Self-Correction Protocol for generic-looking output.
- Image model comparison tester (Seedream 4.0 vs Nano Banana)RoboNuggetsA side-by-side model comparison tester (Seedream 4.0 vs Nano Banana) across 6 use cases rather than one fixed pipeline; corrects Veo3 fast Kie.ai pricing to ~30c/8s (down from ~40c cited in entry 161); notes Seedream natively supports aspect ratio/higher resolution while Nano Banana needs a secondary reframe step.
- Excalidraw diagrams via ClaudeRoboNuggetsClaude draws directly on a live vector canvas (not a generative image model) and self-checks by screenshotting its own output.
Engine shells6A bare model call with no method: they become engine choices inside the buttons
Folded in 6
- Headless local ComfyUI video renderAIOpsEntirely local/self-hosted with zero marginal cost and no data sent externally; hardware-gated (1080p needs a 24GB GPU, which the current Mac fleet can't host); an agent can drive the full install-to-render sequence from one fill-in-the-blank prompt.
- Text-to-video clip via Kling 2.6 (KIE.AI)Free onlineTemplate ships with only a placeholder prompt field ('YOUR_VIDEO_PROMPT_DESCRIPTION') — no real prompt content, submit/poll/download shell only. Same n8n template page (id 12202) as entry 67, the image-to-video variant.
- Image-to-video clip via Kling 2.6 (KIE.AI)Free onlineSame submit/poll/download shell as 66 with placeholders for both prompt and image URL ('YOUR_IMAGE_URL') — the image-to-video sibling recipe on the same n8n template page (id 12202).
- Local image-to-video with Wan 2.2 14B (ComfyUI)Free onlineSelf-hosted ComfyUI local generation, not reachable via any hosted provider; verbatim tuned settings (4 diffusion steps, split step 2, shift 5) plus a fixed Chinese negative-prompt block shared with its text-to-video companion, entry 86 — explicitly NOT the same workflow as 86 despite the shared tuning, since inputs differ (image-seeded vs text-seeded).
- Local text-to-video with Wan 2.2 14B (ComfyUI)Free onlineCompanion recipe to 85, sharing identical Lightning-LoRA tuning settings and the same fixed Chinese negative-prompt block, but text-seeded rather than image-seeded — a legitimately different operator input, not a duplicate of 85.
- Brand video library post/clip (image-to-video from library asset)OursReuses the brand's own existing video-library stills as the start frame rather than generating a fresh one, with a hard 600-character cap on the operator's 'what happens next' motion textarea; live, video library screen.
All found231
UGC video39
UGC video (Claude Skill + Seedance)Partly15-60 second vertical (9:16) iPhone-style UGC ad video, built as multiple 15-second Seedance clips stitched together with a consistent creator character, in either a bet/challenge format or a street-interview format
Seedance 2.0Nano Banana Prochallenge/bet UGCstreet interview- 1Install the Claude Skill · By hand · Claude Opus 4.7 (custom Skill) · Claude desktopDownload the 'Seedance 2.0 UGC Ad Director' Claude Skill zip. In Claude desktop: menu > Customize > Skills > + > Create Skill > Upload Skill. Requires Claude Opus 4.7. Invoke with /seeddance and confirm it announces reading the skill's director structure before trusting its output.
- 2Upload script, get segments · Script · Claude Opus 4.7 (Skill) · Claude desktopUpload the ad script to the skill. It auto-splits the script into 15-second segments (Seedance's per-clip cap) and returns Pinterest search links for a matching creator per segment.
- 3Source reference images · research · PinterestBrowse Pinterest for a clean, no-emoji, no-watermark creator photo per segment, then location/prop reference images (e.g. 'kitchen counter reference', '$100 bill').
- 4Generate the clip · Video · Seedance 2.0 · Max FusionIn Max Fusion > Videos > Seedance 2.0, upload up to 9 reference images and paste the skill's generated prompt (already @image-tagged and containing the line the creator says). For street interview: two distinct voices in one prompt (off-camera interviewer + on-camera subject), outdoor city ambience, camera 'held by interviewer arm's length'.
Prompt
iPhone handheld UGC video, @image1 ... She's standing behind a white quartz kitchen counter... Her right hand slaps 100, a crisp $100 bill with a small grin. On the counter in front of her, a rectangular piece of cardboard with four boxes cut out in a row, four plastic straws poking up through the holes.
- 5Street-interview face-swap (format-specific) · Picture · Nano Banana Pro · Max FusionFor street interview only: screenshot a real street-interview frame, place both the character reference and the background screenshot in Max Fusion Images, and prompt a face/background swap to avoid a recognizable subject.
Prompt
replace the girl on the street for the middle-aged woman, and change the background to be New York.
- 6Iterate via Remix · Video · Seedance 2.0 · Max FusionRefine using Max Fusion's 'Remix' (regenerates from edited prompt without re-uploading references): adjust shot count, add camera language ('medium shot, phone held at eye level'), swap reference images, simplify props that keep breaking consistency.
Prompt
Multiple shots, clips, UGC style. iPhone handheld. It starts with the medium shot. Phone held at eye level... A hand puts a $100 bill... on the table. Camera pans up to the creator from image one. She's standing behind the white quartz kitchen counter wearing a cream linen button-up shirt, sleeves rolled up... A rectangular piece of cardboard with four holes cut in a row... Each rolling down to a different clear wine glass. We have the boxed wine, next to it is a bottle of wine with a white label and three dollar signs on it... She looks right at the phone camera, tilts her head, and gestures with her left hand to the cardboard box and straws.
- 7Extract frame for continuity · edit · Max FusionOn the best clip, use Max Fusion's Extract Frame > Send To > Save Images to capture the last frame, then use that as the creator reference for the next segment's generation (Seedance has no native multi-clip continuity).
- 8Repeat per segment · Video · Seedance 2.0 · Max FusionRepeat steps 3-7 for each 15-second script segment, always re-attaching the extracted creator frame or consistency breaks.
Missing: the Claude Skill's full internal prompt logic is not captured (zip not machine-readable as text); the full first-pass skill-generated prompt was narrated with gaps ('blah blah') on screen, not read in full
AI Ads (Ad Creators Lab) · Claude Skill + Seedance 2.0 AI UGC / Street Interviews (Lessons 13-14)- 1
AI podcast ad (two-speaker)Partlya realistic two-person podcast-clip ad: host and guest seated across a desk, talking-head, consistent voice/accent per speaker across separately generated clips, stitched into a back-and-forth conversation
Seedance 2.0Nano Banana ProKlingpodcast- 1Source a podcast-set reference · research · PinterestSearch Pinterest for a niche-matched podcast-set photo; cross-check in-niche YouTube Shorts for what a typical set looks like. Avoid images that are themselves already AI-generated.
- 2Edit face and background · Picture · Nano Banana Pro (or GPT Image 2 as alt) · Max FusionUpload the screenshot into Max Fusion and edit the face (to avoid using a real person's likeness) and background elements, e.g. plant size, background color, so the set is not recognizable.
Prompt
change the woman's face so it's more round, and give her blonde hair and glasses
- 3Generate speaker 1's clip · Video · Seedance 2.0 · Max FusionGenerate video from the edited still, referencing it with @image1, giving the line to speak plus a detailed voice-direction prompt (accent, mic proximity effect) so voice stays consistent across separately generated clips.
Prompt
Ad Creators Lab is the best place to learn how to make high converting AI ads, hands down. Because you get insights on ads that are crushing it on meta for a seven figure agency, not just what's trending or looks pretty.
- 4Create the second speaker · Picture · Nano Banana Pro · Max FusionTake the same base image, flip horizontally, and change identity/age/ethnicity/hair to create a second character seated opposite, matched in eyeline/plane.
Prompt
flip the creator horizontally and change her to be a woman in her late 30s, brunette, no glasses... looks Latina
- 5Generate speaker 2's clip · Video · Seedance 2.0 / Kling · Max FusionGenerate the co-host's reaction/response line as its own clip, compared across Seedance 2.0 and Kling for lip/voice match quality.
Prompt
Really? What makes you say that?
- 6Stitch conversation · editEdit the separate single-speaker clips together into a back-and-forth podcast conversation in post.
Missing: the Mexican-accented voice-direction prompt for speaker 2 was described only, not quoted in full; no cost/account info
AI Ads (Ad Creators Lab) · Realistic AI Podcasts Tutorial (Lesson 6)- 1
Character-consistent B-roll setRunsa set of 5-10 second B-roll video clips (mirror talk, playing with hair, jewelry in shower, boutique shop, hot yoga) all featuring the same locked character, built from one hero reference image
Nano Banana ProKlingVeo 3.1B-roll- 1Screenshot the cleanest hero frame · By handFrom an existing generated hook video, screenshot the frame where the character faces the camera most directly/flat (angled frames caused proportion errors like an oversized forehead in later generations).
- 2Lock the character reference · Picture · Nano Banana Pro · Max FusionUpload the screenshot as reference and generate a clean, prop/hat-free selfie-style locked reference image, 3 variations to pick from.
Prompt
put this creator in her bedroom as if she is taking a selfie video. Remove the hat and tools.
- 3Generate scene variations · Picture · Nano Banana Pro · Max FusionFrom the locked reference, click Variation and describe each new scene/setting/action/wardrobe. Can feed a separate product image (e.g. jewelry from the brand site) so it appears correctly on the character.
Prompt
have her wearing two gold chains and lots of rings and bracelets, all gold.
- 4Animate each still · Video · Kling (Kling 2.6 named) / Veo 3.1 (costlier alt) · Max FusionGenerate video from each locked still. Kling used for most (fewer credits than Veo 3.1); Veo 3 avoided for B-roll as too expensive for the gain. Keep clips at 5 seconds rather than 10 (10s used more credits and visibly worse motion coherence in a side-by-side test).
Prompt
girl observes her jewelry
AI Ads (Ad Creators Lab) · Character-consistent B-roll (companion to Realistic AI Ad case study)- 1
Sora 2 Pro UGC hookRunsa single 12-second, 1080p vertical UGC-style hook clip (scroll-stopper opening line), framed as native TikTok/FaceTime/GRWM content
Sora 2 ProTikTok-style hookFaceTime-style hookget-ready-with-me- 1Write the format-framed prompt · Video · Sora 2 Pro · Max FusionName character type, setting, prop, and exact line to camera. Critical rule: explicitly state the video's native format ('shot on iPhone as if it's for TikTok', 'selfie FaceTime video', 'get ready with me video') — Sora 2 Pro is trained on organic-format categories and omitting this framing produces noticeably worse output.
Prompt
attractive Gen Z blonde girl that's wearing a yellow builder's helmet. She's outside her house in a sunny day and is holding a hammer. She looks at the camera and says, my boyfriend just spent three hundred dollars on jewelry.
- 2Optional product reference · Video · Sora 2 Pro · Max FusionUpload a background-free product image; the model incorporates it (e.g. character holding it).
- 3Remove burned-in subtitles · edit · Kapwing (inferred from unclear audio, not confirmed)Sora 2 Pro sometimes burns in unwanted on-screen subtitle text; remove with separate software.
Missing: tool used to strip burned-in subtitles is not confirmed by name
AI Ads (Ad Creators Lab) · Sora 2 Pro UGC hook- 1
Research-doc-to-script Arcads hookPartlya 12-second AI UGC hook seeded from a brand research document and a named scripting framework, generated via Arcads on Sora 2, meant as the opening of a full ad
Sora 2 (Pro)gym/locker-room POVnative TikTok- 1Pick a scripting framework · ScriptFrom the course community's Scripting Frameworks library, choose a framework matching the format (named frameworks: Industry Contrarian, Bandwagon Effect, Listicle, How-You-Connect(s)-Without-Why; skip 'Founders' unless doing a literal founder video). Download as plain text.
- 2Teach the framework to the model · Script · ChatGPT · OpenAIUpload the framework doc to ChatGPT and force it to demonstrate understanding with examples before generating real output.
Prompt
This is a script framework for an ad. Break it down in detail
- 3Generate the script · Script · ChatGPT · OpenAIUpload the research document, naming the product, product link, and target persona from the research doc; add tone/format context if relevant (podcast clip, unboxing clip, branded tone).
Prompt
Write three versions of the industry contrarian scripts for persona three... add dark humor where appropriate and use customer language. These ads are native organic TikTok style ads.
- 4Generate the hook video · Video · Sora 2 (Pro) · ArcadsTake the script's hook line into Arcads (a Sora 2 Pro wrapper), upload a reference product image, describe the on-camera character/setting and prop; Sora 2 caps at 12 seconds so only the hook is generated here.
Prompt
bodybuilder doing a TikTok POV handheld video at the gym locker room
- 5Extend into a full ad · Voice · ElevenLabs · ElevenLabsClone the generated speaker's voice in ElevenLabs, then produce the rest of the ad as B-roll + voiceover using that cloned voice. Longer-form alternative noted: Veo 3 via Fal.
Missing: the framework docs themselves (Industry Contrarian, Bandwagon Effect, Listicle, How-You-Connect(s)-Without-Why) are linked course assets, not captured verbatim here
AI Ads (Ad Creators Lab) · Creating high-quality AI UGC videos with ease- 1
Seedance 2.0/2.5 UGC ad video (structured prompt system)RunsA 15-30 second vertical UGC-style ad video of a person talking to camera in their own cloned voice, holding/using a product, produced from one structured text prompt with no manual video editing.
Seedance 2.0Seedance 2.5UGC product adtalking head / opinionpodcast clip (2+ characters)cinematic / storytellingcharacter animation- 1Gather reference assets · researchCollect a character reference image/video, a face reference, a real voice audio sample, any video-motion reference, and product prop image(s).
- 2Build the starting frame · Picture · Nano Banana Pro or GPT Image 2 · ChatGPT / Google AI Studio / HiggsfieldScreenshot a real reference photo/video frame, ask a vision LLM for a highly detailed JSON description of it, then regenerate an editable starting frame with the subject swapped.
Prompt
give me a very very detailed JSON prompt to recreate this image one-to-one. Ignore all text and buttons on the screen.
- 3Prepare the voice reference (voice-ID method) · VoiceSupply a real recorded voice clip as a reference and instruct the model to clone the voice only, not repeat the words in the clip, then supply new dialogue separately.
Prompt
use this audio as the voice ID reference for her voice, do not make her say the words in the audio, just use it as a voice ID
- 4Write the 6-block Seedance prompt · WordsAssemble the prompt in fixed order: Asset Header, Dialogue, 3-5 Style Triggers, brief Production Notes, then the timed Movement/Scene Breakdown, choosing one of five Video Type Templates (A: UGC Product Ad, B: Talking Head/Opinion, C: Podcast Clip 2+ characters, D: Cinematic/Storytelling, E: Character Animation). Hard constraints: max 2,000 characters, max 15s per Seedance 2.0 clip (30s on 2.5).
Prompt
ASSET HEADER (always first, exact template): "The character in @[character_reference_id] does a UGC VIDEO in front of the camera, eyes looking at the lens, face using @[face_reference_id] face. Make her/him say this audio @[audio_reference_id]" (Product prop line if used: "Product shown: @[prop_reference_id]". Omit the audio line if no voice reference.) DIALOGUE block: Dialogue: "[Full spoken script, conversational tone, ellipses for pauses, CAPITALIZE emphasis, ~30-45 words per 15s, up to 55 for fast delivery; label each speaker by name for multi-character clips]" STYLE TRIGGERS: stack 3-5 keywords across Format/Style, Podcast Style (e.g. "Joe Rogan Experience studio feel"), Editing Style (e.g. "SNAP CUT"), Music/Audio, Energy/Vibe. MOVEMENT & SCENE BREAKDOWN, one line per 2-4s beat: "[START]-[END] | [Beat Label]: [Description]" - specify which hand, the full motion arc, eye direction, micro-expressions, posture, prop grip/trajectory, and negatives (e.g. "does NOT pour it"). Be specific on the opening 0:00-0:03 and the closing 2-3s; the middle beats can use broader keywords. PRODUCTION NOTES (brief, e.g. "No zooms, no music, no captions", "Vertical 9:16 format"). VIDEO TYPE TEMPLATES (choose one): A. UGC Product Ad - needs starting image+face+audio+product prop; triggers "UGC style," "authentic," "talking to camera like a friend"; movement: prop retrieval -> hold up -> demonstrate -> reaction; dialogue: hook -> problem/discovery -> product intro -> personal result. B. Talking Head/Opinion - needs starting image+face+audio, minimal props; triggers "Storytime TikTok," "girlfriend energy," "direct to camera"; movement: expressive hand gestures, posture shifts, facial reactions; dialogue: direct address -> emotional reasoning -> punchline/takeaway. C. Podcast Clip (2+ characters) - needs a character image+face ref+audio ref for EACH speaker; triggers "Joe Rogan style," "Flagrant energy," "viral podcast clip editing," "808 bass synced"; movement: reacting to each other, leaning in/back; dialogue: question -> confident answer -> jealous reaction -> punchline destruction; heavy cuts, zoom punches, whip pans, music drops. D. Cinematic/Storytelling - needs character image+face ref+audio ref; triggers "Cinematic," "dramatic lighting," "slow motion," "orchestral swell"; movement: deliberate slow movements, dramatic reveals; dialogue: voiceover narration OR none. E. Character Animation - needs character image+face ref; triggers "Expressive," "animated," "full body movement," "hair physics"; movement: upper body animation, personality-driven gestures; dialogue: personality-first, quirkier scripting allowed. ASSEMBLY ORDER: Asset Header -> Dialogue -> Style Triggers (3-5) -> Production Notes (brief) -> Movement/Scene Breakdown. CHARACTER-COUNT COMPRESSION (if over 2000 chars): drop articles/pronouns, combine actions, use shorthand edit names as verbs (e.g. "SNAP CUT"). Cut first: setting description, middle movement segments, production notes, adjective stacking. NEVER cut: asset header, opening 0:00-0:03, closing 2-3s, dialogue, 2-3 key style triggers.
- 5Generate the clip · Video · Seedance 2.0 or Seedance 2.5 · Higgsfield / Dreamina / Kie.ai / fal.aiSubmit the assembled prompt with all attached references to Seedance 2.0 or 2.5.
- 6Check and iterate · CheckReview output against the prompt; if character count forced heavy compression or a beat under-delivered, adjust the movement breakdown or regenerate. "Let Seedance cook" - broad middle-beat keywords are deliberately left loose for the model to interpret.
The AI Content System · Seedance Content System Pt 1 & 2 / Seedance 2.0 Mastery / Realistic AI UGC With Seedance 2.0- 1
Clone any video with a new presenter (Nano Banana Pro + Kling Motion Control)RunsThe same source video recreated shot-for-shot with a different person, same pose, composition and motion, for testing different presenter angles/demographics on one ad concept.
Nano Banana ProKling 2.6 Motion ControlUGC video recreation- 1Source the video · research · TikTokPick a source UGC video to clone.
- 2Capture the starting frame · editScreenshot the first second/frame of the source video.
- 3Swap the person in the still · Picture · Nano Banana ProRegenerate the still with a different person, same pose and background.
Prompt
remove all buttons and text on the screen and change the person in the image to a brunette woman in her late 30s. Make it a completely different person. Keep the pose the same and the background the same. Don't change anything except for the actual person.
- 4Generate the motion-transferred video · Video · Kling 2.6 Motion Control · Kling official siteFeed the new still plus the original source video into Kling 2.6 Motion Control; it copies the original video's exact motion onto the new image.
The AI Content System · How To Clone Any Video- 1
Sora 2 AI UGC ad video (deconstruct-then-generate)PartlyA 15-second single-shot or up to 25-second Storyboard AI UGC ad video, cloned from a proven reference ad and adapted to the user's own product.
Sora 2Sora 2 ProUGC product reviewsingle-shot admulti-scene Storyboard ad- 1Source a proven reference ad · research · Meta Ad LibraryFind a performing ad on Meta Ad Library and download it.
- 2Deconstruct the ad into a Sora 2 prompt · research · Gemini 3 Pro (custom Gem) · GoogleBuild a Gemini Gem loaded with a UGC-breakdown document and the course's own "Sora 2 Prompting Guide" file, then have it deconstruct the reference video into a full Sora 2 prompt.
- 3Refine the prompt · Words · Claude · AnthropicPaste the Gemini-generated prompt into a Claude project configured to write high-quality Sora 2 prompts and ask it to refine it.
- 4Swap in the user's own product (optional) · Words · Claude · AnthropicUpload the user's product photo to the same Claude project and ask it to recreate the refined prompt using that product, changing dialogue to suit.
- 5Generate · Video · Sora 2 or Sora 2 Pro · OpenAISubmit the prompt to Sora 2 (or Sora 2 Pro) with a starting frame image and the product photo attached (portrait aspect).
Prompt
adjust this prompt and break it up into five-second parts for a full 25-second long video.
- 6Upscale (cheaper route) · Upscale · OpenAI / HiggsfieldInstead of paying for Sora 2 Pro, generate on plain Sora 2 then run through the Sora 2 Upscaler or Higgsfield.
Missing: The course's own "Sora 2 Prompting Guide" file and the Gem's UGC-breakdown knowledge document are referenced but their contents were not captured in this sweep.
The AI Content System · How To Create AI UGC Videos With Sora 2- 1
Gemini Omni AI UGC video (start-frame + voice-reference method)RunsA consistent AI UGC presenter speaking a scripted line to camera, built in up to 10-second scenes, at roughly 4c/10s - Omni's cheapest, fastest UGC route in the course.
Gemini Omni ("Omni Flash")GPT Image 2 / Nano Banana ProElevenLabsUGC talking-to-camera reviewUGC product ad- 1Source a reference video and voice · research · TikTokFind a winning UGC reference on TikTok; match the creator type/environment to the target character and pick a matching voice.
- 2Build the starting frame · Picture · ChatGPT / GPT Image 2 (or Nano Banana Pro) · ChatGPT / HiggsfieldScreenshot the reference frame, get a detailed JSON description from a vision LLM, regenerate as a new 9:16 character image.
Prompt
give me a very very detailed JSON prompt to recreate this image one-to-one. Ignore all text and buttons on the screen.
- 3Smuggle in a real voice as a black-screen video · Voice · CapCutTrim a real voice clip to just under 15s in CapCut, extract audio, delete the video track so the audio plays over a black screen, export as an mp4 (Omni can't import MP3 directly, and a visible reference face would get copied onto the output).
- 4Write the prompt as a screenshot, not typed text · Words · Google DocsWrite the generation instruction in a Google Doc, then screenshot the doc text as an image, to route around Omni's "unable to edit the speech in this video" refusal that fires on a typed instruction.
- 5Generate in Google Flow · Video · Gemini Omni ("Omni Flash") · Google Flow (labs.google) or kie.ai `gemini-omni-video` / APImart `Omni-Flash-Ext` for API useAttach all 3 ingredients (black-screen voice video, start-frame image, prompt screenshot) and give Flow only a tiny typed instruction telling it to read the screenshot and use the start frame as the starting image. Generate 4 times (one "unusual activity" error is common; failures auto-refund) and pick the best.
Prompt
read [screenshot] use [start-frame] as the starting image for the video
- 6Build longer scripts in 10s scenes · Video · Gemini Omni · Google Flow / kie.aiFor anything longer than one clip, make a new voice clip per scene and regenerate each scene fresh with the same character/start frame rather than "extending" (extending degrades quality; if extending is unavoidable, only extend from the first clip).
- 7Finish · edit · VMake / CapCutRemove captions/watermarks if needed, then stitch all scenes together.
- 8Edit an existing video (alternate use) · edit · Gemini Omni · Google FlowBeyond generating fresh, the same screenshot-prompt method can character-swap, re-colour clothing, change spoken language, or remove/add scene elements in an already-existing video by attaching the source video plus a new reference image and routing the instruction through a Google Doc screenshot as in step 4.
Prompt
replace the person in the video attached with this person attached in the image, make sure the [attribute] is exactly the same and fits
Missing: Google's official Omni API is not yet public; kie.ai and APImart are third-party aggregator wrappers. Our own 2026-07-03 test found the API will dub a supplied audio track onto a start-frame character but will NOT make a recorded voice say brand-new words (Google policy) - so a ripped-voice-plus-new-script clip needs ElevenLabs instant-clone-then-dub via kie.ai's black-screen-mp4 route, not a direct voice-swap call; the Flow-web screenshot-prompt bug-workaround is likely unnecessary on the API path where the prompt is a plain text field (unverified).
The AI Content System · How To Make AI UGC Videos With Gemini Omni / Gemini Omni AI Video Editing / Google Omni - AI UGC Video Workflow- 1
Automated end-to-end AI UGC ad factory (Claude Code skill + MCP)PartlyA fully polished, QC'd, captioned, stitched UGC ad video produced unattended by Claude Code from one starting-frame image and one script, at claimed batch scale.
Gemini Omni (via Higgsfield MCP or kie.ai MCP)UGC product ad- 1Connect a generation MCP · By hand · Higgsfield / kie.aiAdd the Higgsfield MCP (higgsfield.ai/mcp) in Claude Settings > Connectors, or the open-source kie.ai MCP/CLI (github.com/felores/kie-cli-mcp) as an alternate/additional route.
- 2Install the video-analysis skill · By hand · Claude Code · Google Gemini APIGive Claude Code the lesson-05 "Claude video analysis skill" PDF plus its Drive zip and tell it to set the skill up end to end (needs a Gemini API key from ai.studio.google.com/api-key); or skip this and use Higgsfield MCP's own native video-analysis tool instead.
Prompt
set me up this video analysis skill end to end
- 3Install the mega-prompt skill · By hand · Claude CodePaste the attached `ai-ugc-factory-setup-prompt.md` into a fresh Claude Code session; it installs a skill file at ~/.claude/skills/ugc-video-producer/SKILL.md, verifies ffmpeg/python3.11+/node20+, clones and starts a native captions server (nicolaigaina/ai-video-captions via `make setup`/`make dev`, explicitly no Docker), detects the connected MCP, then replies READY.
- 4Build the starting frame · Picture · ChatGPT / GPT Image 2 · OpenAIScreenshot a TikTok/Pinterest reference, get a detailed JSON prompt from ChatGPT, regenerate as a 9:16 image, then separately ask it to place the product in the character's hand.
Prompt
a very detailed JSON prompt to recreate this image one-to-one; remove all captions and buttons on the screen; make [her] hold the bag of [product]
- 5Run the full pipeline · Video · Gemini Omni · Higgsfield MCP or kie.ai MCPGive Claude the start-frame image and a script; the ugc-video-producer skill segments the script into 10s Gemini Omni clips (speech starts at frame zero, cuts only at sentence boundaries, ~22 words max/clip), generates each via the connected MCP, QC's every clip (warped hands, garbled label text, lip-sync drift, wrong words), auto-regenerates failed takes (max 3/clip), trims/normalizes/stitches with FFmpeg, burns in animated captions via the local captions server, runs a final QC pass, and exports.
Prompt
return to me the polished end-to-end 20-second video using this starting frame
- 6QC each generated clip · Check · Gemini video-analysis skill or MCP-native analyzerThe skill's QC pass checks every clip against the script line and label text.
Prompt
QC this AI-generated ad clip. Script line for this segment: '<line>'. Product/brand text that must appear correctly: '<exact label text>'. Check for: warped or extra fingers/limbs, garbled or misspelled on-screen or label text, lip-sync drift, face/product morphing between frames, spoken words that differ from the script line, abrupt visual glitches, dead frames. Also report the exact timestamp the FIRST word is spoken and whether the full line finishes before the clip ends. List every defect with a timestamp, or state PASS.
- 7Normalize and stitch · edit · FFmpegFFmpeg trims/scales/pads each accepted take to 1080x1920@30fps, then concatenates all segments.
Prompt
ffmpeg -y -ss <in> -to <out> -i generated/scene_NN_takeK.mp4 -vf "scale=1080:1920:force_original_aspect_ratio=decrease,pad=1080:1920:(ow-iw)/2:(oh-ih)/2,fps=30" -c:v libx264 -preset veryfast -crf 18 -c:a aac -ar 48000 -b:a 192k -avoid_negative_ts make_zero renders/seg_NN.mp4
- 8Caption · edit · faster-whisper (via ai-video-captions server) · self-hosted, localhost:5000POST the stitched video to the locally running captions server, poll status, download the captioned file.
Prompt
curl -s -X POST http://localhost:5000/api/process -F "video=@renders/vN.mp4" -F "style=hormozi" | tee captions/job.json
Missing: Requires local install of ffmpeg/python3.11+/node20+, a cloned third-party GitHub repo (nicolaigaina/ai-video-captions) run natively, and Claude Code operating with elevated/bypass permissions - recorded as the course's own instruction, not something to run unverified.
The AI Content System · Automated AI UGC Content System with Claude Code (Full Breakdown)- 1
Grok Bot research-generate-launch UGC ad systemPartlyOne-prompt AI UGC ad variants generated from a copied reference ad, then (as a separate business step) launched directly to Meta Ads.
Seedance 2.5 (via Higgsfield)UGC product ad- 1Set up the generation bot · By hand · x.aiInstall Grok Bot, create a channel and a bot, add the Higgsfield connector (required) plus optional Google Drive and X connectors.
- 2Source a reference ad · research · TikTok / Meta Ad LibraryFind a reference ad on TikTok or an ad library.
- 3Hand the bot the reference · By hand · Grok BotPaste the link and ask the bot to download it.
Prompt
download this video and let me know when you are done.
- 4Break down the reference · research · Grok BotAttach a video-analysis breakdown SOP file plus the downloaded video and ask the bot to break it down.
Prompt
read this and break down the full video.
- 5Generate the ad · Video · Seedance 2.5 · Higgsfield (via Grok Bot)Insert the user's own product into the breakdown and ask the bot to generate; it calls Higgsfield/Seedance 2.5 and produces the full ad including voice from one prompt.
Prompt
use the [Product]'s [Product] with Seedance 2.5.
- 6Generate A/B variations · Video · Seedance 2.5 · HiggsfieldAsk for multiple variations of the same ad concept using different avatars.
Prompt
generate me three more variations of this ad. / generate all three videos at once.
- 7Launch to Meta (business step, not content-making) · publish · the WAP/WAPT Meta-ads connectorIn a second "management" channel and bot, name the ads and product and ask the bot to push them to Meta Ads via the WAP/WAPT CLI connector; the bot builds an ad set and campaign and the two bots then coordinate to keep the campaign running.
Prompt
naming the three ads and the product, then asking to push them to meta ads using WAPT CLI.
Missing: The video-analysis breakdown SOP file's own contents were not captured in this sweep (Google Drive link only). The launch connector's exact product name is unclear in the source transcript (heard as "WAP," "WAPT," or "WAP CLI").
The AI Content System · Grok Bot System- 1
Claude + Higgsfield MCP reference-breakdown-then-generate (Seedance 2.5)PartlyA realistic Seedance 2.5 UGC ad, generated from a Claude-written prompt that was itself built by analyzing a downloaded TikTok/Instagram reference video.
Seedance 2.5UGC product/review video- 1Source and download a reference video · research · TikTok / InstagramSearch TikTok/Instagram for a reference concept (e.g. "UGC makeup review") and download it without a watermark.
- 2Connect Claude to Higgsfield · By hand · Higgsfield MCPInstall the Higgsfield MCP from higgsfield.ai/mcp into the Claude account.
- 3Get the breakdown and a ready prompt · research · Claude (Higgsfield video analyzer skill via MCP) · Higgsfield MCPAttach a "Seedance 2.5 video breakdown SOP file" and the source video, and ask Claude (via the Higgsfield video analyzer skill through the MCP) to analyze only the first 30 seconds.
Prompt
read the SOP file attached and analyze only the first 30 seconds of the video.
- 4Generate text-to-video first pass · Video · Seedance 2.5 · HiggsfieldCopy the returned prompt into Higgsfield with no character reference image attached (product images are fine).
- 5Build a reusable character · PictureScreenshot a frame from the first generated video and reuse it as the character reference for later generations instead of generating fresh from text each time.
- 6Upscale · Upscale · Higgsfield MCPSend the downloaded Higgsfield output back to Claude and ask it to upscale via the Higgsfield MCP's upscaling tool.
Prompt
upscale this video using the Higgsfield MCP.
Missing: The SOP breakdown file's own contents were not captured in this sweep (Google Drive link only, same link referenced in starter-10).
The AI Content System · Claude Workflow Realism- 1
GPT-6 Astra ad studio (brand research + reference breakdown + Seedance 2.5 generation)PartlyFull Seedance 2.5 ad scripts and generated videos in multiple formats (unboxing, street interview, UGC review), adapted from real reference ads to the user's own brand, researched and generated inside one agent session.
Seedance 2.5unboxingstreet interviewUGC review- 1Set up Astra and connect Higgsfield · By hand · GPT-6 Astra · Codex desktop appOpen the Codex desktop app, start a chat on GPT-6 Astra at medium effort, connect the Higgsfield plugin.
- 2Load the ad-stack skill · By hand · GPT-6 AstraInsert the "Miko Video Ad Stack" zip (from the presenter's community) into the chat and start it.
Prompt
start the video ad stack.
- 3Research the brand · research · GPT-6 AstraGive Astra the brand website and product page link; it browses the site and checks the Meta Ad Library for existing ads, returning a brief on positioning, brand voice, audience, purchase motivations, objections and ad-library findings.
Prompt
here is my product and website.
- 4Supply and analyze reference videos · research · GPT-6 Astra (FFmpeg + Higgsfield video analyzer)Find and download reference ad concepts (e.g. one unboxing, one street interview, one UGC review) and insert each; Astra combines FFmpeg frame-change inspection with the Higgsfield video analyzer for visual and audio analysis.
Prompt
here is the reference. / analyze these videos as well and let me know when you're done.
- 5Adapt to the user's brand · Words · GPT-6 AstraAsk Astra to adapt the analyzed concepts to the user's own product, keeping the same target length.
Prompt
adapt these concepts to our own perfume. We're going to keep the same target length and give me the full scripts and shot plans.
- 6Generate · Video · Seedance 2.5 · Higgsfield (via Higgsfield MCP, called from Astra)Ask Astra to generate the full end-to-end videos; it queues the Higgsfield/Seedance 2.5 jobs, tracks each, and assembles the finished clips.
Prompt
now generate the full end to end videos.
Missing: The "Miko Video Ad Stack" skill's own contents were not captured in this sweep (GitHub link only).
The AI Content System · GPT-6 Astra + Seedance 2.5- 1
Talking-head-over-B-roll UGC recreation (Sora 2 B-roll + ElevenLabs + Infinite Talk + CapCut composite)PartlyA recreated 'talking head over B-roll' UGC video (e.g. a cooking/demo video with a narrator cut into the corner), built from a deconstructed reference, an AI-generated B-roll clip, a separately generated character voice and image, a lip-synced talking clip, then manually composited.
Sora 2Nano Banana Pro/Higgsfield imageElevenLabsInfinite Talk (RunningHub / Wavespeed)talking-head-over-B-roll demo/review video- 1Deconstruct the B-roll · research · Gemini (UGC Deconstructor gem) · GoogleFeed the reference video to a Gemini UGC Deconstructor gem, isolating only the demo/food-making portion, and ask for a Sora 2 prompt.
Prompt
fully deconstruct this video, ignore the parts of the man talking and ignore the man in the corner, only focus on the food making part. / create me a full Sora 2 prompt for this video.
- 2Generate the B-roll · Video · Sora 2 · OpenAIGenerate the B-roll clip from the Sora 2 prompt.
- 3Recreate the character's starting image · Picture · any Nano Banana-style model (Higgsfield used here) · HiggsfieldScreenshot the reference character frame and get a JSON prompt to recreate it, then generate at 4K.
Prompt
give me a full JSON prompt to recreate this image of this old man holding a coconut. Ignore all text and images on the screen and buttons. Just give me a JSON prompt to recreate this exact image.
- 4Design and build the voice · Voice · Gemini (voice-prompt gem) + ElevenLabs Voice Design · Google / ElevenLabsUse a separate Gemini gem to write an ElevenLabs voice-design prompt suited to the character, then build the voice in ElevenLabs Voice Design.
Prompt
create me a high quality voice that suits this old man... make him sound happy and energetic.
- 5Generate the narration audio · Voice · ElevenLabs Text to Speech · ElevenLabsExtract the reference video's own transcript (or write a new script) and generate it as speech with the saved voice.
- 6Generate the lip-synced talking clip · Talking video · Infinite Talk Single · RunningHub.ai (or Wavespeed.ai as a simpler alternative)Attach the audio and the character image to an Infinite Talk tool and generate.
Prompt
a man talking
- 7Composite in CapCut · edit · CapCutLayer the B-roll behind the talking clip, remove B-roll audio, background-remove the talking-head layer, resize/reposition it into a corner, and nudge playback speed so dialogue timing matches the B-roll.
- 8Caption and finish · edit · CapCutAuto-generate captions (Quick template, one word at a time), clean up rows, optionally add background music.
Missing: Steps 7-8 (CapCut layering, background removal, timing, captions) are manual desktop-editor actions with no API shown.
The AI Content System · New UGC Video Breakdown (position 1)- 1
Multi-reference video clip with character consistencyRuns6-8 second video clips (incl. UGC-style handheld and lip-synced talking-head), assembled into a longer sequence, with a consistent character across shots
Seedance 2.0UGC handheldtalking-head with lip syncany format set by the reference sheet + prompt- 1Build a character reference sheet · Picture · image model of choice (illustrated/3D only) · not statedBuild a 3-angle neutral-expression character reference sheet (front, three-quarter, profile, same lighting). Real human photos as character refs are banned since Feb 2026; use illustrated or 3D refs only.
- 2Lock a seed · Video · Seedance 2.0 · Dreamina / Higgsfield / fal.ai / WaveSpeedGenerate a 4-second test clip to lock the seed for the character.
- 3Write each shot prompt · Script · manualWrite each shot prompt in fixed order: subject + action + environment + camera + style + constraints, under 200 words (under 80 for image-to-video). Always state what moves vs holds still. Append the standard quality suffix to every prompt.
Prompt
sharp clarity, natural colors, stable picture, no blur, no ghosting, no flickering
- 4Tag reference files · By hand · Seedance 2.0 upload UI/APITag every uploaded reference file with an @-role: up to 9 images, 3 video clips, 3 audio clips (first frame, character appearance, camera-movement reference, beat-synced music).
- 5Generate the clip · Video · Seedance 2.0 · Dreamina (free) / Higgsfield / fal.ai (free beta) / WaveSpeed ($0.10/s) / JimengGenerate 6-8 second clips with native audio including phoneme-level lip sync (8+ languages), re-injecting the character reference at the start of every scene to counter drift.
- 6Assemble · edit · CapCut or DaVinci ResolveAssemble clips with time-coded prompting (e.g. [0-4s] wide, [4-8s] push-in) into a longer sequence.
AIOps · Seedance 2.0: Multi-Reference AI Video Generation System- 1
Research-to-render UGC ad factoryRunsa finished ~30-second AI-presenter UGC ad, one consistent presenter across clips, from real winning-ad research through to a QA-gated render
Gemini Flash (vision, teardown + QA)Nano Banana Pro (edit variant, character/product sheets)Seedance 2.0Seedance 2.5 (manual lane)talking-to-camera product review/UGC ad- 1Scrape winning ads · research · ApifyRun the TikTok Scraper actor (clockworks/tiktok-scraper) on a niche search term, 20-30 results, returning views/likes/shares/handles/transcripts.
- 2Set up the workbench · By hand · manualCreate ugc-factory/{vault,campaigns}, install ffmpeg/yt-dlp, set .env keys (Apify, Google AI Studio, fal.ai/Higgsfield), and write a CLAUDE.md of standing orders (pipeline order, folder layout, generation laws).
- 3Teardown top ads · research · Gemini Flash (Files API) · Google AI StudioDownload top-outlier ads (top-quartile shares-to-views or 10x the creator's typical) with yt-dlp, run one teardown call per video returning strict JSON: a persuasion record (hook verbatim, hook family, awareness stage, beat map with second+job+on-screen, voice-to-visual per sentence, proof device, objection handled, CTA, one-sentence message), a capture record (device, framing, lighting, cut rhythm, sound, delivery), an evidence grade (proven/candidate/unknown), and the full transcript.
- 4Write the script · Script · not stated (LLM script-writing call) · not statedWrite from customer-complaint language first; 3.5 words/second; ~105 words for 30s; clips under 9s (4-5 per ad); three ranked hook options (callout/confession/offer) with the weakest killed and a reason; one message; a beat table with seconds+word counts; proof device named; CTA last; written the way the presenter talks; brand names respelled phonetically if misreadable.
- 5Generate character and product sheets · Picture · Nano Banana Pro (edit variant, up to 14 references) · not stated (fal.ai/Higgsfield-adjacent)Generate 3 reference images once per campaign: headshot ('iPhone-quality selfie-style', freckles as a drift anchor, minimal makeup), full body (generated WITH the headshot attached, 'the SAME woman as the reference image'), product still (label text quoted verbatim, brand name printed large). Upscale to 1400px short side.
- 6Render the hook clip · Video · Seedance 2.0 (reference-to-video) · fal.ai or Higgsfield (MCP)Render the hook clip alone first using a fixed 5-part skeleton prompt: register line, scene line, one delivery word, the exact quote after 'she says exactly:', hard rules (objects start in hand, count objects scene-wide, one action per clip, faces anchor/products drift, camera-and-sound block at the end).
- 7Human checkpoint · Check · manual reviewJudge face and voice on the hook clip; extract its audio as the voice anchor (or design the voice first in ElevenLabs).
- 8Render remaining clips · Video · Seedance 2.0 · fal.ai or HiggsfieldFire remaining clips in parallel with the same 3 references plus the voice anchor, one variable changed per regenerate loop.
- 9Manual multi-shot lane (2.5) · Video · Seedance 2.5 · Dreamina (manual paste)For multi-shot single-prompt generation with in-prompt cuts, hand off to Dreamina manually: Omni-Modal Reference mode, 9:16, 1080p always.
- 10Assemble · edit · ffmpegConcatenate with loudness normalisation.
Prompt
ffmpeg -f concat -safe 0 -i concat.txt -af loudnorm=I=-16:TP=-1.5 -c:v copy out/ad-30s.mp4
- 11Machine QA gate · Check · Gemini Flash (vision) · Google AI Studio16 evenly spaced frames checked against a fixed checklist (limbs, garments, object permanence, product-label text read per frame, identity, physics, spatial, device) returning strict JSON verdict + defects + suggested prompt fixes.
- 12Human final watch and publish · publish · manualWatch the full ad (the gate is cut-blind and only flags 'broken', never judges 'good'); tag it AI-generated on upload.
AIOps · The AI UGC Factory- 1
Extend and merge UGC video, auto-publishRunsa longer UGC-style video made by extending an existing short viral clip with an AI-generated continuation, then auto-published to social platforms
Kling 2.1 (kling-v2-1-i2v-pro)UGC video extension- 1Read row from Google Sheets · Words · Google SheetsReads the source video URL and a PROMPT/DURATION column for one row (node: 'Get video').
- 2Extract last frame · Video · ffmpeg-api/extract-frame · Fal.aiPOSTs the source video URL to fal.ai's FFmpeg extract-frame endpoint to pull the final frame as a seed image.
Prompt
REQUEST BODY (verbatim, node 'Extract last frame'): {"video_url": "{{ $json.video }}", "frame_type": "last"} - 3Generate extension clip · Video · Kling 2.1 (kling-v2-1-i2v-pro) · RunPodSends the extracted frame plus the row's prompt to Kling 2.1 (RunPod-hosted endpoint kling-v2-1-i2v-pro) to generate a continuation clip; polls via wait/status loop.
- 4Merge clips · edit · ffmpeg-api/merge-videos · Fal.aiMerges the original video with the newly generated clip into one seamless extended video via fal.ai's FFmpeg merge endpoint.
Prompt
REQUEST BODY (verbatim, node 'Merge Videos'): {"video_urls": {{ JSON.stringify($json.urls) }}, "target_fps": 24} - 5Upload to Drive · By hand · Google DriveUploads the merged video to Google Drive as storage before publishing (node 'Upload Video2').
- 6Publish to YouTube · publish · Upload-Post (api.upload-post.com/api/upload)Publishes the video to YouTube via a multipart upload with title and username fields.
- 7Publish to other platforms · publish · Postiz (api.postiz.com/public/v1/upload + n8n-nodes-postiz node)Publishes the same video to TikTok, Instagram, Facebook, X and YouTube via Postiz's native n8n node plus a separate file-upload call.
- 8Update tracking sheet · Words · Google SheetsWrites status and output URLs back to the Google Sheet (nodes 'Update video url' / 'Update video url2').
Missing: generation prompt is a user-filled Sheet column; the real template JSON confirms no example prompt text is shipped anywhere in the workflow
Free online · Extend and merge UGC viral videos using Kling 2.1, then publish on social media- 1
Unhinged POV job-hunting UGC video (GPT prompt chain + Flux + Kling + ElevenLabs + Creatomate)Runsa 5-scene, ~15-second first-person POV "day in the life" TikTok-style video about job hunting/resume building, narrated in an unhinged Andrew-Tate-x-Charlie-Sheen influencer voice, with AI voiceover, captions and a rendered final cut
o3-minigpt-4o-miniQubico/flux1-dev (PiAPI)Kling 1.6 (PiAPI)ElevenLabs TTSCreatomatePOV short-form videoTikTok/Reels- 1Load idea + environment from Sheet · Words · Google SheetsReads a topic idea and an environment_prompt column from Google Sheets for one row.
- 2Generate Image Prompts · Words · o3-mini · OpenAI (via n8n langchain node)Expands the short idea into a detailed first-person POV image prompt (foreground+background) tuned for Flux, in an unhinged 'Andrew Tate viral edge' voice, with worked examples in the prompt itself.
Prompt
You are an advanced, unhinged, hilariously entertaining prompt-generation AI specializing in expanding short POV image prompt ideas into detailed, hyper-realistic prompts optimized for Qubico/flux1-dev. Your task is to take a brief input tied to job seeking, job hunting, or resume building and morph it into a cinematic, immersive prompt locked in a first-person perspective, making the viewer feel they’re living the scene. NEVER include quotation marks or emojis in your response—flux API will choke on them, and that’s a hard no. The topic of this narrative is: {{ $('Load Google Sheet').item.json.idea }} The short prompt idea to expand for this image generation is: {{ $json.response.text }} ONLY GENERATE ONE PROMPT PER IDEA—NO COMBINING. In at least one scene, weave in this environment descriptor: {{ $('Load Google Sheet').first().json.environment_prompt }}, but go wild with unhinged, edgy, funny twists elsewhere (skip the cringe or cheesy garbage). Most job hunting happens on laptops or desktops, so prioritize those over phones. If a phone sneaks in, it’s only showing job-related content like email, LinkedIn, a resume, or a job posting—never a photo or video app. Every prompt has two parts: Foreground: Kick off with First person view POV GoPro shot of... and show the viewer’s hands, limbs, or feet locked in a job-related action. Background: Start with In the background,... and paint the scenery, blending the environment descriptor when required, plus sensory zingers. Top Rules: NO quotation marks or emojis—EVER. This is life or death for flux. Stick to first-person POV—the viewer’s in the driver’s seat, not watching from the sidelines. Show a limb (hands, feet) doing something job-focused—typing, holding a resume, adjusting a tie. Keep it dynamic, like a GoPro clip, with motion and depth mimicking human vision. If tech’s involved (phone, computer), it’s displaying job-hunting gold—email, job boards, resumes—not random trash. No off-topic actions like recording videos or snapping pics—job hunting only, fam. Extra Vibes: Full-body awareness: Drop hints of physical feels—cramping fingers, racing pulse, slumping shoulders. Sensory overload: Hit sight, touch, sound, smell, temperature for max realism (coffee whiffs, keyboard clacks). World grip: Limbs interact with the scene—tapping keys, handing over papers, stepping up. Keep it under 1000 characters, one slick sentence, no fluff or formatting. Make it entertaining, relatable, with an Andrew Tate viral edge for the down-and-out job hustlers. Examples: Input: Updating a LinkedIn profile after a long day Environment_prompt: Tired, cluttered apartment, laptop glow Output: First person view POV GoPro shot of my hands hammering a laptop, cheeto-dusted fingers aching from the grind, the screen flashing my LinkedIn profile with a fresh connection ping; in the background, a trashed apartment lit by the laptop’s ghostly glow, pizza boxes toppling, traffic humming outside, stale takeout stench hitting my nose as my back screams from the hustle. Input: Handing over a resume at a job fair Environment_prompt: Hopeful, busy convention hall, suits everywhere Output: First person view POV GoPro shot of my hand thrusting out a crisp resume, fingers twitching with nerves as it brushes another palm; in the background, a buzzing convention hall packed with suits, coffee fumes and shoe polish in the air, chatter drowning my pounding heart as I lock eyes with the recruiter. NO QUOTATION MARKS. NO EMOJIS. EVER. - 3Generate Image · Picture · Qubico/flux1-dev · PiAPIGenerates a 540x960 vertical still from the expanded prompt via PiAPI's Qubico/flux1-dev txt2img endpoint, appended with an 'iPhone camera, TikTok influencer' realism suffix and a negative prompt against room/video-recording imagery.
Prompt
REQUEST BODY (verbatim, node 'Generate Image'): prompt = "{{ Generate Image Prompts output }} realistic and casual as if taken by an iphone camera by a TikTok influencer"; negative_prompt = "taking a photo of a room, recording a video of a room, photos app, video recorder, illegible text, blurry text, low quality text, DSLR, unnatural" - 4Image-to-Video · Video · Kling 1.6 (PiAPI, mode: pro) · PiAPIAnimates the still into a 5-second clip via PiAPI's Kling endpoint, reusing the same prompt text, with a zoom camera control.
Prompt
negative_prompt = "blurry motion, distorted faces, unnatural lighting, over produced, bad quality"; prompt reused verbatim from the image generation step
- 5Generate Video Captions · Script · gpt-4o-mini · OpenAI (via n8n langchain node)Writes 5 short TikTok captions (Problem>Action>Reward) narrating the job-hunt story in the same unhinged Tate/Sheen persona, hook first, reward last.
Prompt
SYSTEM: DO NOT include any quotation marks in your response. Do not put a quote at the beginning or the end of your response. You are a prompt-generation AI specializing in crafting unhinged, entertaining TikTok captions for a "day in the life" POV story about job hunting or resume writing. Generate five concise, action-driven captions (5-10 words each) that follow a Problem > Action > Reward structure. The first caption should be a shocking or funny hook, and the last should conclude with a satisfying reward. Use emojis sparingly—only one per caption at most, and only when they add impact; skip them if they don’t enhance the message. Guidelines: Perspective: Always first-person POV, immersing the viewer in the story. Tone: Channel Andrew Tate mixed with Charlie Sheen—cursing and sexual innuendos are fair game. Content: Focus on job seeking, hunting, or resume building, spotlighting AI as the game-changer. Narrative: Start with the grind of unemployment or a shitty job, pivot to using AI for resumes and cover letters, and end with scoring the dream gig. Scenes: Highlight raw, emotional moments—skip the boring stuff. Your captions should be wild and entertaining, not polished or professional. The first caption is the hook—make it shocking, hilarious, or ballsy, something Andrew Tate would growl. Use emojis sparingly—max one per caption, only if it hits harder with it. Your response should be a list of 5 items separated by "\n" (for example: "item1\nitem2\nitem3\nitem4\nitem5") USER: ={{ $json.idea }} - 6Generate Script · Script · gpt-4o-mini · OpenAI (via n8n langchain node)Turns the 5 captions into a ~15-second (5x~5s) spoken narration script in the same unhinged influencer voice, no emojis (TTS-bound).
Prompt
SYSTEM: You are an unhinged and hilarious TikTok influencer who's like a mix of Andrew Tate and Charlie Sheen. The user is going to provide you with a topic, and then 5 different parts of a story. Your task is to narrate the story as this hilarious character, who isn't afraid to be edgy or curse or use sexual innuendos. However keep each of the 5 talking points brief, as you only have about 5 seconds to speak during each. The entire length of your narration should be around 15 seconds. Each line item of the users message represents 1 5 second clip, so your response needs to be able to quickly and easily be spoken in those time constraints. Don't say extra things you don't need to. Just quickly tell the story, in order, and make it unhinged, funny, entertaining, and potentially controversially viral. Don't worry about offendeding anyone. Andrew Tate style it. Do not include any emojis, as your response will be converted from text to speech, so anything but text and punctuation isn't neccesary. Also, don't make your jokes overly corny, speak in a witty, edgy, funny way, but no corny dad jokes or anything cringe. USER: ={{ Generate Video Captions output }} - 7Generate voice · Voice · ElevenLabs TTS (voice id onwK4e9ZLuTAKqWW03F9) · ElevenLabsConverts the narration script to speech via ElevenLabs TTS on a fixed voice id.
- 8Render Final Video · edit · CreatomateAssembles 5 video clips + voice audio + 5 scene-title text overlays into the final rendered video via a Creatomate template.
Missing: none — all generating prompts confirmed verbatim from the real template JSON (api.n8n.io/api/templates/workflows/3121), corrected from the first pass which wrongly showed prompt:null for every step
Free online · Generate 100 viral UGC videos automatically and publish to social media- 1
Unhinged POV job-hunting UGC video, auto-published to 5 platformsRunsthe same 5-scene unhinged POV job-hunting UGC video as fw-n8n-100-job-hunt-ugc-piapi-elevenlabs-creatomate, plus an AI-generated social caption and automatic publishing to TikTok, Instagram, YouTube, Facebook and LinkedIn
o3-minigpt-4o-minigpt-4oQubico/flux1-dev (PiAPI)Kling 1.6 (PiAPI)ElevenLabs TTSCreatomatePOV short-form videoTikTok/Reels/YouTube/Facebook/LinkedIn- 1Steps 1-8: identical core pipeline · Video · o3-mini / gpt-4o-mini / Qubico/flux1-dev / Kling 1.6 / ElevenLabs TTS / Creatomate · PiAPI, ElevenLabs, CreatomateRuns the exact same idea->prompt->image->video->captions->script->voice->render chain as fw-n8n-100-job-hunt-ugc-piapi-elevenlabs-creatomate; node parameters are byte-identical (confirmed by direct comparison of the raw JSON).
Prompt
IDENTICAL to entry fw-n8n-100-job-hunt-ugc-piapi-elevenlabs-creatomate steps 2, 5 and 6 (Generate Image Prompts / Generate Video Captions / Generate Script) — see that entry for full verbatim text; not repeated here to avoid duplication.
- 9Get Audio from Video / transcribe · Words · n8nExtracts audio from the finished render and transcribes it (used as the caption-writing input).
- 10Generate Description for Videos in TikTok and Instagram · Words · gpt-4o · OpenAI (via n8n langchain node)Writes a social caption/description for the video, styled on three worked example Instagram captions provided in the prompt, based on the transcribed audio.
Prompt
SYSTEM: You are an expert assistant in creating engaging social media video titles. USER: I'm going to upload a video to social media. Here are some examples of descriptions that have worked well on Instagram: Follow and save for later. Discover InfluencersDe, the AI tool that automates TikTok creation and publishing to drive traffic to your website. Perfect for entrepreneurs and brands. #digitalmarketing #ugc #tiktok #ai #influencersde #contentcreation Discover the video marketing revolution with InfluencersDe! . . . #socialmedia #videomarketing #ai #tiktok #influencersde #growthhacking Don't miss InfluencersDe, the tool that transforms your marketing strategy with just one click! . . . #ugc #ai #tiktok #digitalmarketing #influencersde #branding Can you create another title for the Instagram post based on this recognized audio from the video? Audio: {{ Get Audio from Video text }} IMPORTANT: Reply only with the description, don't add anything else. - 11Upload + publish to 5 platforms · publish · Upload-Post (api.upload-post.com/api/upload)Uploads the video with the generated caption/title to TikTok, Instagram, YouTube (title truncated to 70 chars), Facebook (to a fixed page id) and LinkedIn via Upload-Post's multipart API.
Missing: none for the new nodes; core pipeline gaps are none (see paired entry)
Free online · Generate 100 viral UGC videos automatically and publish to social media (multi-platform publish variant)- 1
Full UGC ad pipeline: character image to Seedance 2.0 videoPartlyA short (5-10s) UGC-style ad video of an AI-generated 'creator' character talking to camera, showcasing or reacting to a product (testimonial, product showcase, unboxing, before/after, or lifestyle style), automated end-to-end through the Higgsfield web app via browser automation.
Higgsfield Soul 2.0 (character image)Seedance 2.0 (image-to-video)testimonial/talking-headproduct showcasereaction/unboxingbefore/afterlifestyle/day-in-the-life- 1Write the character image prompt · WordsGenerates a production-grade photorealistic character prompt for a UGC-style creator, built in a fixed layer order (subject description -> age & features -> expression & pose -> clothing & style -> lighting -> camera & framing -> background -> technical quality tags), explicitly avoiding words like 'perfect'/'flawless'/'8K' that push results toward an uncanny-valley AI look.
Prompt
Beautiful young woman, 23 years old, warm brown eyes, light brown hair with loose beach waves past shoulders. Natural skin texture with light sun-kissed glow, subtle freckles across nose. Mid-conversation expression looking directly at camera, genuine warm smile, slightly raised eyebrows. Wearing a casual white ribbed tank top, small gold hoop earrings, minimal natural makeup. Soft golden hour sunlight from the right side, warm tones. Close-up portrait, head and shoulders, shallow depth of field f/1.8, slightly above eye level selfie angle. Clean bright apartment background with soft bokeh. Photorealistic, 4K, ultra detailed skin texture, natural color grading, shot on Sony A7III
- 2Generate the character image on Higgsfield Soul 2.0 · Picture · Higgsfield Soul 2.0 · Higgsfield.aiPlaywright automation navigates to higgsfield.ai/image/soul-v2, types the character prompt into the image-prompt textbox, asks the user to confirm (uses credits), clicks Generate, waits ~20s, and verifies the new image is the first item in the History grid.
Prompt
<character image prompt>
- 3Load the image into the Seedance 2.0 video page · Video · Higgsfield.aiNavigates to higgsfield.ai/create/video?model=seedance_2_0, opens the upload-media dialog, switches to its 'Image Generations' tab, and selects the most recently generated image (always the first 'Check eligibility' item in the grid), then closes the dialog.
- 4Write and enter the UGC video prompt · Video · Seedance 2.0 · Higgsfield.aiClears the Lexical rich-text prompt editor (a contenteditable div, cleared via Control+A/Backspace, not a plain textbox) and types one of five style-specific motion prompt templates referencing the source image as @image1, describing action, expression, lighting and camera feel appropriate to the chosen ad style.
Prompt
UGC-STYLE TEMPLATE (used when the orchestrator's own 5 UGC styles are chosen): "The woman from @image1 holds up a product toward the camera, turning it to show different angles. Excited genuine expression, eyes wide. Ring light illumination, clean background. She points at the product features, nodding enthusiastically. Close-up selfie framing, iPhone-quality video, authentic UGC feel." (one of five UGC style templates in ugc-video-auto/SKILL.md: Testimonial/Talking Head, Product Showcase shown here, Reaction/Unboxing, Before/After, Lifestyle/Day-in-the-Life) The SAME automation can instead call any of 15 independent, much larger style sub-skills (seedance-auto-generate/SKILL.md lists all 15 by slash-command), each producing full production-grade Seedance 2.0 prompts with its own 2-Second Hook Framework, camera/lighting/sound-design guidance and 5 worked examples. Two verbatim worked examples pulled from those files: (07-ecommerce-ad, "Luxury Fashion Outfit Showcase", 25s): "Generate a 25-second e-commerce fashion advertisement video for a high-end designer jumpsuit using Seedance 2.0 on Higgsfield... 2-Second Hook (0-2s): Black screen. Text appears 'New Season. New You.' Simultaneously, a hand reaches into frame from the left, holding the folded jumpsuit in premium pearl-white fabric... Product Showcase (2-10s): The jumpsuit unfolds across the frame. Seedance 2.0 on Higgsfield generates smooth 360-degree rotation showing: Front facing / Side angle / Back detail / Close-up macro shot of beading... Product Shot with Text (10-12s): text overlay 'Italian Silk Blend' and price '$449'. Lifestyle Integration (12-22s): woman wearing the jumpsuit at an upscale evening event, champagne glass in hand, confident expression, camera captures full body/waist detail/fabric texture... text overlay 'Designed for You. Made for Tonight.' Final CTA (22-25s): 'SHOP NOW' with pulsing glow, 'Exclusive Online', 'Free Shipping Worldwide'. Technical Specs: 9:16, warm golden lighting on lifestyle / cool studio white on product shots, 120 BPM upbeat trending audio, smooth rotations/slow walks/subtle zoom-ins." (11-social-hook, "Satisfying Transformation Hook", 30s): "Create a 30-second satisfying transformation video optimized for TikTok with a hook that stops the infinite scroll. HOOK STRUCTURE (First 2 seconds): [0.0-0.3s] Open on extreme close-up (macro lens) of GRIMY, STAINED SURFACE. Viewer cannot recognize what surface is... Sound: ASMR scratching sound with slight tension music underneath. [0.3-0.8s] Zoom out slightly... reveal surface is a previously-white shoe. On-screen text at 0.5s: 'THIS IS DISGUSTING'... [0.8-1.5s] Quick cut to person dipping shoe brush into cleaning solution... text at 1.2s: 'WATCH WHAT HAPPENS'... [1.5-2.0s] scrubbing in fast-motion (2x speed), paint-like transformation brown/yellow disappearing... reveal shoe is now BRILLIANTLY WHITE... FULL VIDEO (2-30s): continued scrubbing multi-angle (2-10s), final rinse in slow-motion (10-15s), before/after side-by-side (15-25s), final shot wearing the restored shoe with end text 'SAVE THIS FOR LATER' or 'COMMENT IF YOU NEED THIS' (25-30s). TECHNICAL SPECS: 9:16 vertical, original ASMR sounds + music bed, music starts 0.3s/builds through 2s/peaks 1.5s, two text overlays in the first 2 seconds."
- 5Generate and monitor the video · Video · Seedance 2.0 · Higgsfield.aiAsks the user for confirmation (video costs more credits than images), clicks Generate, then periodically checks the History tab; Seedance 2.0 videos typically take 1-3 minutes.
Missing: The generation itself runs on Higgsfield's own paid web app (Soul 2.0 image credits + Seedance 2.0 video credits), driven by Playwright browser automation against a logged-in session - not a direct API call, so it needs a live browser session and a Higgsfield account, not just an API key.
Free online · GitHub — AKCodez/higgsfield-claude-skills: all 19 files enumerated via the trees API; read in full: 'ugc-video-auto' (UGC orchestrator), 'ugc-hot-girl' (character prompt), 'higgsfield-image-auto' + 'seedance-auto-generate' (the two browser-automation halves), plus the '07-ecommerce-ad' and '11-social-hook' style sub-skills for worked examples; the remaining 13 of 15 style sub-skills (01-cinematic, 02-3d-cgi, 03-cartoon, 04-comic-to-video, 05-fight-scenes, 06-motion-design-ad, 08-anime-action, 09-product-360, 10-music-video, 12-brand-story, 13-fashion-lookbook, 14-food-beverage, 15-real-estate) were fetched and confirmed to share the same structure (2-Second Hook Framework, camera/lighting/sound-design guidance, 5 worked examples each) but not fully transcribed here- 1
UGC talking/product video ad from a single prompt chainRunsa UGC-style talking/product video ad built from a still image, animated with natural handheld motion, and paired with an ElevenLabs voiceover script
Nano Banana Pro / Flux Pro / Midjourney (image)Kling AI / Runway (image-to-video)ElevenLabs (voice)ChatGPT/Claude (script)UGC selfie/product review- 1Generate the static UGC-style image · Picture · Nano Banana Pro / Flux Pro / MidjourneyGenerate a still image following UGC authenticity principles (imperfection, authenticity, emotion-first) — e.g. a selfie-style shot of a person holding the product.
Prompt
iPhone selfie of a [age] [gender] holding [product] next to their face with a satisfied smile, bathroom mirror reflection, morning routine context, authentic and unposed.
- 2Animate the image into a UGC-style video · Video · Kling AI / RunwayAnimate the still into a short video using a motion template that deliberately keeps it looking like casual, unprofessional smartphone footage.
Prompt
Template: "A person naturally [action] in a [setting], casual smartphone video quality, slight camera shake, natural movements, UGC style, not professional footage." Worked example: "Person's hands slowly lifting a product out of a box, turning it to show the label, natural room lighting, slight camera shake like someone is filming with one hand, casual and authentic."
- 3Write the script and generate voiceover · Voice · ChatGPT/Claude (script) + ElevenLabs (voice) · ElevenLabsUse ChatGPT/Claude with a 15-second UGC script framework (hook, benefit + personal anecdote, natural CTA) to write the script, then generate the voiceover with ElevenLabs using a conversational preset.
Prompt
Write a 15-second UGC-style ad script for [product] in a casual, conversational tone. Include a strong hook in the first sentence, mention one specific benefit with a personal anecdote, and end with a natural-sounding call to action.
Missing: exact ElevenLabs voice/model settings not specified beyond 'conversational preset'
Free online · AI Prompting Guide for UGC Content Creators (2026)- 1
Single-shot text-to-video (Kie.ai Seedance) — DEAD CODEPartlyUGC video generated directly from a script, bypassing the engine picker.
bytedance/seedance-2-fastbytedance/seedance-2-pro- 1Propose · VideoproposeVideo → consumer/video.mjs handlePropose builds a video prompt dynamically.
- 2Generate (bug) · Video · bytedance/seedance-2-fast / -pro · Kie.aihandleGenerate bypasses resolveProvider entirely and hardcodes kieSubmit/kiePoll against Kie.ai Seedance, regardless of which engine the operator picked.
Missing: Built but not wired to any button in the UI (VideoHero/VideoApproveBar imported but never rendered); even if wired it has a real bug — it ignores the engine picker and always hardcodes Kie.ai.
Ours ·- 1
Scripted vertical short (Short studio)PartlyA scripted vertical video short: 4-5 beats, 60-130 words, ≥3 scene hints, in a stylistic preset that can include claymation/pixar-3d/stop-motion.
claymationpixar-3dstop-motion- 1produce_short_scripted · VideomakeTheVideo() dispatches verb produce_short_scripted; short.ts validates the script (4-5 beats, 60-130 words, ≥3 scene hints, style list).
Missing: Provider not named in this repo.
Ours ·- 1
Shorts full lifecycle (script→render→voice/caption/remix→crosspost)PartlyA scripted short with its full post-production lifecycle: render, remix, voice swap, caption fix, crosspost, and viral-harvest.
- 1Render · VideohandleRender() → renderShort() verb produce_short_scripted, ~$2.80 gate, validates beats/words/scene hints.
- 2Remix / voice swap / caption fix · editremixShort(), swapVoice(), fixCaptions() — post-production fixes on the rendered short.
- 3Crosspost / harvest · publishcrosspostShort() publishes; harvestViral() pulls performance signal back in.
Missing: No provider named anywhere in this repo for the render step.
Ours ·- 1
Sora 2 Two-Step Prompt Pipeline (Gemini deconstruct → Claude refine)PartlyA UGC/image-to-video Sora 2 clip, ≤12-15s, built by deconstructing a real reference video into a prompt then refining it.
Gemini ProClaudeSora 2- 1Deconstruct · Video · Gemini Pro · GoogleA Gemini Gem (built from sora2-gemini-gem-instructions.md), Pro mode, 'deconstruct this video' on an uploaded reference — returns a script transcription, authenticity cues, and a literal Sora-optimized UGC prompt template.
- 2Refine · Words · Claude · AnthropicPaste the Gemini output into a Claude Project configured with prompting-guide.md rules, 'refine this prompt'.
- 3Generate · Video · Sora 2 · Sora.com / Key / HiggsfieldGenerate via Sora.com, Key (no-watermark API platform), or Higgsfield (≤12s).
Missing: No auth/pricing given for any of the three generation platforms.
Ours ·- 1
Veo 3.1 Frame-to-Video UGC Pipeline (Snubroot + Google Flow)ManualChainable 8-second UGC clips built from a still start frame, animated via Google Flow, chained by reusing the last frame as the next clip's first frame.
Nano Banana ProVeo 3.1 Fast- 1Load Snubroot framework · By handLoad a Claude Project with the external Snubroot Veo-3 Meta Framework (github.com/snubroot/Veo-3-Meta-Framework), timestamp-prompting section manually deleted.
- 2Starting frame · Picture · Nano Banana ProGenerate the persona + product starting frame.
- 3Build Veo prompt · Words · ClaudeClaude generates the Veo prompt dynamically using the loaded framework.
- 4Generate + chain · Video · Veo 3.1 Fast · Google FlowGoogle Flow, Frames-to-Video mode, Veo 3.1 Fast, 4 outputs, pick best of 4, chain by screenshotting the last frame as the next clip's first frame, repeat. Text-to-Video banned for ads; Ingredients-to-Video only for unboxing.
- 5Stitch · edit · CapCut / PremiereStitch chained clips.
Missing: Depends on the un-reproduced Snubroot framework content — not runnable on its own from these files.
Ours ·- 1
Seedance 2.0 — Cockpit field-template ('Michael's method')PartlyA field-by-field A-roll/B-roll Seedance prompt (Reference/Subject/Setting/Wardrobe/Product/Camera rig/Action/Angle/Motion/Dialogue/Tone/Realism), up to 15s with Mode-B chaining for longer.
bytedance/seedance-2-fast- 1Fill the field template · Video · bytedance/seedance-2-fast · Kie.aiGlobal specs block + per-field A-roll template + B-roll template + up to 9 tagged reference images (@image1-4+) + a literal QC-negatives block, 15s cap with Mode-B chaining for longer.
Prompt
no cuts, no montage, no captions... no invented packaging (QC-negatives block, literal in file)
Missing: Complete for prompt construction, partly for execution mechanics.
Ours ·- 1
UGC Ads Full Production (Higgsfield-centered)RunsA multi-scene UGC ad: an influencer holding a product, animated via Veo 3.1 in 4s chunks reusing the same reference image across scenes.
Nano Banana ProChatGPT Image 1.5Veo 3.1- 1Base influencer image · Picture · Nano Banana Pro / ChatGPT Image 1.5Generate a base influencer image.
- 2Product integration · Picture · Nano Banana Pro / ChatGPT Image 1.5Add the product into the influencer's hand without changing the face.
Prompt
Keep the same influencer... Add a realistic [PRODUCT]... Do not change the face.
- 3Animate · Video · Veo 3.1 · HiggsfieldVeo 3.1 animate via Higgsfield Frames mode, 4s, 1080p, one full worked Coca-Cola example given; iterate checklist; reuse the same reference image across multiple scenes.
- 4Stitch · editStitch multiple scenes.
Ours ·- 1
AI Persona Ad Pipeline (ICP → Face-Lock → Start Frame → Video)RunsA UGC persona ad: an ICP-driven persona locked by a JSON face-preservation protocol, given a mandatory action-transfer start frame, then animated and voiced.
Nano Banana ProVeo 3.1Sora 2MiniMax- 1Base generation · Picture · Nano Banana ProNano Banana Pro base generation from a JSON prompt, with a mandatory imperfection block and matching negative prompt.
Prompt
visible pores, fine lines, natural skin texture, realistic subsurface scattering
- 2Face-Lock Protocol · Picture · Nano Banana ProInject a literal JSON face-preservation block into every subsequent generation of this persona.
Prompt
{"face_preservation": {"preserve_original": true, "accuracy_level": "100% identical to reference"}} - 3Start Frame via Action Transfer · Picture · Nano Banana ProMandatory Start Frame built from persona + action reference + product-in-hand (Action Transfer method).
- 4Animate · Video · Veo 3.1 / Sora 2Veo 3.1 or Sora 2 animate, chained via last-frame reuse.
- 5Voice + finish · Voice · MiniMaxMiniMax voiceover, Adobe Podcast Enhance, CapCut edit, Topaz upscale.
Ours ·- 1
Meta 'Native & Stealth' Format Library (4 sub-formats)PartlyFour distinct video sub-formats: AI podcast clip, AI street interview, UGC testimonial rant, and pseudo-documentary/news segment.
Nano Banana ProInfiniteTalkMiniMaxElevenLabsVeo 3.1 FastKling 3.0 Omnipodcast clipstreet interviewtestimonial rantpseudo-documentary/news- 1A) AI Podcast Clip · Video · Nano Banana Pro + InfiniteTalk/MiniMax/ElevenLabsNano Banana Pro still + InfiniteTalk/MiniMax/ElevenLabs lipsync.
- 2B) AI Street Interview · Video · Nano Banana Pro + Veo 3.1 FastNano Banana Pro + Veo 3.1 Fast POV.
- 3C) UGC Testimonial Rant · Video · Nano Banana Pro + Veo 3.1 FastNano Banana Pro 'messy room' scene + Veo 3.1 Fast chained 8s chunks.
- 4E) Pseudo-Documentary/News Segment · Video · Kling 3.0 Omni · Kie.aiSame method as Kling 3.0 Omni Multi-Shot (see our-kling3-omni-multishot).
Missing: Each format's tool chain is named, no filled worked example per format.
Ours ·- 1
AI-Influencer network — Pipeline B trend-driven video (fully executable)RunsA trend-driven UGC video for a locked AI-influencer persona: scripted, evaluated, voiced, lip-synced, B-roll added, and published — end to end with literal API bodies.
ClaudeClaude Haiku (evaluator)ElevenLabsWaveSpeed InfiniteTalkWaveSpeed Kling 2.1 Propersonal_storyindustry_insightbuild_in_public- 1Route content type · WordsContent-type routing (personal_story/industry_insight/build_in_public) against a Hook Formula Library table.
- 2Generate script · Script · ClaudeLiteral script-generation prompt template.
- 3Evaluate + retry · Check · Claude HaikuEvaluator-Optimizer loop with a literal evaluator prompt, threshold >=7 avg / no criterion <=4, max 2 retries.
- 4Voice · Voice · ElevenLabs · ElevenLabs APIElevenLabs API call with a literal request body. Explicit rule: do NOT generate an AI image for the persona — use Face Seed Direct.
- 5Lipsync · Talking video · InfiniteTalk · WaveSpeed APIWaveSpeed InfiniteTalk API call, literal, $0.15/5s, up to 10min.
- 6B-roll · Video · Kling 2.1 Pro · WaveSpeed APIWaveSpeed Kling 2.1 Pro API call, literal, for B-roll.
- 7Post-process · editLiteral FFmpeg post-processing: 9:16/1080p/30fps, strip metadata.
- 8Approve + publish · publish · Telegram, Late APITelegram approval-button template, then Late API publish with mandatory ai_generated: True.
Missing: HEDRA_API_KEY flagged as missing from .env if adapted to use Hedra. Static-image and carousel variants of the same publish pipeline exist but without literal per-image prompts.
Ours ·- 1
Medical/Professional Persona UGC (Pharmacist/Doctor authority)RunsA UGC video of a professional-authority persona (pharmacist/doctor) speaking a script line-by-line, generated manually per line (full automation flagged 'Coming Soon').
ClaudeGeminiNano Banana ProVeo 3- 1Describe reference · Picture · ClaudeAsk Claude for a hyper-detailed JSON prompt describing a stock reference image, with optional descriptor overrides.
Prompt
Give me a hyper-detailed JSON prompt for this image.
- 2Generate persona · Picture · Nano Banana Pro · Google Flow / HiggsfieldGenerate from the JSON via Gemini/Nano Banana Pro (or Flow/Higgsfield).
- 3Script · Script · ClaudeClaude writes the script, fed past winning scripts as context.
- 4Animate line-by-line · Video · Veo 3Veo 3 Frames-to-Video, literal formula 'Make him say: [line]', mandatory 'No background music'; physical-action variant also given.
Prompt
Make him say: [line]. No background music.
- 5Stitch · edit · CapCutCapCut stitch of all line clips.
Ours ·- 1
Infinite UGC video (custom character + product)Partlyone multi-scene UGC-style ad video of any requested total length (built from chained 8-second Veo3 clips stitched together), featuring a consistent AI-rendered character promoting a product sent in as a photo via Telegram.
GPT-image (OpenAI 4o image model)Veo3 FastVeo3 QualityGPT-4.1FFmpegUGC vertical video, length is a multiple of 8s up to the user's requested duration- 1Receive trigger photo · By hand · TelegramUser sends one photo (product and/or reference character composited together) to a Telegram bot along with text instructions (what to promote, character description, desired total video length).
- 2Fetch image file · Words · TelegramHTTP GET to Telegram resolves the bot ID and file ID into a downloadable file path for the sent photo (mid resolution, 'photo two').
- 3Describe reference image · research · GPT-4o (vision, 'analyze image' operation) · OpenAIAn OpenAI 'analyze image' node describes the product (brand name, colors, visuals) and/or character (name, outfit colors, visual description) present in the photo.
- 4Acknowledge to user · By hand · TelegramBot replies on Telegram: 'Got it. I'm now creating your video.'
- 5Craft reference-image prompt · Script · GPT-4.1 · OpenAIAn image AI agent (with a think tool) writes one image prompt depicting the reference as accurately as possible, in an amateur/iPhone-photo realism style, output in YAML with a locked aspect ratio (3:2 or 2:3), autofix-format enabled.
- 6Generate reference image · Picture · GPT-image (OpenAI 4o image gen) · Kie.aiHTTP POST sends the image prompt and the original photo URL to the 4o image model via Kie.ai to generate a single start-frame reference image (e.g. character holding the product).
- 7Wait/check image · Check · Kie.aiWaits 60s and polls until the reference image is ready (success flag = 1).
- 8Craft scene/video prompts · Script · GPT-4.1 · OpenAIA video AI agent (with a think tool) writes N scene prompts (N = ceil(requested total seconds / 8), default 3 if unspecified) with continuous dialogue across scenes, the brand name mentioned only in the first scene, the character never physically touching the product, and no copyrighted-character depictions (Veo3 will reject those).
- 9Split scenes · edit · n8nn8n Split Out node turns the single item containing N scenes into N separate items so each can be sent to Veo3 individually.
- 10Generate video clips · Video · Veo3 Fast or Veo3 Quality · Kie.aiFor each scene, HTTP POST sends the scene's prompt, model, aspect ratio and the reference image (as the image-to-video start frame) to Veo3 via Kie.ai, batched ~3 seconds apart.
- 11Wait for all clips · Check · Kie.aiWaits a fixed 600 seconds (no per-item retry loop, to avoid partial-group merge errors), then fetches all scene results by task ID.
- 12Aggregate clip URLs · edit · n8nn8n Aggregate node collects the N separate clip-result items back into one item containing all N video URLs, in order.
- 13Merge clips · edit · FFmpeg (merge videos) · fal.ai (transcript says "file.ai" again here, same ambiguity as n6/n9)HTTP POST to an FFmpeg 'merge videos' function stitches the N clips together in sequence into one continuous video.
- 14Wait/check merge · Check · fal.aiWaits then polls until the merged video URL is ready.
- 15Deliver video · publish · TelegramSends the finished merged UGC video back to the user on Telegram.
Missing: image-agent and video-agent system prompts sit in an unharvested community template file; the merge-step provider name is ambiguous in the transcript ('file.ai'); no true last-frame-to-next-clip scene continuity yet (each scene is generated independently from the same single reference image, not chained), per source's own stated limitation.
RoboNuggets · n15 - Endless UGCs, any length, any character- 1
UGC ad video autopilot (Nano Banana + Veo3)Runsshort UGC-style ad video
Nano BananaVeo3 fast8s clips- 1Generate first frame · generate-image · Nano Banana · fal.aiNano Banana image-to-image generates a UGC-style first frame from a reference/product image
- 2Animate into video · generate-video · Veo3 fast · Kie.aiVeo3 fast animates the generated first frame into an 8s video clip
- 3Combine clips · merge-videoCombine multiple generated clips into one ad video
- 4Deliver · deliver · TelegramSend finished video back via Telegram bot
Missing: node-level HTTP body params only described via diff against an earlier ChatGPT-4o-image version, not fully quoted
RoboNuggets · n18 - Nanobanana Autopilot UGCs- 1
Infinite UGC video agent (Sheet-driven, multi-model)Runsmulti-scene UGC video ads
Seedream 4.0Nano BananaChatGPT-imageVeo3 fastmulti-scene, any aspect ratio per sheet input- 1Collect inputs · input · Google SheetsGoogle Sheet row provides character, setting, voice, script, reference image inputs
- 2Craft prompts · agent · OpenAI (agent) · OpenAIOpenAI AI agent (with think tool, structured-output parser, and sheet-dedup tool) crafts a scene-0 image prompt plus per-scene video prompts
- 3Generate scene-0 image · generate-image · Seedream 4.0 / Nano Banana / ChatGPT-image · Kie.aiSwitch node routes scene-0 image generation to Seedream, Nano Banana, or ChatGPT-image via Kie.ai
- 4Animate each scene · generate-video · Veo3 fast · Kie.aiVeo3 fast image-to-video animates each scene using scene-0 image as starting frame
- 5Merge scenes · merge-video · fal.aifal.ai merges all scene clips into the final video
- 6Log · log · Google SheetsLog prompts, choices and output URLs to Google Sheets (Sheet1 + Sheet2 human-in-the-loop tracking); separate bulk-run trigger sub-workflow available
Missing: full agent system prompt not quoted verbatim
RoboNuggets · n21 - The Ultimate Agent for Infinite UGCs- 1
Unrestricted-content video generator (Seedream + Wan2.5 vs Veo3)Runsvideo for use cases frequently rejected by stricter models (e.g. fictional characters, foreign-dialogue animation, viral faceless content)
Seedream 4.0Wan2.5Veo310s clips with audio (Wan2.5)- 1Generate starting frame · generate-image · Seedream 4.0 · Kie.aiSeedream generates the starting frame image
- 2Animate (parallel branch A) · generate-video · Wan2.5 · Kie.aiWan2.5 image-to-video generates a 10s clip with audio, near-zero content restrictions
- 3Animate (parallel branch B, comparison) · generate-video · Veo3 · Kie.aiVeo3 image-to-video generates a comparison clip (frequently rejected — 3 of 7 test cases rejected)
Missing: use-case-specific prompts for the 7 tested scenarios not fully quoted
RoboNuggets · n22 - Wan2.5 x Seedream Unrestricted Videos- 1
Live-action character-swap selfie videoRunsa short selfie-style video sequence (multiple stitched scenes) where the user appears alongside live-action versions of fictional characters, played by real actors, on a fake film set.
ChatGPT (unspecified model)Nano Banana ProKling 2.5selfie-style short videomulti-scene stitched video- 1Ideate characters and actors · researchAsk ChatGPT (or any LLM) to suggest which fictional characters to feature and which real actors should portray them; chat back and forth to swap any character or actor.
- 2Generate selfie stills · Picture · Nano Banana Pro · Kie.aiFeed a user photo plus a per-character prompt (written by ChatGPT) into Nano Banana Pro via Kie.ai to create a hyper-realistic selfie-style still of the user next to the actor-as-character on a fake film set.
Prompt
Create a selfie-style image of this young woman together with Goku from Dragon Ball Z in a hyper-realistic version, portrayed by Brad Pitt. They are on the filming set of the real-life version of Dragon Ball Z. Brad Pitt has long, yellow, very natural-looking hair. Both are casually posing for the camera in a natural, friendly way. Brad Pitt is wearing Goku's classic combat outfit. The film set includes Dragon Ball Z details. She is not holding anything in her hand. No mobile phones. No phones in hands. She is wearing the same attire as in the reference photo
- 3Animate first/last-frame clips · Video · Kling 2.5 · Kie.aiUse Kling 2.5's first-and-last-frame mode (via Kie.ai) to turn each still into a clip where the user walks/transitions between scenes, giving seamless transitions when stitched.
Prompt
Selfie-style video. The young woman politely thanks the actor and walks to another area of the set and poses with another actor. She is not holding anything in her hand. No mobile phones. No phones in hands.
- 4Stitch scenes · editStitch all generated clips together in a standard video editor to form the final multi-scene video.
- 5Automate at scale (optional) · By hand · n8n + AirtableThe author automated all 17 scenes of the real build in under 30 minutes using an n8n workflow driven by an Airtable base (rows define each character/actor pairing), instead of running each step manually.
RoboNuggets · b4 - How I Created the Live Action Naruto Video- 1
Infinite UGC ad generatorPartlya batch of N UGC-style vertical ad videos, image-to-video, from one reference product/actor photo plus a creative brief, each with its own scene
GPT-image-1 (gpt4o-image)Veo3-fastUGC advertical batch- 1Set brief · By hand · n8nEdit Fields node collects reference image URL, number of videos, dialogue script (or 'suggest for me'), model, aspect ratio, special requests.
- 2Analyse reference · Picture · OpenAI (not named) · OpenAIVision call describes the reference image.
- 3Generate N scenes · Words · not named · n8nAI Agent produces N scenes, each with image_prompt + video_prompt + aspect_ratio + model.
- 4Split scenes · edit · n8nSplits the scene array into one item per scene.
- 5Generate scene image · Picture · GPT-image-1 · Kie.aiGenerates the scene's image from the reference photo and image_prompt.
- 6Poll image · Check · Kie.aiPolls until the image is ready.
- 7Generate scene video · Video · Veo3-fast · Kie.aiAnimates the scene image using the video_prompt.
- 8Poll video · Check · Kie.aiPolls until the video is ready.
- 9Deliver · publish · Box.comUploads the finished clip.
Missing: no verbatim AI Agent prompt in source; referenced n8n template JSON and sample product photo not present locally.
RoboNuggets · R38 | Infinite UGC Adverts AI System- 1
AI influencer content generatorRunsa set of finished social posts (still images and/or short talking-to-camera videos, vertical or horizontal) of a consistent AI character wearing a specified outfit in a specified setting, auto-scheduled to Instagram
Nano BananaSeedream v4 Edit (bytedance/seedream-v4-edit)Veo 3.1 (V3 fast)GPT-4.1talking-to-camera poststatic character postverticalhorizontal- 1Character creation (manual, one-off) · Picture · Gemini or ChatGPT image gen · Google Gemini / OpenAICreates the starting character reference image.
Prompt
subject: Young male with light tan skin tone, dark short hair styled loosely to one side, and calm, confident facial expression. clothing: Oversized bright pink hoodie with drawstrings visible, paired with casual dark pants. pose: Seated casually with knees bent and one arm resting on thigh, relaxed and natural posture. setting: Minimalist studio backdrop in soft pink gradient, evenly lit. lighting: Balanced, soft studio lighting with minimal shadows for a clean look. camera: Medium shot at eye level, straight-on framing capturing upper body and facial features clearly. mood: Youthful, trendy, approachable vibe with a hint of confidence and modern streetwear style. style: Realistic, high-resolution portrait with subtle depth and texture detail in clothing and skin. composition: Centered subject, uncluttered background emphasizing color harmony between outfit and backdrop.
- 2Multi-angle reference set (manual, one-off) · Picture · Seedream v4 Edit · Kie.aiGenerates multiple reference images (front/side/3-quarter/back, various expressions) of the same character.
Prompt
setting: Professional photo studio with clean white background and even studio lighting pose: Character standing naturally in various stances, showing full body from head to toe angles: front view, side view, three-quarter view, and back view expressions: neutral, smiling, serious, and candid variations camera: DSLR or mirrorless camera, eye-level framing, consistent distance and focus lighting: balanced studio lighting with soft shadows and white floor reflection style: realistic, high-resolution studio photography with clean composition and minimal background reference: use the provided character reference image for face, clothing, and body details
- 3360-degree rotation video (manual, one-off) · Video · Veo3 (V3 API) · Kie.aiGenerates a 360-degree rotation video of the character for extra angle captures.
Prompt
The character stands in a well-lit studio and rotates 360 degrees in place, showing every angle of his body and outfit. No dialogue, no extra movement—just a smooth, continuous rotation in a neutral studio background.
- 4Intake form · By hand · n8nCollects character image, setting image, optional item image, image count, video count, creative direction text, aspect ratio.
- 5Upload assets · edit · Kie.aiConverts uploads to base64 and uploads to Kie.ai's file store.
- 6Generate post prompts · Words · GPT-4.1 · OpenAIGiven the theme, creative direction, character brief and reference URLs, produces N structured posts (title, caption, post type, image prompt, video prompt).
- 7Split posts · edit · n8nUngroups the AI agent's post array into individual items.
- 8Generate image · Picture · Nano Banana · Kie.aiCreates the post image from the image prompt plus three reference images.
- 9Poll/fetch image · Check · Kie.aiFetches the finished image by task id.
- 10Log to Google Sheet · Words · Google SheetsAppends title/caption/image/id/publish-status row.
- 11Filter for video posts · Check · n8nKeeps only posts where type=video and the image succeeded.
- 12Generate video · Video · Veo 3.1 (V3 fast) · Kie.aiAnimates the generated image into a talking/acting clip, using it as the first frame.
- 13Poll/fetch video · Check · Kie.aiFetches the finished video by task id.
- 14Update sheet row with video · Words · Google SheetsUpdates the row with the video output and status.
- 15Auto-publish · publish · Blotato → InstagramEvery 8 hours, grabs the next 'scheduled' row, posts to Instagram, marks it done.
Missing: the AI agent's system prompt and the 'character brief' GPT output structure are described but not quoted verbatim; would need to reconstruct or source the referenced n8n template/custom GPT for exact text.
RoboNuggets · R44 | The AI Influencer Toolkit- 1
Talking head11
RIS talking avatar from a locked Sora faceRunsa new talking-head clip of a previously generated (locked) Sora 2 Pro character face saying new lines, without regenerating the face
RISElevenLabs V3 Alphatalking head- 1Select the locked character image · By hand · Max FusionIn Max Fusion Images, select the previously locked character still (re-uploading a person image directly into Sora 2 Pro gets flagged/rejected). Optionally save as a reusable 'actor'.
- 2Generate the speech track · Voice · ElevenLabs V3 Alpha · ElevenLabs (integrated in Max Fusion)Design a new voice with ElevenLabs V3 Alpha or reuse a saved cloned voice. Paste the script line; insert inline emotional/expression tags (e.g. curious tone, examining tone, disappointed tone) to shape delivery.
Prompt
I'm going to wear them both for one week straight and see which one survives.
- 3Generate the talking video · Talking video · RIS · Max FusionIn RIS, go to Audio Script > Generate Speech, combine the locked face with the ElevenLabs audio; optionally apply minimal 'expressive hands'/gesture settings.
AI Ads (Ad Creators Lab) · Creating engaging talking avatars with Sora 2 Pro and RIS- 1
Sora 2 podcast-style ad clip (purpose-built GPT prompter)PartlyA short two-person podcast-style ad clip discussing a product, generated from a one-line product brief.
Sora 2Sora 2 ProSeedance 2podcast clip- 1Open the custom GPT prompter · Words · custom GPT · OpenAIOpen the course's own purpose-built custom GPT, trained on its own Sora 2 prompt structure ("content system V2 GPT").
- 2Give the brief · Words · custom GPT · OpenAIGive the GPT a one-line product brief, or let it use its own creativity.
- 3Generate · Video · Sora 2 / Sora 2 Pro / Seedance 2 · OpenAI / Kie.ai etc.Paste the GPT's output prompt directly into Sora 2 (or Sora 2 Pro, or Seedance 2) and generate.
Missing: The custom GPT's own prompt structure/training is the course's proprietary tool and is not disclosed in the lesson; only its output pattern is shown.
The AI Content System · Sora 2 - Podcast Clips- 1
Daily AI-persona talking-head social video (n8n queue)PartlyA daily talking-head social video from an AI persona, queued from the ads dashboard.
ElevenLabsHeyGen- 1Run Now · Words · n8nOperator presses Run Now; triggerRun() POSTs to the n8n webhook social-content-trigger.
- 2Pipeline (outside repo) · Video · Claude, ElevenLabs, HeyGen · n8nPer UI copy: Intelligence Bank → Claude script → ElevenLabs voice → HeyGen avatar → Late API publish.
- 3Queue read · Check · SupabaseGET reads the social_content_queue Supabase table, rendered in the dashboard.
Missing: Generation pipeline is entirely inside n8n, not auditable from this repo.
Ours ·- 1
Authority thought-leadership video (Story-Sell framework)PartlyA thought-leadership talking-head video following a Story-Sell script framework, queued from the ads dashboard.
HeyGenVeo 3.1Sora 2Seedance 1.5 ProKling 2.1- 1Submit key points · WordsOperator enters key_points in a textarea; forwarded verbatim.
- 2Hand off · Video · n8nPOST to n8n webhook authority-video.
- 3Pipeline (outside repo) · Video · picker: Veo 3.1 / Sora 2 / Seedance 1.5 Pro / Kling 2.1 · n8nPer UI copy: Perplexity research → Claude script → HeyGen avatar → B-roll model (operator-picked).
Missing: Generation pipeline is inside n8n, not auditable from this repo.
Ours ·- 1
InfiniteTalk Long-Form UGC (WAN2GP on RunPod)RunsA 30s+ continuous-lipsync talking video from one start image plus an audio file.
Wan2.1 InfiniteTalk 14B- 1Deploy pod · By hand · RunPodRunPod deploy template 'wan2gp', GPU A40 (~$0.40/hr), login admin/gpuPoor2025 (overridable via env vars), port 7862.
- 2Configure + generate · Video · Wan2.1 InfiniteTalk 14B · RunPod (self-hosted WAN2GP)Dropdowns Wan2.1 → InfiniteTalk 14B → Single/Multi Speaker; inputs = start frame (Nano Banana Pro recommended) + audio file (ElevenLabs/MiniMax, optionally Adobe Podcast-cleaned); generate; always stop the pod after.
Ours ·- 1
AI-Influencer network — Pipeline A (repurpose + trend avatar)PartlyTwo sub-workflows: real-video repurposing into captioned clips, and a trend-driven HeyGen avatar video of a real person (Simon).
WhisperClaudeHeyGenElevenLabs- 1A-1 Repurpose real video · Video · Whisper, Claude · Late API, TelegramWhisper transcription → Claude extraction → FFmpeg cut+caption → Telegram approval → Late API publish.
- 2A-2 Trend-driven avatar · Video · HeyGen, ElevenLabsHeyGen avatar of Simon + ElevenLabs voice clone, scheduled cadence.
Missing: Both sub-workflows described only, not literal.
Ours ·- 1
Personal-brand AI avatar video pipelinePartlyAn avatar talking-head video of Simon plus auto-clipped vertical B-roll shorts.
ElevenLabsHeyGenKling 2.6 / Seedance- 1Clone voice · Voice · ElevenLabs eleven_turbo_v2_5 · ElevenLabs APIElevenLabs clone, Simon's voice ID 2GErPEnQqbyeqhZQPM6r, model eleven_turbo_v2_5.
- 2Render avatar · Video · HeyGen · HeyGen APIHeyGen API render, avatar from a 2-5min base video, ~2-4min render.
- 3B-roll · Video · Kling 2.6 / Seedance · Kie.aiKie.ai (Kling 2.6/Seedance) for B-roll — explicitly not Sora/Veo3.
- 4Auto-clip + distribute · edit · OpusClip, Late APIOpusClip auto-clips into 10-15 vertical shorts; Late API distributes. Runs on a FastAPI/Hetzner endpoint, no Zapier/Make.
Missing: Tool chain and env-var names given, no literal prompts/API bodies.
Ours ·- 1
Longform AI talking-character videoRunslongform (5-6 min) talking-character video with voiceover
ElevenLabs (voice)InfiniteTalk (video)SeaArt or Gemini (character image)480p720p- 1Source character image · generate-image · Gemini or SeaArt · Kie.ai (SeaArt) / Google (Gemini)Generate the character image via Google Gemini (free, no aspect-ratio control) or SeaArt (aspect-ratio + 4K + batch + 'edit' feature to place character into a new scene)
- 2Generate voiceover · generate-voice · ElevenLabs (e.g. Turbo v2.5) · Wavespeed AIHTTP POST with Bearer-token auth, body {text, voice_id}; poll via switch node checking status attribute (completed/processing/error)
- 3Generate talking video · generate-video · InfiniteTalk · Wavespeed AIHTTP POST body {audio, image, resolution: 480p/720p, optional prompt e.g. 'minimal movement'}, poll via switch node; ~30min total generation time for a 5-6min video
Prompt
minimal movement
- 4Optional upscale · Upscale · Topaz Labs · fal.aiOptional upscale via Topaz Labs
Missing: exact InfiniteTalk pricing per resolution not itemised beyond the ~60c/generation total
RoboNuggets · n26 - Longform AI Characters with InfiniteTalk- 1
Higgsfield Supercomputer UGC talking-head review videoPartlyA short AI-generated UGC-style talking-head product review video, built end to end from a product name and a couple of clarifying answers.
SoulUGC talking-head review video- 1Brief the agent · Words · HiggsfieldAsk Supercomputer to make a UGC video; it asks clarifying questions (product, UGC type, length) via a built-in ask-user-question tool.
Prompt
Make a UGC video Seedance 2.0.
- 2Generate the character · Picture · Soul · HiggsfieldSupercomputer preloads its UGC workflow: it generates a starting-frame character image first, before animating, and asks the user to approve the character-generation prompt.
- 3Write the script and storyboard · Script · GPT Image 2 · HiggsfieldSupercomputer writes a monologue/script for the review, then generates a storyboard image from it, showing the full prompt before sending.
- 4Animate the final video · Video · HiggsfieldSupercomputer animates the approved character/storyboard into the final talking-head video with voice/monologue.
Missing: The specific video/lip-sync model used for the final animation step is not named in the source; the course notes visible AI tells (product magically appearing/duplicating in the presenter's hand, an off-brand scream at the start) — quality was judged 'not fully there yet' from a one-shot run.
RoboNuggets · b17 - Higgsfield Supercomputer- 1
Baby podcasts AI machine (Hedra lip-sync)Partlya lip-synced 9:16 video of a 'babyified' toddler version of a real TikTok creator, speaking that creator's own remixed audio in a podcast-studio setting
Hedra (character-2, video generation with lip-sync)ChatGPT image editing (for the baby-portrait source image, manual)babyified talking-head / podcast remix- 1Scrape source TikTok videos (Part 1: Original Vids Agent) · research · Apify (TikTok scraper by Clockworks)On a monthly schedule trigger, an Apify TikTok scraper (Clockworks actor) pulls up to 100 of a given TikTok profile's videos from the last 30 days.
- 2Log scraped videos to Google Sheets · edit · Google SheetsAppend each scraped video's caption, video URL, audio URL (music.meta.playUrl), original date, and production status='for production' to a Google Sheet, plus an auto row-number ID and a randomized image_url formula pointing at a manually-prepared list of baby-portrait images.
- 3Prepare the baby-portrait source image (manual, one-off) · By hand · ChatGPT image generation (model unspecified in lesson) · OpenAI (ChatGPT web app, manual) + freeimage.host (hosting)Upload a photo of the real creator to ChatGPT with a fixed prompt to generate a photo-realistic vertical 9:16 'babyified' toddler portrait of them at a podcast desk with a microphone and studio lighting; download and upload the result to freeimage.host, then paste the hosted image URL into the sheet's image list. Done once per account, reused across many videos.
Prompt
Create a photo-realistic, vertical 9:16 portrait of a baby version of this person. Keep their facial structure and profile recognizable. Make it cute and stylized like a toddler. Show them seated at a podcast studio desk with a professional microphone in front of them, with studio lighting, to clearly emphasize the podcast setting.
- 4Get next video for production · edit · Google SheetsGet the first sheet row where production status = 'for production' (one row per run, daily schedule trigger).
- 5Create audio container · edit · Hedra (api.hedra.com/web-app/public/assets)Create an empty audio-file container on Hedra to receive the TikTok's original audio.
Prompt
{ "name": "audio-file", "type": "audio" } - 6Fetch and upload the source audio · edit · Hedra (api.hedra.com/web-app/public/assets/{id}/upload)GET the TikTok's original audio file from the sheet's audio URL, then POST-upload it into the created Hedra audio container.
- 7Create image container and upload baby portrait · edit · Hedra (api.hedra.com/web-app/public/assets)Create an empty image-file container on Hedra, then fetch the sheet's babyified portrait image URL and upload it into that container.
Prompt
{ "name": "image-file", "type": "image" } - 8Create video (Hedra lip-sync animation) · Talking video · Hedra character model, ai_model_id d1dd37a3-e39a-4854-a298-6510289f9cf2 · Hedra (api.hedra.com/web-app/public/generations)Hedra animates the uploaded baby-portrait image, lip-synced to the uploaded audio, into a 720p 9:16 video of a baby podcast host speaking with expressive reactions and head movement.
Prompt
{ "type": "video", "ai_model_id": "d1dd37a3-e39a-4854-a298-6510289f9cf2", "start_keyframe_id": "{{ $('Create image container').item.json.id }}", "audio_id": "{{ $('Create audio container').item.json.id }}", "generated_video_inputs": { "text_prompt": "A baby podcast host sitting in front of a microphone, speaking with calm intensity and natural focus. Expressive facial reactions, animated head movements, and strong eye contact with the camera. A professional podcast studio setup with polished lighting in the background", "resolution": "720p", "aspect_ratio": "9:16" } } - 9Poll for completion · Check · Hedra (api.hedra.com/web-app/public/generations/{id}/status)GET the generation status repeatedly (wait node + IF loop) until status = complete, then route to logging; otherwise loop back and wait again.
- 10Log finished video and mark done · edit · Google SheetsUpdate the sheet row (matched by ID) with production status='done', the finished video URL, and publishing status='for publishing'.
- 11Optional: auto-publish · publish · Blotato (per the referenced publishing-agent lesson)The lesson points to a separate 'ultimate publishing agent' workflow that reads the video URL from the sheet and posts it to YouTube/Instagram/TikTok on a schedule.
Missing: the baby-portrait source image is made manually via an interactive ChatGPT image-edit step (not automated in this workflow) and must be prepared once per account before the automation can run; Hedra requires a paid plan (lesson notes a $10/month minimum tier) to use the API.
RoboNuggets · R30 | The Baby Podcasts AI Machine- 1
Longform AI talking-head explainer + B-roll systemPartlya multi-minute (up to 10 min per generation click, chainable to 5-min segments) talking-head explainer video with an AI character narrating a script, plus optional B-roll cutaway clips
Fish Audio S1InfiniTalkNano Banana ProKling v2.6 Pro (image-to-video, B-roll)talking-head explainer16:99:16- 1Prep inputs · By hand · Nano Banana Pro (for the character image) · Kie.ai / GeminiGathers a driving character image (via a reverse-engineered SEALCaM prompt from a reference photo) and writes the video request (topic, tone, length).
- 2Create script · Script · LLM via n8n AI agent (model unspecified) · n8nGenerates a timed, segmented spoken script.
Prompt
the full 'System Prompt for AI Agent (Talking Head)' YAML system prompt (~230 lines total combined with the B-roll prompt below), including the SEALCaM framework definition, was harvested verbatim by the researching agent from the source file
- 3Create voice · Voice · Fish Audio S1 · fish.audioConverts the script to MP3 narration.
- 4Create video (talking head) · Talking video · InfiniTalk · WavespeedAnimates the driving image with the audio.
Prompt
Natural movements, natural eyebrow motion, minimal gestures, and a professional smiling tone
- 5Create B-roll (bonus) · Video · Nano Banana Pro (image), Kling 2.6 (video) · Kie.ai / WavespeedA separate AI agent breaks the narration into ~20-second segments, writes SEALCaM image and video prompts, generates start-frame images, then animates them.
Prompt
the full 'System Prompt for AI Agent (B Roll)' verbatim SEALCaM-based system prompt was captured by the researching agent
- 6Edit/finish · edit · none (external NLE)Manual assembly in Premiere/CapCut/Canva; not automated.
Missing: the n8n workflow JSON itself was not opened/harvested as readable text (only referenced), so exact node wiring/polling logic is unverified, though every model call is directly API-reachable today.
RoboNuggets · R52 | Longform AI Explainers System- 1
Animation11
Talking product/object avatarPartlyan 8-10 second clip in which an inanimate product (e.g. a jar of moisturizer) is anthropomorphized with a face and delivers a first-person ad-copy line to camera
Seed DreamNano Banana ProKling 3.0 ProGoogle Veo 3.1Sora 2 ProSeedance 1.5 Protalking object- 1Ideate the object's angle · Script · custom ChatGPT GPT (not named) · ChatGPTUse a pre-built ChatGPT GPT (found online, linked in the course, not built by the presenter) to brainstorm the misconception/category-failure the object represents (a competitor-product analogy).
Prompt
The object I want to show is an over-the-counter moisturizer, and I want to talk about how it uses seed oils and is bad for your skin.
- 2Generate the character still · Picture · Seed Dream / Nano Banana Pro · Max FusionUse Seed Dream for speed/cost when the image doesn't need legible text; use Nano Banana Pro (or Nano Banana 2) specifically when the image needs readable on-object text (e.g. an ingredients list).
- 3Generate and compare across engines · Video · Kling 3.0 Pro / Google Veo 3.1 / Sora 2 Pro / Seedance 1.5 Pro · Max FusionAnimate the locked still, comparing the same prompt head-to-head across four engines. Rule: describe what IS happening, not instructions ('She is doing this' not 'Have her do this'). Match clip duration carefully to script length or dialogue gets cut off/rushed.
Prompt
I'm your over-the-counter moisturizer. I'm packed with cheap seed oils to keep my price low. Sure, I feel silky for a minute, but your skin hates heavy oils like these. I give you clogged pores and irritation.
Missing: the GPT's own full script/image prompt output was not read aloud/quoted in the transcript
AI Ads (Ad Creators Lab) · Talking Objects Tutorial (Lesson 24)- 1
Crochet/claymation storyboard-to-video adRunsa full crochet/knitted-yarn stop-motion 'diorama' style video ad, built scene by scene from a script via a Claude Skill, then edited in CapCut
Seedream 4.5Nano Banana ProSeedance 1.5 ProKling 3.0 ProElevenLabscrochet/claymation diorama- 1Install the Crochet Ad Visuals Skill · By hand · Claude (Opus) · Claude DesktopClaude Desktop > Customize > Skills > upload the provided skill zip. Trigger with /crochet add visuals and upload the script.
- 2Answer the character question · By hand · Claude (Skill) · Claude DesktopThe skill asks whether a consistent character is needed across visuals, with three choices.
Prompt
Before I break this down, one quick question: Do you need a consistent character to appear across all the visuals? A. Yes — here's my character description: [user describes] B. Yes — suggest one based on my script C. No, each visual can have its own character or no character Reply with A/B/C and I'll generate the visual breakdown.
- 3Get the phrase-by-phrase visual breakdown · Script · Claude (Skill) · Claude DesktopThe skill splits the script into phrases and offers ~5 bulleted visual-idea options per phrase; the user picks one per phrase.
- 4Get the full storyboard prompts · Script · Claude (Skill) · Claude DesktopFor each chosen visual, the skill returns a director's note (Type A/B/C), a start-frame prompt, an optional end-frame prompt, and an I2V (video) prompt, all built from the skill's fixed 5-part image structure and 4-part I2V structure.
- 5Generate the character reference · Picture · Seedream 4.5 / Nano Banana Pro · Max FusionCopy the character prompt into Max Fusion Images with Seedream 4.5 (cheapest) or Nano Banana Pro.
Prompt
Stop-motion animation diorama scene, handmade knitted and crocheted miniature world. Full-body small yarn-textured [demographic] character, [age], [hair], wearing [outfit in yarn/knit terms], standing in a neutral pose against a simple crocheted backdrop. Every element is a handmade fabric prop physically crafted from wool and yarn.
- 6Generate start/end frames · Picture · Seedream 4.5 / Nano Banana Pro · Max FusionCopy the start-frame prompt into Max Fusion Images, re-attaching the character reference every time (forgetting this visibly dropped consistency); optional inline chat edits for small tweaks (e.g. 'make him look even more sad').
Prompt
Stop-motion animation diorama scene, handmade knitted and crocheted miniature world. Full-body small yarn-textured dark-skinned male character, early 30s, buzz cut, wearing a fitted red knitted crewneck, standing in a simple crocheted apartment living room, one arm raised and bent with his nose pressed directly into his inner wrist and forearm — actively sniffing it. His face is neutral, eyebrows just starting to furrow. Every element is a handmade fabric prop physically crafted from wool and yarn. Simple crocheted apartment background softly out of focus. The entire scene is fully crocheted and knitted — every surface, character, object, and skin is made of visible yarn stitches and fabric loops. No realistic skin, no smooth surfaces, no photorealistic textures anywhere in the frame. Every element, including hands, arms, faces, and backgrounds, must have visible knit and crochet stitch texture throughout. Stop-motion animation diorama scene, handmade knitted and crocheted miniature world. Lighting: soft cool flat natural light from a window to the left. Negative: realistic skin, smooth skin, photorealistic hands, photorealistic faces, realistic textures, smooth surfaces, CGI render, claymation, clay texture, Pixar style, Disney style, 3D animation, plastic texture, flat cartoon, airbrushed, anime, illustrated, 2D, studio lighting, ring light, HDR, symmetrical lighting, harsh shadows, perfectly exposed.
- 7Generate the scene-to-scene video · Video · Seedance 1.5 Pro (cheap test) / Kling 3.0 Pro (quality) · Max FusionCopy the I2V prompt into Max Fusion Videos. Generate first on Seedance to check timing/motion cheaply, then re-generate on Kling for final quality. Always include the character-consistency clause and mandatory claymation-style clause.
Prompt
The knitted character sniffs his inner wrist once, pulls his face back with a visible grimace — disgusted and deflated. He lowers his arm slowly and shakes his head. The apartment background and lighting stay completely still. Cool flat window light consistent throughout. Camera locked. Ensure the entire scene remains fully crocheted and knitted throughout every frame — all characters, objects, and surfaces must retain visible yarn stitch texture with no realistic or smooth surfaces at any point. Non-CGI. Non-cinematic. Animations must start at the first frame. Non-disney.
- 8Repeat per phrase, edit in CapCut · edit · ElevenLabs (voiceover) · CapCutRepeat generation per phrase, always re-attaching the character reference. In CapCut: drag in the ElevenLabs voiceover (pre-labelled to match clips), nudge/trim to line up action with speech, use Speed > Curve for 'flash out/flash in' velocity-ramp transitions timed to spoken words, add a downloaded 'liquid glugging' SFX and built-in sound effects (Drinking, cloth hits) synced to on-screen actions.
AI Ads (Ad Creators Lab) · The Ad We Are Making (Crochet) / Claude Skill Setup / Script Visualisation & Image Creation / Generating Videos & Transitions / Editing Claymation Step-By-Step (Lessons 7-10, 20, 22)- 1
3D animated explainer ad generator (packaged Replit + KIE.AI app)RunsA 3-7 clip, ~30-second vertical 3D-animated explainer ad, scripted, voiced, generated and stitched end to end from one product photo and a brand kit.
KIE.AI (unspecified underlying video/voice models)3D animated explainer ad- 1Get a KIE.AI API key · By hand · KIE.AISign up at kie.ai/api-key, top up credits, copy the key.
- 2Import the pre-built app into Replit · By hand · ReplitImport the course's `ai-animation-ads.zip` into a new private Replit project.
- 3Launch the app · By hand · Replit AgentPaste the fixed launch prompt verbatim into Replit Agent to start the pre-built Node.js server with no code changes.
Prompt
This is an AI Animation Ad Generator web app: a single static HTML frontend served by a small zero-dependency Node.js server. It's already fully built. Do not modify any source code. Just help me get it running. Setup steps: 1. No dependencies to install. The Node server uses only the built-in standard library. There is no package.json to npm install, no build step. Confirm Node 18+ is available. 2. No environment variables or secrets to add on the Replit side. The API key (KIE.AI) is entered later through the app's Settings panel and stored in a local config file the server reads. 3. The app runs with `node animation-ads-server.js` and listens on the PORT env Replit provides (defaults to 4174). Please set up a workflow to start it. The .replit config is already included and handles all of this. Once the server is running, open the Webview's "Open in new tab" link, then go to Settings in the app and paste in my KIE.AI key. Do not change any source code. Just configure the environment, start the app, and confirm the Webview is live.
- 4Add the key and build a brand kit · By handPaste the KIE.AI key into the app's Settings; in Brand Setup fill in brand name, one-line pitch, brand color, ad voice (Cinematic/Playful/Bold/Dreamy/Luxury) and product photos.
- 5Analyze the brand · Words · KIE.AI (via the app)Click "Analyze brand" to get 4 distinct AI-pitched ad concepts (title, premise, hook, shot list), or write a custom brief.
- 6Render · Video · KIE.AI (via the app)Pick a clip count (3/5/7), generate one or several concepts. The app auto-scripts each clip's scene + voiceover line, renders each clip in parallel (9:16, 720p, conditioned on the product photo), QC-checks voice completeness and re-renders any cut-off line, then stitches all clips into one finished ad.
The AI Content System · AI Animation Ad Generator - Playbook- 1
Multi-style AI animation ad (Omni Ad Engine SOP via Claude)PartlyA 30-45 second animated ad in any chosen visual style (Lego, Pixar, claymation, paper collage, etc.), delivered as a full set of paste-ready 5-second Gemini Omni prompts plus an ElevenLabs voiceover script, built from one product photo with no reference image needed by the video model.
Gemini Omni (Veo)ElevenLabsLego/brick-builtPixar/glossy 3Dclaymation/stop-motionpaper collageanime/2D celfelt & plushwatercolor/ink storybookminiature/tilt-shiftneon/cyberpunkretro/80s VHScustom build-your-own style- 1Load the SOP into Claude · Words · Claude · AnthropicPaste the attached "THE OMNI AD ENGINE" SOP file into a Claude chat (or set up as a Claude skill); Claude now acts as an AI performance-ad director.
- 2Give the product and pick a style · Words · Claude · AnthropicAttach a product photo and ask for an AI animation ad; either name a style or ask Claude to recommend 2-3 fitting Style Skins.
- 3Get the full prompt package · Words · Claude · AnthropicClaude returns a Concept Header, a Locked Product Block (replaces a reference image, since Omni is prompted text-only for consistency), an interval-by-interval breakdown of 5-second Omni prompt blocks following a 7-beat narrative formula, then a full ElevenLabs voiceover script.
Prompt
INTERVAL FORMAT (every 5-second Omni clip): 00:XX - 00:XX | Interval N: <punchy title> Visual Composition: <the full frame - the style's materials & colors, the layout, where each element sits, camera angle (default 9:16 vertical), lighting, depth, the palette mood for this beat>. The Motion: <the ONE hero action in these 5s, in order, with timing/feel - what moves, transforms, enters, assembles, reveals; how the camera moves (push-in, pull-back, macro zoom); any short text card that slams in; the style's signature motion verbs>. Visual Assets: <the discrete pieces/overlays Omni should build in this style - characters, props, diagram parts, headline, arrows, stamps, drop shadows>. [Narration over this interval]: "<exact VO words for these 5s>" (kept for the editor; NOT pasted into Omni) LOCKED PRODUCT BLOCK - <Product name> (paste verbatim into every interval it appears in): "<form factor> rendered in <the chosen style's materials & exact colors>, <the wordmark/logo: exact text, font feel, color>, <accent bands/badges: exact text + color + shape>, <1-3 on-pack claims as small legible lettering>, <product-adjacent props (gummies/capsules/device) in exact color & finish>, <style construction notes: e.g. 'brick-built with stud texture' / 'glossy 3D Pixar render' / 'sculpted matte clay'>, soft contact shadow, on <background>, top-down or hero-center, 9:16." 7-beat narrative formula: Hook -> "Here's how"/root tease -> Root-cause explainer -> Why others fail -> Product reveal + mechanism -> Payoff/proof -> CTA. Output order: Concept Header -> Locked Product Block (+ Character Block if needed) -> Interval-by-Interval Breakdown -> Direction (music arc, SFX kit, caption treatment, palette arc) -> Full Voiceover Script for ElevenLabs.
- 4Generate each interval in Gemini Omni · Video · Gemini Omni (Veo) · Google Flow (flow.google / labs.google)Paste each 5-second interval's prompt (often combined two at a time into a 10-second block, with the Locked Product Block and concept context repeated) into Google Flow / Gemini Omni and generate.
- 5Stitch and finish (per the SOP, not shown live) · edit · ElevenLabs + an editorStitch all clips, record and layer the ElevenLabs voiceover, add music/SFX/burned-in captions per the SOP's Direction section, export.
Missing: Step 4 (Google Flow generation) is shown only via the browser UI in this lesson, with no API/MCP route mentioned. Step 5 (stitch/VO/music/captions) is specified by the SOP but not demonstrated.
The AI Content System · Multi Style AI Animation Ad Generator- 1
Video-as-code motion graphics renderRunsdeterministic motion-graphics / explainer / title-card MP4s, authored as plain HTML/CSS with data-start/data-duration/data-track-index labels, not live-action generation
title cardmulti-scene explainer- 1Install · By hand · HyperFramesInstall prerequisites (Node 22+, FFmpeg, headless Chrome auto-downloads) then install official agent skills.
Prompt
npx skills add heygen-com/hyperframes
- 2Plan · Words · Claude Code / Codex / Cursor / Gemini CLIPlan the composition from a plain-English brief.
- 3Write the composition · edit · HyperFramesWrite the video as HTML tags with data-start, data-duration, data-track-index labels per element; write animation in CSS or GSAP, parked as a paused, seekable timeline on window.__timelines (mandatory for deterministic renders).
- 4Add media and lint · edit · HyperFramesAdd media and captions, lint.
- 5Preview · Check · HyperFrames dev serverPreview in-browser with live reload; catches errors before rendering.
- 6Render · Video · HyperFrames CLIHeadless Chrome captures every frame, FFmpeg stitches to MP4.
AIOps · HyperFrames: HeyGen's Video-as-Code Framework with Agent Skills- 1
Stylized anime video with consistent character/stylePartlya short stylized anime-look video clip (e.g. a girl in a cyberpunk jacket in rain/neon city) with subtle, style-preserving motion
Nano Banana ProTextideoanime character video- 1Generate anchor frame · Picture · Nano Banana ProGenerate a single anime-style still image to use as the anchor/start frame, using LoRA-style quality and negative-prompt tags to lock the anime look and avoid 3D/photoreal drift.
Prompt
(masterpiece, best quality:1.2), 1girl, solo, cyberpunk jacket, glowing circuitry, rain soaking clothes, neon city background, depth of field, looking at viewer, cinematic lighting, volumetric fog, <lora:NanoBananaPro_v1:1>
- 2Negative prompt (paired with step 1) · Picture · Nano Banana ProApplied alongside the step-1 prompt to suppress unwanted 3D/photoreal/low-quality/anatomy artifacts.
Prompt
(worst quality, low quality:1.4), 3d, photorealistic, monochrome, zombie, distortion, bad anatomy
- 3Animate the anchor frame · Video · Textideo · TextideoUpload the anchor frame to Textideo and write a motion prompt following the formula [Subject Action] + [Camera Movement] + [Atmosphere] to add subtle life (blinking, breathing, hair movement, weather) without breaking the anime style.
Prompt
Girl blinking slowly, breathing, rain falling in background, hair swaying in wind, slow camera zoom in, high fidelity, no morphing.
Missing: Textideo's exact API/pricing not given; motion scale and guidance scale are UI sliders whose exact API parameter names aren't confirmed
Free online · I Ditched Runway for Anime — Here Is the Superior Stack- 1
Documentary/history-style Short — per-shot styled animationRunsA vertical (9:16) documentary-style Short built shot-by-shot: each storyboard shot gets a brand-styled start frame, then a short animated clip (Ken Burns fallback if generation fails), assembled into one Short.
gpt-image-2Omni-Flash-ExtSeedance 2.0stylized vertical documentary- 1Start frame · Picture · gpt-image-2 · Codex CLI / muAPIFor each storyboard shot: FREE Codex CLI gpt-image-2 first, muAPI gpt-image-2 fallback. Takes frame_candidates per shot (default 2); a headless-claude vision judge promotes the winner. A shot marked 'continues' skips generation — it reuses the previous shot's last frame (chaining).
Prompt
Scene: {depict}. Subject: exactly the moment above, rendered as: {frame_style}. Important details: the named subject and action are unmistakably the focus; vertical composition with the subject in the upper two-thirds of frame; rich physical texture; historically accurate costume, props and setting for the era. Use case: opening frame of one shot in a stylized vertical documentary Short. Constraints: {negative}; portrait 9:16 orientation; no text anywhere. - 2Host frame · By hand · SupabaseUpload the start frame to the Supabase social-assets public bucket so the video provider can fetch it by URL.
- 3Animate · Video · Omni-Flash-Ext / Seedance 2.0 · APImart / Kie.aiDefault door: Omni (APImart), frame-mode i2v. Shots the storyboard flags 'complex' route to Seedance 2.0 via Kie.ai instead (capped by motion.max_complex, default 2). One retry per shot with a simplified motion prompt if the first attempt fails (one-variable change, word-capped at 80 per the i2v course rule).
Prompt
{motion}. Overall animation character: {motion_grammar}. Something visibly happens during the shot — never a frozen scene that only drifts past the camera. The scene: {depict}. Keep the established art style exactly. No dialogue, no speech, no on-screen text. One continuous shot. - 4Fallback · editIf generation fails on any shot, that shot falls back to a Ken Burns pan/zoom on its still frame (or scene image) — the Short still ships for review rather than blocking the batch.
Ours ·- 1
3D/Stylized Character RecreationRunsA 3D/stylized (e.g. Pixar-like) animated character video, built scene-by-scene from a reference screenshot.
ClaudeNano Banana ProVeo 3- 1Describe reference · Picture · ClaudeAsk Claude for a detailed JSON-format prompt from a screenshot/reference.
Prompt
Give me a detailed prompt for this image in JSON format
- 2Generate character · Picture · Nano Banana ProNano Banana Pro generate in portrait mode.
- 3Animate · Video · Veo 3Veo 3 Frames-to-Video with a literal action prompt, e.g. character stands up, appending 'No background sounds.'
Prompt
[Character] stands up. No background sounds.
- 4New scenes · Picture · Nano Banana Pro + Veo 3For new scenes, return to Nano Banana with a literal new-scene prompt, then back to Veo with a new literal action prompt; repeat per scene. Unwanted elements removed via literal negative prompts on both sides.
- 5Finish · edit · CapCutCapCut stitch + voiceover + captions.
Ours ·- 1
Medical/Biological 3D AnimationRunsA hyper-realistic 3D biological/mechanism visualization video (e.g. liver absorption, capsule dissolving, collagen binding, blood cells).
Nano Banana Proreve.comChatGPT ImageKling AI 2.5 TurboVeo 3- 1Generate base image · Picture · Nano Banana Pro / reve.com / ChatGPT ImageLiteral base-prompt formula with 4 full worked examples (liver/absorption, capsule dissolving, collagen binding, blood cells).
Prompt
A hyper-realistic 3D visualization of [specific biological process or product mechanism]
- 2Animate · Video · Kling AI 2.5 Turbo / Veo 3Literal motion-prompt formula with 4 examples, via Kling AI 2.5 Turbo or Veo 3.
- 3Polish · edit · CapCutCapCut polish: music, text overlays, logo.
Ours ·- 1
Higgsfield MCP brand book to animated logoRunsA brand book from a logo, a 6-panel logo-animation storyboard, and a short animated logo video, all generated through Claude talking to Higgsfield's models via its MCP connector.
Nano Banana 2GPT Image 2Seedance 2.0brand bookstoryboardlogo animation videolanding page mockup image- 1Install Higgsfield MCP skill and connect account · Words · HiggsfieldClone the robonuggets/higgsfield-skill into Claude Code, then have Claude open the browser to log in and connect the user's Higgsfield subscription (credits, not pay-per-call).
Prompt
I'm giving you a skill. Get the files with: git clone https://github.com/robonuggets/higgsfield-skill Follow the README to set it up. Recommend how it would best apply to our setup in plain language, and ask me a few questions to clarify my intent before you install it.
- 2Generate a brand book from a logo, two ways · Picture · Nano Banana 2 / GPT Image 2 · HiggsfieldUsing the design-system skill and an attached logo, Claude generates two brand-book iterations for comparison, one image per model, including colour hex codes and typography identification.
Prompt
I want to create a brand book for this logo, please. Can you use your design system skill to create two iterations of the brand book. Use Nano Banana 2 for one of the images and GPT Image 2 for the other image just so that we can compare them, please.
- 3Generate a 6-panel logo-animation storyboard · Picture · GPT Image 2 · HiggsfieldFrom the winning brand-book direction, Claude generates a 6-panel storyboard image showing how the logo evolves, with a description of what happens per panel.
Prompt
Create a storyboard for a logo animation for Spiderhead AI, and use GPT Image 2 for that so that we have like a six-panel storyboard of how the logo will evolve.
- 4Animate the storyboard · Video · Seedance 2.0 · HiggsfieldClaude sends the storyboard image as a reference to a video model to animate the logo reveal, checking resolution/duration with the user before spending credits.
Prompt
Can you now send this to Seedance 2.0 to animate this via the Higgsfield MCP? Make sure to send the storyboard as a reference and for Seedance to follow the instructions as per the storyboard and make it as clean and visually arresting as possible.
- 5(Branch) Generate a landing page mockup · Picture · GPT Image 2 · HiggsfieldIn the same chain, Claude also generates a landing-page mockup image using the design system and colour palette, which can then be handed to a fresh Claude Code session to build as a real, interactive website.
Missing: Per-call Higgsfield credit costs are not surfaced by the MCP; the design-system skill referenced is not itself harvested in this batch.
RoboNuggets · b15 - Automate EVERY AI Model with Higgsfield MCP- 1
Logo animationPartlyone short animated video turning a static logo into motion, with a scene timeline (0-2s intro, 2-60s main, final scene) and adjustable motion intensity per scene
Veo3Veo3-fastlogo animation- 1Read input row · By hand · Google SheetsReads a Google Sheet row with the logo URL and status = for production.
- 2Set creative direction · Words · n8nCaptures direction text, aspect ratio, model choice, dialogue in an Edit Fields node.
- 3Analyse logo · Picture · OpenAI (not named) · OpenAIVision call produces a YAML description: colour scheme, brand name, font style, visual description.
- 4Generate video prompt · Words · not named · n8nAI Agent builds a JSON video prompt with a scene timeline and intensity levels.
- 5Generate video · Video · Veo3 or Veo3-fast · Kie.aiImage-to-video animation of the logo.
- 6Poll and save · Check · Kie.ai + Google SheetsPolls for the finished clip and writes the output URL back to the sheet.
Missing: the AI Agent's system/user prompt is only narrated in prose, never quoted; would need reconstructing from the described fields. Referenced n8n template JSON not present locally.
RoboNuggets · R37 | LogoRobo - The Logo Animator AI Agent- 1
Cinematic ad15
Style-locked cinematic short film pipelineRunsa cinematic, style-locked short film/sequence: still keyframes generated first, kept ones animated as 3-5s shots, chained and assembled into a scored, upscaled master
fal-ai/flux-2-profal-ai/nano-banana-2bytedance/seedance-2.0 image-to-videofal-ai/kling-video/v3/pro/image-to-videofal-ai/lyria2 (music)fal-ai/topaz/upscale/videoMidjourney (operator pattern, no API)dark fairy talemythic graphic-novel epicsmartphone-UGC realismany style-contract-locked genre- 1Extract a style contract · Check · not stated (vision model, 'cinematographer' role prompt) · not statedFeed 5-10 reference stills to a vision model with a fixed extraction prompt (lens feel, lighting, palette, texture, contrast, blocking, atmosphere, era markers), merge into one locked 100-130 word paragraph.
Prompt
<role>You are a cinematographer analyzing a film still for exact technical reproduction.</role> <task>Name every visual parameter of this frame in precise, prompt-ready language.</task> <parameters> - Lens feel: focal length impression, depth of field, distortion, vintage or modern glass - Lighting: direction, source (practical, natural, artificial), hardness, ratio - Palette: dominant colors, shadow tint, highlight tint, saturation level - Texture: film grain, halation, bloom, digital cleanliness or analog dirt - Contrast: lifted or crushed blacks, dynamic range, exposure bias - Blocking: subject placement, foreground/midground/background layers - Atmosphere: fog, smoke, dust, weather, air quality - Era markers: anything dating the image (stock, fashion, technology) </parameters> <output_format>One paragraph, 100-130 words, comma-separated visual language ready to paste into an image prompt. No film titles, no director names, no camera brands. Describe what the image looks like, never what it references.</output_format>
- 2Generate still keyframes · Picture · FLUX.2 + Nano Banana (full automation) or Midjourney (operator pattern: human pastes/downloads, agent writes prompts) · fal.ai (API) or Midjourney (manual)Use the 'frame director' template: one story beat in, 10 distinct compositional keyframe prompts out, each appending the style contract verbatim. No motion yet.
Prompt
<role>You are a frame director. You take ONE story beat and return 10 distinct cinematic keyframe prompts: different compositions and camera positions of the same beat. Film stills, never posters, never portraits, never concept art.</role> <inputs>The beat; the locked style contract; the aspect.</inputs> <rules> - Append the style contract verbatim to every prompt. Never paraphrase it. - Each prompt explores a different compositional language: extreme low angle, overhead, foreground obstruction, reflection, silhouette, frame-within-frame, deep vanishing point, negative space, compressed telephoto, handheld close, POV, layered fg/mg/bg, subject partially hidden, environmental scale. - Respect the beat's role (setup, escalation, payoff) in how much you reveal. - Creature or reveal beats: show less — silhouette, fog, one fragment of anatomy. </rules> <output_format>10 entries: title (3 words), composition concept (1 line), the prompt (one paragraph ending with the style contract + aspect).</output_format>
- 3Review and re-round · Check · Claude CodeAgent reviews frames against the brief and writes the next round.
- 4Animate a kept keyframe · Video · fal-ai/kling-video v3 pro, bytedance/seedance-2.0 image-to-video · fal.ai (queue API) or HiggsfieldAnimate as an isolated 3-5 second shot using a 3-part motion grammar (named camera move + in-scene event + explicit 'no frozen figures'); chain shots by reusing the last frame of shot N as the reference for shot N+1.
- 5Run the fleet · Video · as step 4 · two custom Claude Code skills + a submit scriptThrottled parallel generation (capped concurrency, one retry max, 3-4 candidates per shot), logging every generation with kept/killed verdicts to a jsonl file.
- 6Assemble and score · edit · fal-ai/lyria2 (music) · ffmpeg + fal.aiAssemble from a text-file cut list (file, duration, mute flag) with ffmpeg; score with a music model briefed in 'movements'.
- 7Upscale the master · Upscale · fal-ai/topaz/upscale/video · fal.aiUpscale once on the final master (not per-shot).
AIOps · Turn Claude Code Into a Film Studio: Agentic AI Video Production Pipeline- 1
Full static + video ad pipelineRunsa finished ad: a hero product image, edited/refined, animated into a short video, voiced over, subtitled and upscaled to 4K
GPT Image 2Nano Banana 2Nano Banana Pro EditGPT Image 2 EditSeedance 2.0Happy Horse 1.0VEED SubtitlesBytedance Upscalerproduct hero shotlifestyle shotshort ad video- 1Generate hero image · Picture · GPT Image 2 · fal.aiGenerates the primary product studio shot from a fully written scene description (lighting, lens, label text, materials).
Prompt
A studio product shot of the Lumora Glow Drops bottle on a clean cream studio backdrop, soft three-point lighting with a clean highlight down the frosted glass, the brushed-copper dropper cap catching a small glint, the label reading "LUMORA" in a thin uppercase serif with "GLOW DROPS" beneath and a small line "Vitamin C Brightening Serum · 30ml", shot on an 85mm lens at f/4, sharp front to back, neutral warm grade.
- 2Generate alt lifestyle image (faster/cheaper option) · Picture · Nano Banana 2 · fal.aiAlternative hero-style shot for quick variations, same product in a lifestyle setting instead of studio.
Prompt
A bright lifestyle shot of the Lumora Glow Drops bottle on a sunny bathroom shelf beside a folded sage hand towel, water droplets on the tile behind it, punchy saturated daylight, a small "FRESH BATCH" tag hanging off the shelf, shot at eye level.
- 3Edit pass 1 — background swap (written adjustment) · edit · Nano Banana Pro Edit · fal.aiWhole-image edit instruction that changes only the background, explicitly telling the model what to keep unchanged.
Prompt
Keep the bottle, label, and lighting exactly as they are. Replace the cream background with a soft sage-green gradient. Change nothing else.
- 4Edit pass 2 — masked label refinement · edit · GPT Image 2 Edit · fal.aiSurgical, mask-limited edit that sharpens and deepens only the label text/foil inside a selected region.
Prompt
Within the masked label area only, sharpen the "GLOW DROPS" lettering and deepen the copper foil tone. Leave everything outside the mask unchanged.
- 5Animate to video (single shot) · Video · Seedance 2.0 · fal.aiAnimates the refined still into a slow camera-move video clip.
Prompt
A slow, smooth push-in toward the Lumora bottle, soft morning light warming the frosted glass and the copper cap, the scene calm and still, quiet room ambience.
- 6Add voiceover to the video · Voice · Happy Horse 1.0 (video-edit endpoint) · fal.aiEdits the existing video in place, preserving all original motion/framing, and adds a spoken voiceover line.
Prompt
Keep all original motion and framing, and add a soft, calm female voiceover that says, 'New Lumora Glow Drops, a vitamin C brightening serum. Wake up brighter.'
- 7Add subtitles · edit · VEED Subtitles · VEEDBurns styled subtitles onto the voiced video.
- 8Upscale to 4K · Upscale · Bytedance Upscaler · fal.aiUpscales the finished subtitled video to 4K for delivery.
Free online · How To Create Ads With AI- 1
Prompt-optimized Kling video with TTS voiceover, captions and 9-platform publishRunsa short cinematic video from a raw prompt idea, GPT-rewritten for Kling, with a matching TTS voiceover mixed in, burned-in captions, an auto-written social caption and YouTube-style title, and automatic publishing to 9 platforms
gpt-4o-mini (via GPT-4 labelled agent)gpt-4oKling 1.6 (PiAPI)OpenAI TTSshort cinematic videomulti-platform social video- 1Extract Prompt & Caption · Words · n8nPulls a raw videoPrompt and captionIdea from the trigger input via a Code node.
- 2Transform Prompt for Kling (GPT-4) · Words · gpt-4o-mini (bridge node; agent node labelled 'GPT-4' in the UI) · OpenAI (via n8n langchain agent + lmChatOpenAi bridge)Expands the short raw prompt into a cinematic, sensory-rich Kling-ready prompt, with 3 worked examples in the system prompt. The node is labelled 'GPT-4' but its underlying chat-model bridge node ('OpenAI Model Bridge') is actually configured to gpt-4o-mini.
Prompt
SYSTEM MESSAGE (verbatim, node 'Transform Prompt for Kling (GPT-4)'): ⚙️ System Instructions – AI Agent: You are a prompt optimization assistant for Kling AI, a cutting-edge text-to-video generation platform. 🎯 Mission: Transform brief user inputs into cinematic, high-quality video prompts rich in sensory details, spatial context, and dynamic motion—perfectly tuned for Kling AI. ✍️ Instructions: Expand short ideas into visually immersive scenes. Use concrete visual elements: colors, textures, lighting, motion, atmosphere, camera angles. Keep descriptions concise but vivid, like a scene from a movie script. Output only the final prompt—no explanations, no formatting tips. 🌟 Examples: User Input: “Rainy street at night” Final Prompt: Neon signs reflect in puddles on a dimly lit city street. Raindrops ripple across a parked motorcycle's chrome. A figure in a trench coat walks past flickering lights as steam rises from a nearby sewer grate. The camera tracks from ground level, capturing water splashing underfoot. User Input: “Mountain hike” Final Prompt: A hiker climbs a rugged mountain trail at sunrise, golden light hitting snow-capped peaks in the distance. Wind rustles through pine trees as birds soar overhead. The camera pans slowly from behind, revealing a vast valley below bathed in morning mist. User Input: “Cozy library” Final Prompt: A warm library room with wooden shelves lined with old books. Dust floats in the sunbeams streaming through tall windows. A cat naps on a leather armchair while a hand flips through a weathered novel. The camera gently rotates around the room, revealing small glowing lamps. USER MESSAGE: ={{ $json.videoPrompt }} - 3Generate Video via Kling API · Video · Kling 1.6 (PiAPI, mode std) · PiAPISends the transformed prompt to PiAPI's Kling video_generation endpoint.
Prompt
prompt field = Transform Prompt for Kling output verbatim; negative_prompt left empty
- 4Generate Voice-Over Script · Script · gpt-4o · OpenAI (via n8n langchain node)Writes a tight ~18-22 word voiceover script sized to exactly 7 seconds of narration, with one worked example.
Prompt
Task: Craft a short-form voice-over script that perfectly fits a 7-second video duration. The script should align with the following topic: {{ Extract Prompt & Caption captionIdea }} Requirements: Match the average spoken word count for a 7-second voice-over (approximately 18–22 words, depending on pacing). The script must be tight, impactful, and written in a natural spoken tone suitable for narration. No intro text or labels — return only the plain voice-over script with no headers or commentary. ✅ Example (For Context Only): Topic: "Why cold showers boost productivity" Generated Output: Cold showers shock your system awake, boost focus, and kickstart circulation—training your brain to embrace discomfort. - 5Convert Script to Audio (TTS) · Voice · OpenAI TTS · OpenAIConverts the voiceover script to speech via OpenAI's audio endpoint.
- 6Upload audio + merge with video + add captions · edit · Cloudinary + video-merge/caption APIUploads the TTS audio to Cloudinary, merges audio into the Kling video, then burns in subtitles/captions on the merged video, polling each stage for completion.
- 7Generate Social Caption from Voiceover · Words · gpt-4o · OpenAI (via n8n langchain node)Writes a short, value-driven social caption built directly on the voiceover script, one sentence per line, no quote characters.
Prompt
📲 Prompt: Caption Generator for Voiceover-Based Social Posts Your Role: Create a concise and engaging social media caption that builds directly on the voiceover script below: {{ Generate Voice-Over Script output }} 🧠 Caption Guidelines: Keep it short, compelling, and value-driven. Avoid generic motivational fluff — focus on real, actionable insight or highlight a problem/solution pattern. Structure: One sentence per line. Note: Do not use this character: " in the result. Return a single short paragraph with no line breaks and no special characters. Note: Do not use this character: " in the result. - 8Generate YouTube-Style Title · Words · gpt-4o · OpenAI (via n8n langchain node)Writes a punchy, curiosity-driven YouTube title under 70 characters based on the voiceover script.
Prompt
Act as a YouTube Title Expert. Based on the following video description: {{ Generate Voice-Over Script output }} Generate a short, punchy, and curiosity-driven YouTube video title that makes people want to click. Make it feel urgent, valuable, or surprising — and avoid generic or boring phrases. Keep it under 70 characters. Return only the title, no explanations. Note: The title must be free of special characters and the character ". Return only a plain text title. - 9Upload + publish to 9 platforms · publish · BlotatoUploads the final video to Blotato and publishes to Instagram, YouTube, TikTok, Facebook, Threads, Twitter/X, LinkedIn, Bluesky and Pinterest.
Missing: none — all generating prompts confirmed verbatim; only correction is the 'GPT-4' label on the Transform Prompt agent node actually resolving to gpt-4o-mini via its model bridge node
Free online · Create cinematic Kling AI videos with voiceover, captions, and auto-publish to 9 platforms- 1
Batch ad-variant video generator (1 template x CSV = N ads)RunsDozens of on-brand, message-matched performance/UGC-style video ad variants (15-30s) rendered from one Remotion motion-graphics template driven by a CSV of hooks/offers/CTAs, exported to every placement aspect ratio for A/B testing.
Remotion (code-driven video rendering, not a generative model)performance adUGC-style adA/B test variant set- 1Build the data-driven ad template · Video · RemotionAuthor ONE Remotion composition that reads a single `variant` object as props (hook, benefit, proof, cta, bg, accent) - nothing testable is hardcoded, and every animated value is a pure function of useCurrentFrame() so renders are deterministic frame-by-frame (no CSS transitions or JS timers, which desync the renderer).
- 2Fill a CSV test matrix (one variable at a time) · ScriptPopulate a CSV (id,hook,benefit,proof,cta,bg,accent) with one row per variant, holding every column constant except the single field under test - e.g. 5 hook strings against the same body to run a clean hook test, then the winning hook x 3 CTAs for a CTA test.
Prompt
VERBATIM worked example CSV rows (id,hook,benefit,proof,cta,accent,bg columns), each CTA mirroring its hook's promise: id=v01_problem | hook="Spending 2 hrs/day on reports?" | benefit="Build one in 60 seconds" | proof="Save 11 hrs/week" | cta="Save 2 hours — try free" id=v02_interrupt | hook="Your reporting tool is lying to you." | benefit="See the real numbers instantly" | proof="Save 11 hrs/week" | cta="See the real numbers — free" id=v03_curiosity | hook="Nobody talks about this reporting trick." | benefit="One click builds it" | proof="Save 11 hrs/week" | cta="Try the trick — free" HOOK -> CTA MESSAGE-MATCH WORKSHEET (fill one row per ad BEFORE writing the CSV; the CTA must pay off the exact promise the hook made): "Spending 2 hrs/day on reports?" (problem-first) -> "Save 2 hours — try free" "Your reporting tool is lying to you." (pattern interrupt) -> "See the real numbers — free" "Nobody talks about this reporting trick." (curiosity gap) -> "Try the trick — free" "Cut report time in half this week." (immediate benefit) -> "Cut it in half — free" "If you build reports by hand, stop." (direct callout) -> "Automate yours — free" Bad pairing to avoid: a "stop wasting 2 hours" hook ending on a generic "Shop now" — the promise and the payoff don't connect, so even a great hook converts poorly. TESTING ORDER (isolate one field per wave): 1. Hook (biggest lever on CPA, test 5-8 hooks against one identical body) -> 2. Offer/benefit (take the winning hook, vary the core promise) -> 3. CTA (winning hook+offer, vary the closing action) -> 4. Format/aspect (same winning ad, compare placements). Budget guide many teams use: ~60% to proven winners, ~30% to variations of winners, ~10% to fresh concepts; run each wave ~1 week / several hundred impressions per variant before declaring a winner. COMMON PITFALLS (from the source's checklist): changing two variables in one variant at once; hook clips in 9:16 running past ~60 characters (doesn't fit the frame in time); CTA hidden behind the platform's own UI overlay (captions/buttons/handle stack in the 9:16 bottom third); visual drift across variants (anything besides the tested field changing); animation flicker caused by non-frame-driven values (CSS transitions/JS timers instead of useCurrentFrame()).
- 3Render one MP4 per CSV row · Video · Remotion CLIA script converts each CSV row into a typed props JSON file, then Remotion renders one MP4 per row from the single template.
- 4Export every variant in every placement aspect ratio · Video · Remotion CLIRe-renders each variant at 9:16 (Reels/Stories/TikTok), 4:5 (Meta Feed), 1:1 (broad reach) and optionally 16:9 (YouTube), composing all critical content inside a shared 1:1 center-safe-zone so one master reframes cleanly to every crop.
- 5Verify before batch-rendering · Check · Remotion CLI (still renders)Renders cheap PNG stills at the hook, mid and CTA frames using the SHIPPED row's real props (not just defaults) to check fidelity (exact hook/offer/proof/CTA text, right brand colours) and artifacts (text overflow, off-canvas, CTA sitting in the 9:16 bottom third, missing fonts) - checks one representative variant in every target aspect before batch-rendering the full CSV x aspect matrix, since a layout bug repeats across every row.
Free online · GitHub — iart-ai/ad-video-skills: skill 'ad-creative-video'- 1
Beat-synced product launch / sizzle filmRunsA 15-60s premium launch/hype video (hook -> tease -> reveal -> feature montage -> end-card CTA) with the brand reveal snapped precisely to the music's drop, exported to 16:9/9:16/1:1.
Remotion (code-driven video rendering)GSAP (timeline sequencing)launch videosizzle reelteaser/trailer- 1Lock the soundtrack and mark the drop · CheckSound design leads picture: the music track is chosen and locked FIRST, its beats and exact drop timecode measured, before any visual is cut - the reveal must land exactly on the drop frame, never scored to a near-finished cut.
- 2Build the hook -> tease -> reveal arc · Video · Remotion + GSAPStructures the film across five beats scaled to the total duration (hook 0-3s stops the scroll, tease 3-9s builds curiosity beat-synced, reveal 9-13s snaps the product/logo to full opacity exactly on the drop with a micro-overshoot for impact, feature montage 13-25s cuts one benefit per shot on the beat, end card 25-30s holds logo+tagline+CTA clean for 2s+).
Prompt
VERBATIM shot-by-shot template, 30s film (scalable): 00:00–00:03 HOOK — Visual: single striking frame, extreme close detail/impossible motion. Motion: slow push-in or a fast whip that resolves; NO logo yet. Audio: sub-bass swell begins, sparse. Cut: on first audible beat at ~03s. 00:03–00:09 TEASE — Visual: fragments/silhouettes of the product, never the full thing. Motion: light leaks, partial reveals, parallax; tension rising. Audio: riser building toward the drop. Cut: quick cuts accelerating into the drop. 00:09–00:13 REVEAL (THE DROP at 00:09) — Visual: product/logo SNAPS to full on the drop frame. Motion: scale 1.18→1.0 over 0.18s + white flash-out; settle clean. Audio: drop hits + impact stinger on the exact frame. Hold: let it sit ~1.2s before moving. 00:13–00:25 FEATURE MONTAGE — Visual: one benefit per shot, ~0.6–1.0s each (≈8–10 shots). Motion: bold keyword pop + one supporting visual; consistent enter/exit. Audio: hard cuts on kick/snare; energy sustained. Pattern: KEYWORD + UI/product moment, repeat with rhythm. 00:25–00:30 END CARD — Visual: logo + tagline + CTA, centered, still. Motion: minimal — a soft settle, then hold ≥2s. Audio: music resolves; clean sonic tag. CTA: one action (URL or button), nothing competing. RE-PROPORTIONING: 15s = Hook 0-2s, Tease 2-5s, Reveal 5-7s, Montage 7-12s (4-5 shots), End card 12-15s. 60s = same hook (still ≤3s), longer tease + richer montage (14-16 shots, grouped in waves), end card ≥3s. The hook never gets longer — only the montage scales. BEAT-SYNCED REVEAL (GSAP, verbatim): DROP = 9.0s; tween `.tease` from {scale:1.0, opacity:.6} to {scale:1.08, opacity:1, duration:DROP, ease:'power1.in'}; at DROP, set `.logo` opacity 1 and tween scale 1.18→1.0 over .18s ease 'power3.out'; tween `.flash` opacity .9→0 over .25s ease 'power2.out'; fade `.tease` opacity to 0 over .12s; hold `.logo` at scale 1.0 for 1.2s dwell. - 3Cut the kinetic feature montage · edit · RemotionOne benefit per shot (~0.6-1.0s each), hard cuts on the beat, each shot pairing a bold keyword with a single supporting visual; keeps one consistent motion language (same enter curve, same exit) across the whole montage so speed reads as confidence, not chaos.
- 4Multi-aspect export · Video · Remotion CLIRenders the 16:9 master first, then reframes (not letterboxes) to 9:16 and 1:1, keeping the hero/logo/CTA inside a shared 1:1 center-safe-zone so one project serves every placement.
- 5Verify the reveal lands on the drop before encoding · Check · Remotion CLI (still renders)Renders frame-exact stills at the hook, reveal and end-card frames with the real shipped props, checking that the logo snaps precisely on the drop frame, text/colours are exact, and nothing sits in the 9:16 bottom third - verifies the reveal + end card in EACH target aspect before encoding all of them, since a reframe bug repeats across every aspect.
Free online · GitHub — iart-ai/ad-video-skills: skill 'launch-video'- 1
Animated testimonial / review quote videoRunsA 6-12s vertical/feed social-proof video: a customer quote revealed line-by-line in reading order, one key phrase highlighted, a star rating filling to the real (including fractional) score, and an author block (name/role/company/avatar) signing off - templated so one design renders a whole list of reviews.
Remotion (code-driven video rendering)testimonial videoquote cardreview videosocial proof clip- 1Structure the quote card anatomy · Video · RemotionLays out five elements in strict reveal order: quote-mark motif (signals 'this is a quote' before words appear, frame 0) -> quote body (staggered line-by-line reveal, ~6-10 frames apart, using spring() driven by useCurrentFrame so it's deterministic) -> one emphasis span on the single phrase that sells (lands after its line settles, via a left-to-right highlight sweep) -> star rating (fills left-to-right AFTER the quote reads, using a masked/wiped star layer so it handles fractional scores like 4.5 without a half-star sprite) -> author block (name/role/company/avatar slides up and holds still last).
Prompt
VERBATIM timing map (9:16, 30fps, ~10s): frame 0-8 quote mark scales in, card settles. Frame 8-48 quote body reveals line-by-line, each line entering at frame (8 + i*6). Frame 44-54 key-phrase highlight sweep (starts after its line settles, at hlStart = 8 + lines.length*6 + 4, sweeping over 10 frames). Frame 56-70 stars fill left-to-right to rating/max (starts at fillStart = hlStart + 14, over 14 frames). Frame 72-88 author block slides up and holds (starts at authStart = fillStart + 16). Frame 88-300 everything holds still (screenshot-able) — stillness in the back half is what makes a testimonial read as fact, not ad. Review prop schema (verbatim TypeScript): { quote: string (verbatim — never paraphrased or trimmed mid-sentence); highlight?: string (EXACT substring of quote, one phrase only); rating: number (0–max, halves allowed e.g. 4.5); max?: number (default 5); name: string (required — a quote with no name is not social proof); role?: string; company?: string; avatar?: string (real photo URL, or omit — never a stock face) }. Validation rule (verbatim function): validate(r) pushes errors for: empty quote; missing name ("no anonymous testimonials"); rating outside 0–max; and critically, `if (r.highlight && !r.quote.includes(r.highlight)) errs.push('highlight is not a substring of the quote')` — this blocks the render rather than silently dropping the emphasis. Line-splitting rule (verbatim regex logic): split the quote on clause boundaries (`. ! ? , ; —`) via `/[^.!?,;]+[.!?,;]?\s*/g`, then pack lines to ~32-34 chars; never split a short quote into one-word-per-line ("reads as a ransom note, not a testimonial"). Orphan-prevention: bind the last two words of a line with a non-breaking space via `s.replace(/ ([^ ]+)\s*$/, ' $1')` so no single word dangles alone on the last line. Star-rating rule: width-clip method (`rating/max*100%`) is exact for any decimal, no rounding to nearest half-star; never animate stars before the quote is read; never show more filled stars than the data supports — an invented score is worse than none; omit the star row entirely if a review has no rating. - 2Template x reviews - one video per quote · Video · Remotion CLIMakes every field (quote, highlight, rating, name, role, company, avatar) a prop on one Testimonial component, then renders one MP4 per review JSON file in a folder, so a list/CSV of reviews becomes a batch of on-brand videos with only the words and score changing.
- 3Verify reveal order and real score before batch-rendering · Check · Remotion CLI (still renders)Renders stills at the quote, stars and author frames with a real shipped review's props, checking the quote is verbatim with exactly one emphasized span, the star fill matches the REAL rating (not rounded), and name/role/company sit clear of the 9:16 bottom third - verifies one review in every target aspect (9:16/1:1/16:9) before batch-rendering the whole list.
Free online · GitHub — iart-ai/ad-video-skills: skill 'testimonial-video'- 1
Viral marketing ad video via VEO3 (JSON prompt template + GPT-4.1 rewrite)Runsa short (8s-ish) 16:9 or 9:16 marketing/ad concept video (e.g. gamer-room unboxing, car commercial) generated by VEO3 from a JSON-structured prompt
Veo3Veo3 FastGPT-4.1product unboxing adcar/brand commercial concept- 1Maintain a JSON prompt-template library keyed to real viral ads · WordsThe linked GitHub repo's prompt_library.py indexes 12 named templates (Amazon Gamer Room, Chewy Pet Supplies, IKEA, Jeep Basecamp, Lego Porsche, Louis Vuitton Handbag Reveal, Nike Ad, NYC Skyline, Rimowa Suitcase, Stridex Sneaker Boutique, Tennis Spin, Tesla Showroom), each imported from prompts.py and each traceable to the original viral X/Twitter post it was reverse-engineered from. Structure varies by template: some (e.g. Amazon Gamer Room) use a rich schema with a timeline of timestamped action+audio beats; others (e.g. Tesla Showroom) use a flatter scene/camera/lighting/motion schema.
Prompt
amazon_gamer_room = { "name": "Amazon Empty Room Assembly", "description": "Cinematic hyperlapse of Amazon box self-assembling gaming setup in empty Scandinavian room", "base_style": "cinematic, photorealistic, 4K", "aspect_ratio": "16:9", "room_description": "An empty, large, sunlit Scandinavian room with white walls and light wood floors.", "camera_setup": "A single, fixed, wide-angle shot. The camera does not move for the entire 8-second duration.", "key_elements": ["A sealed Amazon box with logo visible"], "assembled_elements": ["bed with dark blue duvet", "blue throw blanket", "metal table", "neon lamps", "wardrobe", "shelves", "gamer PC", "Gamer monitor and keyboard", "rug", "curtains", "gamer headsets"], "negative_prompts": ["no people", "no text overlays", "no distracting music"], "timeline": [ {"sequence": 1, "timestamp": "00:00-00:01", "action": "In the center of the otherwise empty room, a sealed Amazon box sits on the floor and begins to tremble gently.", "audio": "Low, subtle rumbling sound. The echo of a large, empty room."}, {"sequence": 2, "timestamp": "00:01-00:02", "action": "The box seams burst open with a puff of cardboard dust.", "audio": "A sharp 'POP' sound, followed by tearing cardboard."}, {"sequence": 3, "timestamp": "00:02-00:06", "action": "Hyper-lapse: From the fixed wide perspective, furniture pieces fly out of the box and assemble themselves, creating all the items from the 'assembled_elements' list.", "audio": "A cascade of satisfying, fast-paced ASMR sounds: whirring, clicking, wood snapping into place."}, {"sequence": 4, "timestamp": "00:06-00:08", "action": "The final piece\u2014the blue throw blanket\u2014gracefully lands on the newly formed bed. The room is now perfectly furnished and serene. All motion ceases.", "audio": "All chaotic sounds stop. A single, soft 'fwoomp' as the blanket lands. The sound of a furnished, quiet room."} ] } tesla_showroom = { "description": "Cinematic shot of a minimalist Tesla-branded crate magically opening to reveal a fully formed Tesla vehicle and an instantly assembled, sleek Tesla-themed showroom around it. No text.", "style": "cinematic", "camera": "fixed wide angle, with subtle zooms on key transformations", "lighting": "controlled, high-tech, transitioning from dim to bright and clean", "room": "empty futuristic space transforming into a minimalist Tesla showroom", "elements": ["Tesla-branded crate (glowing seams)", "Tesla vehicle (e.g., Model 3/Y/Cybertruck)", "charging station", "minimalist display panels", "sleek showroom furniture", "ambient lighting elements"], "motion": "crate panels retract smoothly and silently, car revealed, showroom elements rise/unfold precisely and rapidly", "ending": "pristine, inviting Tesla showroom with car as centerpiece", "text": "none", "keywords": ["16:9", "Tesla", "magic assembly", "showroom", "innovation", "futuristic", "no text", "clean design", "reveal"] } - 2Rewrite idea + template into a VEO3 prompt · Words · GPT-4.1 · OpenRouter (also supports Claude, DeepSeek as alternates)GPT-4.1 takes the user's own ad idea plus a chosen inspiration template and edits the structured JSON object to fit the new idea, keeping the original field structure unless told to change it. Output is a title + finished VEO3 prompt.
Prompt
System instruction: "Edit a structured prompt object (JSON) based on the provided user creative idea. Adapt the inspiration prompt to the user creative direction including style, room, background, elements, motion, ending, text, keywords. Do not modify any part of the structure unless the user explicitly requests it. Maintain original structure and intention unless asked otherwise." User message format: "Create a VEO3 prompt for this ad idea: [user_idea]. Follow and get instructions directly from this prompt: [inspiration_prompt]"
- 3Generate the video · Video · Veo3 / Veo3 Fast · Kie.aiSubmit the structured prompt to Kie AI's VEO3 API endpoint to generate the video, then poll the API for completion.
- 4Log results · CheckSave generated video URLs and metadata (title, prompt used, cost) to an Excel file for tracking/reference.
Missing: 10 of the 12 named templates in prompt_library.py were not individually transcribed (only amazon_gamer_room and tesla_showroom were opened in full) — would need to read the rest of prompts.py (381 lines total) to capture them
Free online · Create Viral AI Marketing Ads for Almost Nothing with Veo3 (+ kaymen99/viral-ai-vids prompt library)- 1
Video ad creative frameworks (hooks, scripts, formats, A/B testing) — Claude skillPartlythe creative planning layer for a video ad (chosen hook formula + script structure + format/aspect-ratio plan + A/B test plan), usable standalone as a Claude Code skill or to drive Creatify API video generation from a product URL
Meta/TikTok/YouTube Shorts/YouTube pre-roll/LinkedIn ad- 1Pick a hook formula · ScriptChoose from five proven hook approaches (Pattern Interrupt, Question, Bold Claim, POV, Stat/Authority) with platform recommendations and audience-fit guidance for the ad's opening line/frame.
Prompt
Pattern Interrupt: "Stop scrolling if you [audience identifier]" / "[Shocking visual] + Wait, what?" / start mid-action (unboxing, applying product, dramatic reveal) — best for TikTok, Reels, broad audiences. Question Hook: "What if I told you [counterintuitive claim]?" / "Why are [demographic] switching to [product category]?" / "Did you know [surprising statistic about their problem]?" — best for educational products, B2B, considered purchases. Bold Claim: "[Product] replaced my [expensive alternative]" / "I saved [specific amount] by switching to [product]" / "The only [product category] that actually [bold promise]" — best for competitive markets, DTC, price-advantaged products. POV Hook: "POV: You finally found a [product] that [solves problem]" / "POV: It's 6am and you're [using product in aspirational way]" / "Me when I discovered [product] exists:" — best for lifestyle products, Gen Z/Millennial, TikTok. Stat/Authority Hook: "[Number] people can't be wrong" / "Dermatologists say [claim] — here's why" / "After testing [number] [products], this one won" — best for health/beauty, B2B, premium products, skeptical audiences.
- 2Pick a script body structure · ScriptChoose from five complete message frameworks (Problem-Agitate-Solve, Feature Cascade, Social Proof Stack, Before/After, Day-in-the-Life) with timing breakdowns for 15s and 30s ad formats.
Prompt
Problem-Agitate-Solve (PAS): 1. Problem (2-3s) — name the pain point clearly. 2. Agitate (3-5s) — make it feel urgent, show consequences of not solving it. 3. Solve (5-8s) — introduce product as the natural solution. Best for pain-point-driven products, health/wellness, productivity tools. Feature Cascade: 1. Hero feature (3-4s). 2. Supporting feature 1 (2-3s) — credibility/utility. 3. Supporting feature 2 (2-3s) — secondary concern. 4. Proof (2-3s) — quick social proof or demo result. Best for tech, SaaS, feature-rich products. Social Proof Stack: 1. Testimonial/review quote (3-4s). 2. Visual proof (3-4s) — before/after, demo, results screenshot. 3. Volume proof (2-3s) — "Join [X] customers" or star ratings. 4. Urgency (2-3s) — limited offer or scarcity. Best for e-commerce, subscription products, competitive markets. Before/After: 1. Before state (3-5s). 2. Transformation moment (2-3s). 3. After state (3-5s). 4. How to get it (2-3s) — CTA. Best for beauty, fitness, home improvement, cleaning products. Day-in-the-Life: 1. Morning routine (3-4s). 2. Key moment (3-4s) — the 'aha' moment or primary use case. 3. Result/feeling (3-4s) — emotional payoff. 4. Invite (2-3s) — "Try it yourself" CTA. Best for lifestyle products, UGC-style ads, TikTok/Reels.
- 3Plan multi-format delivery · WordsUse the format decision matrix to pick aspect ratios and platform targets for the same core ad.
- 4Plan A/B tests · CheckApply the built-in A/B testing methodology to prioritize which variables to test first, with minimum spend thresholds and metrics to track.
- 5Generate the video (optional, via Creatify API) · Video · Creatify APIIf automating beyond the planning layer, use the Creatify API for URL-to-video generation, ad cloning, script generation, custom templates, product showcase videos, display banners, or batch product-catalog processing, following the batch-workflow architecture (async polling, webhook handling, rate limiting, error recovery).
Missing: step 5 (actual Creatify API video generation) still has no literal model-level prompt beyond passing a product URL/catalog — that stage is code/API orchestration (async polling, webhook handling, rate limiting) rather than a text-to-model prompt, so it correctly stays null; running the API stage end-to-end requires a Creatify API key
Free online · video-ad-generator (Claude Code skill)- 1
Realistic home-video film scene with a talking personPartlyA 20-second realistic film clip in two scenes: a woman talking to a friend's camera through a car door (her own voice, early-2000s camcorder look), then a six-shot road-trip montage cut to music ending on her line.
Claude Opus 5.5GPT Image 2.5Seedance 2.5home video / friend filmingroad-trip montage- 1Claude writes every prompt · Words · Claude Opus 5.5Give Claude Opus 5.5 one idea and a few screenshots of the style you like. It writes the character sheet, start frames, camera moves, emotions, dialogue timing, sound and music. The trick: tell it WHO holds the camera, WHERE they sit and HOW they move.
- 2Character sheet · Picture · GPT Image 2.5One wide sheet of the same woman: six turnaround views, five expressions and detail close-ups. Her face comes from a photo (image_1); two screenshots give the camera look only (image_2, image_3).
Prompt
Create ONE master character + frame reference sheet for the woman in <<<image_1>>>. She must be the exact same person in every panel: same face, bone structure, brown eyes, freckles across nose and cheeks, natural full lips, long sun-streaked light-brown hair parted in the middle with caramel highlights and loose flyaways. Same thin gold chain with small cross pendant, brown cord bracelets on right wrist, faded cropped white baby tee with pale blue and coral sun-and-waves print, light-wash frayed denim cutoff shorts. Identity is locked, no drift between panels. <<<image_2>>> and <<<image_3>>> are STYLE ONLY: copy their early-2000s compact camera home-video grade, not their person, not their text overlays. LAYOUT: one wide 16:9 sheet on a warm off-white paper background, thin grey panel borders, small clean black labels under each panel in a simple sans-serif. Three rows. ROW 1 — TURNAROUND (neutral studio, soft even daylight, plain light-grey backdrop, full body, relaxed standing A-pose, arms slightly away from body, bare feet or white sneakers): FRONT / 3/4 LEFT / LEFT PROFILE / BACK / RIGHT PROFILE / 3/4 RIGHT Clean true-to-life color here so wardrobe and proportions read clearly. Consistent height and scale across all six. ROW 2 — FACE + DETAIL (tight head-and-shoulders close-ups, same soft daylight, real skin with pores, freckles, peach fuzz, slight sheen, no retouching): NEUTRAL / SOFT PRIVATE SMILE / EXCITED DISBELIEF (brows up, eyes wide) / BIG GENUINE LAUGH (teeth showing, cheeks lifted) / PLAYFUL "I TOLD YOU" (head tilt, closed-mouth smile) / DETAIL PANEL split in four: eye macro, necklace and cross pendant, cord bracelets on wrist, tee print and denim fray. REALISM EVERYWHERE: real skin texture, natural asymmetry, individual hair strands and flyaways, real fabric creases, correct hands with five fingers, no beauty filter, no plastic skin, no CGI look. Sheet title top-left in small caps: "CHARACTER + FRAME SHEET". No other text beyond the panel labels. No watermark, no logos, no captions from the style references, no extra people.
- 3Start frame for scene 1 · Picture · GPT Image 2.5From the sheet, a candid photo taken from inside a parked car: she stands at the open door on a cliff road by the sea. The old-film look is written into the prompt, not a filter.
Prompt
Use the uploaded character sheet as the strict identity reference. Keep her exact face, bone structure, brown eyes, light freckles across nose and cheeks, natural full lips, and long straight light-brown hair with sun-kissed caramel highlights parted in the middle. Keep the thin gold chain necklace with the small cross pendant. SCENE: A candid photo taken from inside a parked silver sedan, looking out through the open rear passenger door. The door frame, the grey fabric headliner, the black door panel with its chrome handle, and part of the seat belt frame the shot on both sides. She stands just outside the open door on a rocky clifftop road above the ocean, both arms raised, one hand gripping the top of the door frame and the other resting on the roof edge, leaning slightly toward the camera. Relaxed, unposed, calm expression, looking straight into the lens. WARDROBE: Fitted cropped white cotton baby tee with a faded, sun-washed vintage graphic print in pale blue and coral red, slightly worn and soft. Light-wash low-rise denim cutoff shorts with frayed hems. Thin brown leather cord bracelets on one wrist. Hair slightly messy and windblown, a few strands across her face. BACKGROUND: Deep blue Mediterranean sea filling the frame behind her, sparkling sun glints on the water, a flat horizon line at chest height, pale hazy sky, dark rocks and sandy stones at the cliff edge below. LIGHT: Harsh direct midday summer sun, bright and slightly overexposed. Sun-warmed skin with a natural sheen and slight glow on the cheeks and nose, soft shadows under the chin, a faint reflection of light off the white car interior. CAMERA AND LOOK: Early 2000s point-and-shoot film photo, 35mm film, 28mm wide lens with slight barrel distortion at the edges, shot from chest height inside the car. Warm, faded, sun-bleached color grade, lifted blacks, low contrast, creamy highlights that gently blow out in the sky, slightly yellowed whites, soft halation around bright edges, fine visible film grain, a touch of softness and imperfect focus. Looks like a real candid snapshot from a summer road trip, not a studio shot. REALISM: Real skin texture with visible pores and freckles, no retouching, no plastic skin, natural flyaway hairs, real fabric wrinkles on the tee, real dust and wear on the car interior. Aspect ratio 4:3. No text, no watermark, no logos.
- 4Scene 1: she talks to the camera · Video · Seedance 2.5One locked 10-second take. The sheet locks who she is, the start frame locks the shot. Timed beats, her two spoken lines, facial emotions per word, camera-mic sound, a long list of what must not happen.
Prompt
=== REFERENCES === <<<image_2>>> is the exact START FRAME and master reference for the scene, car, location, framing, light and film grade. Frame one must match it pixel for pixel: camera inside a parked white sedan looking out through the open front passenger door, grey fabric headliner along the top, grey seat belt and B-pillar at screen-left, the open door with black panel and chrome handle at screen-right, rocky dry clifftop, glittering deep-blue sea, flat horizon, hazy pale bright sky, harsh midday sun. <<<image_1>>> is the IDENTITY LOCK for the woman. She must be this exact person in every frame: same face and bone structure, warm brown eyes, freckles across nose and cheeks, soft flush on cheeks, natural full lips, long sun-streaked light-brown hair with caramel highlights, middle part, loose flyaways. Jewelry: thin gold chain with a small cross pendant and a small round gold coin pendant; brown multi-strand cord bracelets on her left wrist. Wardrobe: faded cropped white baby tee with a vintage print (pale blue stripe band, coral-red sun disc with a small seagull, blue waves), light-wash frayed denim cutoff shorts. Use the sheet's expression panels (Soft Private Smile, Excited Disbelief, Big Genuine Laugh, Playful "I Told You") as her emotional range. Identity, hair, jewelry and wardrobe never change or morph. === CAMERA LOCK (MOST IMPORTANT) === One continuous 10-second take, 1080p. Zero cuts, zero camera switches, zero time jumps. ONE camera angle and ONE framing for the entire clip, exactly the composition of <<<image_2>>>. The camera does NOT rotate, pan, tilt, zoom, dolly, reframe or change position at any moment. The headliner, B-pillar, seat belt, open door, horizon line and sea stay in the same place in the frame from first frame to last. Only the woman moves inside this fixed frame. The camera is held by a friend sitting in the driver's side of the car, filming through the open passenger door on a 2003-era compact digital camera held steady at chest height. The only camera motion is subtle human handheld life: gentle breathing sway, faint micro-tremor, a few millimeters of drift. Never jerky, never stabilized-smooth, never a deliberate move. === 0.0-3.0s: STANDING, FIRST LINE === Frame one is identical to <<<image_2>>>. She hangs from the door frame and roof edge with both arms raised, leaning slightly toward the lens, calm direct gaze. Half a second of stillness, looking into the lens. Then a slow private smile starts at one corner of her mouth, like she has been dying to tell someone this. A small audible breath, then she speaks, soft and warm, relaxed American English, slightly raspy, intimate, like telling her best friend a secret: "I've been testing out the new Opus 5.5 model..." (pronounce "Opus five point five") On "testing out": eyebrows lift slightly, eyes brighten and widen a touch. On "Opus 5.5": a tiny proud head tilt, quick glance down and back to the lens. Eyes glossy and alive, one natural blink. Sea breeze pushes loose strands across her face, the tee flutters, sun glitter shimmers on the water. === 3.0-5.0s: SITTING DOWN INTO THE FRONT SEAT (SAME FRAME) === Still smiling, she lowers her arms from the door frame, turns slightly and sits down sideways onto the front passenger seat, right in front of the lens, one hand on the seat edge for balance. She swings her legs in and settles, then turns her upper body toward the camera. Her head and shoulders now sit lower and closer in the same fixed frame, face near the center, with the open door, sea and sky still visible behind her exactly where they were. Hair falls forward as she sits, then swings back. Tee and denim crease naturally, pendants swing and settle, the seat gives under her weight with a soft creak. The camera does not follow her. It stays locked on the same angle. Autofocus alone reacts: a brief half-second hunt as she moves closer, one pulse, then it locks crisp on her eyes. Exposure dips slightly as her face moves into the shaded interior light, then settles. She tucks her hair behind her ear with her fingertips, glances down, lets out a small breathless laugh, then lifts her eyes straight back into the lens. === 5.0-8.5s: SEATED, THE PAYOFF LINE === Seated, close to the lens, turned toward it, she leans in slightly, voice lower and slower, almost a confession, then building: "...and the shit you can do with it is absolutely insane." Emotion arc: "and the shit you can do with it" — Excited Disbelief: slow side-to-side head shake, eyes narrowing slightly, lips pressing between words like she still can't believe it. A tiny inhale beat. "is absolutely insane" — her face opens into the Big Genuine Laugh: brows lift, eyes widen and sparkle, cheeks rise, teeth visible, a real laugh breaks on the last word. Her free hand lifts in one loose open-palm gesture near her collarbone and drops. Lived, not performed. Real warmth, real excitement, soul in the eyes. Natural blinks, tiny eye drifts off lens and back, breath visible in chest and shoulders. Exact lip-sync with natural mouth shapes and tongue movement. Camera stays on the same angle; only the faint handheld breathing sway continues, with a tiny extra bob on her laugh as if the friend laughed too. === 8.5-10.0s: THE HOLD === The laugh settles into Playful "I Told You": soft closed-mouth smile, head tilted, one shoulder lifting in a small shrug. One slow blink, a quiet laugh through the nose, she bites her lower lip to hold back the smile. A breeze through the open door moves a few strands across her cheek; she leaves them. Same fixed angle to the last frame. End on her warm face mid-smile. === FILM LOOK (MATCH <<<image_2>>>) === Early-2000s compact camera home footage. Warm, faded, sun-bleached palette; slightly yellowed whites; lifted milky blacks; low contrast; soft creamy highlights clipping and blooming in the sky and sea; gentle halation glow around bright edges and her hair; fine organic grain and light sensor noise in the shadows; slight overall softness with sharp eyes; mild barrel distortion at the frame edges; subtle chromatic fringing on high-contrast edges; faint highlight smear on the brightest sea glints. Harsh direct sun outside, soft warm bounce light on her face once seated inside the car, warm rim light from the open door catching her hair. Nostalgic summer road-trip memory. === REALISM === Real skin: pores, freckles, sun sheen on nose and cheekbones, slight flush, faint under-eye texture, peach fuzz glowing in backlight, natural asymmetry. No retouching, no smoothing, no beauty filter. Real hair physics with individual strands and flyaways. Fabric wrinkles, denim fray threads moving. Jewelry catches light as she moves. Five fingers on every hand, correct anatomy. === AUDIO === On-camera built-in mic only. Her voice clear and natural, a little outdoor air while standing, warmer, closer and more intimate once seated. Small breaths before lines, soft laughs, a seat creak and fabric rustle as she sits, steady sea breeze, faint distant waves, a hint of wind buffeting the mic. No music. === NEGATIVES === No camera rotation, no pan, no tilt, no zoom, no reframing, no angle change, no scene change. No cuts, no jump cuts, no slow motion, no speed ramps, no gimbal smoothness, no dolly or crane moves. Camera never visible, no selfie arm. No extra people. No face, hand, jewelry, tee print or clothing morphing. No plastic skin, no lip-sync drift, no frozen face between lines, no anatomy errors, no duplicate fingers. No on-screen text, captions, date stamps, logos or watermarks. No lens flare. Film look consistent from first to last frame.
- 5Pictures for scene 2 · By hand · GPT Image 2.5Two more pictures: an aerial start frame of the white car on the coast road, and an inside-the-car picture of both people (the man is new, with his own look).
- 6Scene 2: six-shot montage · Video · Seedance 2.5Six shots in 10 seconds, each cut on a drum hit of made-up surf-rock music: drone follow, side tracking, low pass, dashboard, back seat, then her closing line.
Prompt
=== REFERENCES === <<<image_2>>> is the exact START FRAME: aerial view of the winding coastal cliff road, the older white 2000s sedan driving along it, dry golden hillside at left, green coastal scrub, sandy beach, rocks and rolling white surf at right, beach houses and palms on the far hills, sun low over the sea with a glittering sun path. Frame one matches it pixel for pixel. <<<image_1>>> is the master reference for the car interior, both identities, wardrobe and interior light. Woman: long sun-streaked light-brown hair, freckles, brown eyes, thin gold chain with small cross and coin pendant, brown cord bracelets on her left wrist, white baby tee with a coral sun, seagull and blue waves print, seat belt across her chest, passenger seat. Man: short dark brown hair, tanned olive skin, short full beard and mustache, faded teal-blue palm-leaf print shirt open at the collar, light-wash jeans, silver watch, driving. Dark grey fabric headliner, small sunroof, black door panels, grey seats, simple older dashboard. The white sedan is the same car in every shot. Identities, wardrobe and car never change or morph. === FORMAT === 10 seconds, 1080p, six shots cut on the drum hits of the music, like a real summer road-trip film shot by a friend with a drone and a handheld camera. Every shot has real camera life: natural sway, small corrections, organic motion, never CGI-smooth. === SHOT 1 — 0.0-2.0s: AERIAL FOLLOW === Starts exactly on <<<image_2>>>. The drone glides forward and slowly descends behind the white sedan as it cruises along the curve. Wheels turning, soft shadow sliding over cracked asphalt and the faded double yellow line. Waves roll in and break along the beach, the sun path shimmers, scrub sways in the sea wind. Slight real drone float and drift, a soft flare passing as the angle shifts. === SHOT 2 — 2.0-3.5s: SIDE TRACKING === Cut on the beat. Camera shooting from a second car driving alongside, level with the sedan, sea side of the road. The white sedan fills the middle of the frame in clean profile, wheels spinning, tiny road bumps rocking the body. Through its open windows her hair is flying out. Foreground scrub whips past as a blur, the sea and sun glitter slide behind. Gentle handheld bounce from the camera car. === SHOT 3 — 3.5-5.0s: LOW FRONT PASS === Cut on the beat. Camera low at the road edge, just above the asphalt, facing oncoming traffic. The sedan comes around the curve toward the lens, headlights off, windshield catching a streak of sun, and rushes past close on the left. Dust and road grit kick up, the camera shakes slightly from the air push, motion blur on the body and wheels. Cliffs and sea behind. === SHOT 4 — 5.0-6.5s: INTERIOR FROM THE FRONT === Cut on the beat. Camera mounted low on the dashboard facing back through the windshield area toward both of them, a little dust and a soft reflection on the glass edge. Both in frame side by side: him at the wheel on the left of frame, calm, eyes on the road, one arm on the open window; her on the right, head resting back, eyes half-closed, hair blowing across her face. Sun patches slide across them as the road curves, sky and sunroof light filling their faces. Car vibration gives a subtle constant micro-shake. === SHOT 5 — 6.5-8.0s: INTERIOR FROM THE BACK SEAT === Cut on the beat. Handheld from the middle of the back seat, between the two headrests, looking forward through the windshield. The winding coastal road and the bright sea ahead, the dashboard, his hand on the wheel and the silver watch, his eyes glancing up in the rear-view mirror. Her hair streams back toward the lens in the wind. She reaches over and rests her hand on his forearm on the gear shift, cord bracelets catching the light. The camera rocks gently with the road. === SHOT 6 — 8.0-10.0s: PASSENGER CLOSE, THE LINE === Cut on the beat to the framing of <<<image_1>>>: camera close to her on the passenger side, wide lens, looking across toward him. Window down, hair whipping across her face in real strands. She turns to the lens, pushes the hair aside, a slow warm smile breaks, and she says softly, relaxed American English, a little breathy over the wind: "Told you... it has no limits." Exact lip-sync, natural mouth shapes. He glances over at her with a small amused smile, then back to the road. She holds the smile, eyes sparkling, a soft laugh through the nose, hair crossing her cheek. End on her face mid-smile, the sea bright in the window behind. === LIGHT AND LOOK === Exteriors: warm low afternoon sun, glittering sea, soft haze on the far hills, long car shadow. Interiors: match <<<image_1>>>: clear natural afternoon daylight, true blue sea and pale blue sky through the windows, direct sun landing as soft warm patches on faces and arms with natural shadows, cabin in cool shade, soft sunroof light. Overall a natural cinematic 35mm travel-film feel with a light vintage touch: gentle contrast, soft natural saturation, real skin tones, very light fine grain, slight softness at the edges, mild wide-angle distortion inside the car. Color stays consistent across all six shots, no heavy yellow tint, no sepia. === REALISM === Real skin on both: pores, freckles, light sheen, peach fuzz, beard texture, natural asymmetry, no retouching. Real hair physics in strong wind. Real asphalt cracks, dust and grit. Real ocean motion and foam. Fabric creases, seat belt twist, dust on the dashboard, a faint smudge on the windshield. Correct hands with five fingers. Car proportions, color and interior identical in every shot. === MUSIC AND SOUND === Warm, hazy indie surf-rock track around 96 BPM: a lazy reverb-soaked electric guitar line, soft brushed drums, mellow bass, nostalgic summer mood, no vocals. Every cut lands exactly on a drum hit. Under the music: soft wind and distant waves on the aerial; engine hum and wind rush on the side tracking; a loud whoosh, tire roar and gravel crunch on the low pass; inside the car, wind buffeting through the open windows, road hum and tire noise. On Shot 6 the music dips slightly so her line sits clear and close, then swells back under her smile to the end. === NEGATIVES === No text, captions, subtitles, logos, badges, license plate text or watermarks. No extra people. No other cars except the implied camera car, which is never visible. No face, hand, jewelry, shirt print or car morphing. No plastic skin, no lip-sync drift, no anatomy errors, no duplicate fingers. No slow motion, no speed ramps, no CGI-smooth camera, no warping road or water. Car color and model identical in every shot.
- 7Join and caption · editJoin the two scenes and add word-by-word captions (seen in the video, not described in the post).
Missing: Claude's own instruction is not shown; the scene 2 pictures and the man's sheet have no prompts; the edit is not described.
Outside posts · X post by @abxxai, 2026-09 (x.com/abxxai/status/2102775755646337530)- 1
Veo3 branded ad video (viral-prompt remix)Partlyone branded 8-second Veo3 ad video (vertical or horizontal) built by customizing a viral "$100k ad" Veo3 prompt for a given brand from a short creative brief.
Veo3 FastVeo3 QualityGPT-4.1vertical 9:16 adhorizontal ad- 1Set direction · By hand · n8nUser manually fills an edit-fields node: a creative brief (e.g. "I want an ad for Volkswagen"), a prompt-inspiration pasted from the Notion viral-prompt library, aspect ratio, and model choice (Veo3 Fast/Quality). Trigger is manual click, not scheduled.
- 2Refine prompt · Script · GPT-4.1 · OpenAIAn AI agent (with a think tool and access to prior Google Sheet entries to avoid duplicate ideas) rewrites the inspiration prompt for the requested brand, producing a title and final Veo3 prompt in a forced structured-output format.
- 3Generate video · Video · Veo3 Fast or Veo3 Quality · Kie.aiHTTP POST sends the refined prompt, model choice and aspect ratio to Veo3 via Kie.ai; request is queued.
- 4Wait and poll · Check · Kie.aiWaits then polls the task by ID (get request) until a result video URL is returned.
- 5Log to sheet · Words · Google SheetsAppends a row with ID, title, prompt, production status and final video URL to a Google Sheet.
Missing: the viral ad prompts themselves (the actual reusable IKEA/Tesla/sneaker/Jeep/Remoa prompts) sit in an external Notion library, not harvested into this file; the ads-factory agent's system prompt not harvested.
RoboNuggets · n9 - The Veo3 $100k Ads Factory- 1
Claude Code + Seedance 2.0 cinematic promo/motion-graphics videoPartlyA cinematic, motion-graphics-style promotional video (screenshot-driven, e.g. an app/SaaS product ad) generated end-to-end by a Claude Code agent driving the Seedance 2.0 video model.
Seedance 2.0 (ByteDance)reference(screenshot)-to-video promo cliptext-to-video clipimage-to-video clip (first-frame mode, added via docs)batched multi-project promo reel- 1Install Seedance skill and connect API · By hand · Claude CodeClaude Code is given the RoboNuggets Seedance skill (git clone github.com/robonuggets/seedance-skill), which documents how to call Seedance 2.0 and good prompting practice. The user's model-provider API key (fal.ai in the walkthrough) is supplied and stored in a .env file so Claude Code can call the model directly.
Prompt
I'm giving you a skill called Seedance Skill. Get the files with: git clone https://github.com/robonuggets/seedance-skill Follow the README to set it up. Recommend how it would best apply to our setup in plain language, and ask me a few questions to clarify my intent before you install it.
- 2Propose creative directions · Script · Claude (Claude Code) · Claude CodeGiven a folder of product screenshots, Claude Code (using the skill's prompting guidance) proposes multiple distinct creative directions for a promo video (e.g. an iPhone-screen animation, a tablet version, a hexagon/particle motion-graphics concept) for the user to pick from before generating.
Prompt
Hey, so these are screenshots for [app]. Can you suggest to me a couple of directions on how we can make a promotional video for these using Seedance? Can you give me three options so that I can select from those?
- 3Generate reference-to-video clip(s) · Video · Seedance 2.0 · fal.aiClaude Code writes the Seedance prompt for each chosen direction (referencing the supplied screenshots) and calls Seedance 2.0's reference-to-video mode via the provider API, polling until each clip is ready and saving the output MP4s.
- 4(Optional) Batch produce across multiple projects · Video · Seedance 2.0 · fal.aiGiven a list of project websites/App Store links instead of local screenshots, Claude Code browses each site, sources hero imagery, crafts a Seedance prompt per project, and generates a promo clip for each — producing a small reel of intro/title videos in one pass.
- 5(Optional) Add image-to-video capability to the skill · By hand · fal.aiThe provider's LLM-ready docs page for Seedance's image-to-video (first-frame) mode is copied and given to Claude Code to extend the installed skill, so future runs can also use a single starting image instead of only reference-to-video.
Prompt
I want to include this image to video capability as part of this Seedance skill. Here's the documentation from [provider]. Can you integrate it, please?
Missing: Exact Seedance prompt text as actually sent to the API is not shown verbatim, only paraphrased narration; the skill file itself (github.com/robonuggets/seedance-skill) was not read/harvested, so its internal prompting rules and default settings (resolution, duration, aspect ratio defaults) are unconfirmed; small on-screen text rendering is flagged as unreliable, especially at low source-image resolution.
RoboNuggets · n46 - Claude + Seedance 2.0- 1
Cinematic product advert agent systemPartlya finished ~30-second cinematic product advert (multi-scene, image-to-video, with music and voiceover tracks) built from a character/setting/product core elements board, AI-scripted scenes, then manually edited together
Nano Banana ProVeo 3.1 (fast)SunoElevenLabs Multilingual v2cinematic ad9:16 short-form ad- 1Core 'money shot' hero image (manual, iterative) · Picture · Nano Banana Pro · Gemini (chat interface)Interactively refines one hero image of character + product in the target scene inside a Gemini chat.
- 2Core Elements Board generation · Picture · Nano Banana Pro · Kie.aiPlaces character (multi-angle), setting (multi-view) and product (multi-angle incl. macro) into one wireframe-templated board image.
Prompt
Create a Brand board based on the layout structure provided in Image 1. The final image should be a portrait 9:16 aspect ratio. CHARACTER Section: Main Image: Fill the large "Main character image" box on the left with a high-quality, detailed portrait of the subject provided in Image 2. Wireframe Replacement: Replace the two wireframe head placeholders at the top right with close-up headshots of the subject from Image 2. The left box should be a direct front view, and the right box should be a side profile view, matching the angles suggested by the wireframes. Pose Variations: Fill the two rectangular boxes below the headshots ("Different angle and pose 1" and "2") with distinct views of the subject. Maintain visual consistency with the subject in Image 2 (same features, clothing, style), but vary the pose and expressions significantly between these two boxes to show range. SETTING Section: Main Image: Fill the large "Main setting image" box with the environment provided in Image 3. Imaginative Variations: Fill the two smaller boxes on the right ("Different view 1" and "2") with radically DIFFERENT architectural or spatial perspectives that imaginatively expand on the environment shown in Image 3. These views must show different parts of the location, not just a slight rotation of the main view, while maintaining the same atmosphere and visual style. Different view 1 should be different from Main setting image and Different view 2 Different view 2 should be different from Main setting image and Different view 1 PRODUCT Section: Accuracy Requirement: For all product images, ensure the item is represented with extreme high fidelity. The rendering must accurately reflect the exact materials, textures, finishes, and design details present in the source image. Lighting should be photorealistic to highlight these details. Main Image: Fill the large "Main product image" box with a clear, attractive main view of the item provided in Image 4. Detailed Variations: Fill the two smaller boxes on the right with two distinct, alternative angles of the product from Image 4. Include at least one extreme macro close-up that emphasizes the texture and build quality of the item. Different view 1 should be different from Main product image and Different view 2 Different view 2 should be different from Main product image and Different view 1 Image labels: Image 1 - Template Image 2 - Photo of the character Image 3 - Photo of the setting Image 4 - Photo of the product - 3Creative direction brief (via custom GPT) · Words · 'Creative Direction Ideator GPT' (custom GPT) · OpenAI (ChatGPT custom GPT)Uploads the elements board plus a request to get a full creative summary, music-track description and a 5-scene breakdown.
- 4Load inputs to Airtable · Words · AirtableUploads the core image, elements board and written creative direction into an Airtable project row and scenes tab.
- 5Generate scene prompts + music prompt + script · Words · LLM agent (specific model not stated) · OpenRouter or Google AI StudioGiven the creative direction, elements board and core image, outputs per-scene starting/ending-image prompts, a transition/video prompt, a music prompt and a voiceover script.
- 6Review/approve scene prompts · Check · AirtableHuman inspects each scene's image prompts, edits as needed, marks approved.
- 7Generate start/end frame images per scene · Picture · Nano Banana Pro · Wavespeed or Kie.aiRuns approved prompts through Nano Banana Pro to create starting and ending frame images per scene.
- 8Review/regenerate images · Check · Airtable + Wavespeed/Kie.aiHuman inspects start/end frames, edits prompt text directly if needed, reruns unapproved rows.
Prompt
the camera zooms in close to her right eye
- 9Generate video per scene · Video · Veo 3.1 (fast) · Kie.aiAnimates the transition between the approved start and end frame images.
- 10Generate music · music · Suno · Kie.aiRuns the music prompt through Suno, producing 2 candidate tracks.
- 11Generate voiceover · Voice · ElevenLabs Multilingual v2 · Wavespeed (elevenlabs/multilingual-v2)Runs the approved script and chosen voice ID through ElevenLabs.
Prompt
Your eyes create the world you pull into reality
- 12Manual video edit (final assembly) · edit · none (external NLE: Premiere/CapCut/Canva/DaVinci)Human edits scene clips, music and VO together, trims unusable clip tails, adds logo. Explicitly not automated by design.
Missing: Creative Direction Ideator GPT content not harvested (external link only); scene-prompt agent's system guideline not quoted verbatim; full ad script not quoted verbatim (only one line). Final assembly is explicitly manual by course design.
RoboNuggets · R50 | The Cinematic Adverts Agent System- 1
Winning-ad repurposer / creative cloner (n8n)Partlya new multi-scene ad video that replicates the shot-by-shot structure, camera work and pacing of an existing viral or winning reference ad, but swapped to the user's own character/product/brand, plus matching background music
Gemini 3Nano Banana ProKling v2.6 Pro (image-to-video)Sunocinematic adcloned ad structure- 1Source winning reference ad (manual research) · By hand · MagicBrief / InstagramFinds and downloads a high-performing ad via MagicBrief (filter by industry/theme/winners/runtime) or Instagram + SortFeed extension.
- 2Load project inputs to Airtable · Words · AirtableUploads the reference video, character/product reference images and a written swap request (what to swap, what to keep, scene count, text-overlay flag, aspect ratio).
Prompt
have the Asian model in image one wear the different clothes featured in image two ... ignore the clothes from the Prada ad ... create only eight scenes ... no text overlays
- 3Analyze reference video (scene breakdown) · research · Gemini 3 · Google AI StudioFeeds the reference video into Gemini, which breaks it into scenes using the SEALCaM framework and describes the music.
- 4Generate scene prompts from analysis + creative direction · Words · LLM agent · OpenRouter or Google AIGiven the video analysis and the swap request, generates per-scene image prompt, video prompt, music prompt and script, all following the SEALCaM field order.
Prompt
SEALCaM Framework Definition: S – Subject: What the camera is optically prioritizing within the frame. E – Environment: The physical or constructed space surrounding the subject. A – Action: Observable motion within the frame, including subject and camera movement. L – Lighting: The lighting setup and exposure characteristics shaping the image. Ca – Camera: The capture device, lens choice, framing, angle, and movement strategy. M – Metatokens: Technical and stylistic capture cues related to production quality and presentation. Further guidelines: S: Use shot-focused terminology: primary subject, secondary subject, foreground element, background element E: Use production terms: location type, set design, spatial depth, background treatment A: Use blocking and motion terms: subject movement, camera movement, environmental motion L: Use lighting terms only: key light, fill, rim, practicals, contrast ratio, exposure level, color temperature Ca: Must include: Camera type (cinema camera or stills camera, e.g. ARRI Alexa, RED, Sony FX series, DSLR, mirrorless); Lens type and focal length (e.g. 35mm prime, 85mm portrait lens); Framing and angle (wide, medium, close-up; eye-level, low-angle); Camera motion (locked-off, handheld, dolly, pan, tilt, tracking) M: Use visual production qualifiers only: realism level; texture and grain; motion cadence; render or capture quality; platform or delivery cues if visible SEALCaM fields must appear in the exact order: S, E, A, L, Ca, M
- 5Review/approve scene prompts · Check · AirtableHuman inspects prompts before generating images.
- 6Generate start-frame images per scene · Picture · Nano Banana Pro · WavespeedRuns each scene's image prompt with character/product reference images through Nano Banana Pro.
- 7Generate video per scene · Video · Kling v2.6 Pro (image-to-video) · WavespeedAnimates each starting frame using the scene's video prompt.
- 8Generate music · music · Suno · Kie.ai (suno-api)Runs the agent-generated music prompt through Suno, producing 2 tracks.
- 9Manual final edit · edit · none (external NLE: Premiere/Canva/CapCut)Human stitches scene clips and music together. FFmpeg partial automation is possible but not used for low-volume runs.
Missing: the full Gemini video-analysis prompt and the scene-generation agent's system guideline are both referenced as separate external documents (linked, included in the n8n template) not present in this local lesson file.
RoboNuggets · R51 | The Creative Cloner AI Agent System- 1
Antigravity creative cloner (agentic, scripted)Runsa video clone of a reference ad, built as a packaged Python-script agent skill (no n8n): recreates a viral ad's structure with the user's own product/character swapped in, concatenated with a music fade
Gemini (video analysis)Nano Banana ProKling 2.6cloned ad structure- 0Create folder structure · By hand · Antigravity agentSets up inputs/, outputs/, tools/, .agent/ (skills/, workflows/, .env) folders.
Prompt
Create a project folder structure for a "Creative Cloner" agent that recreates viral videos with new products/characters. Create the following folders in my current workspace: 1. `inputs/` - Where I'll put reference videos and photos 2. `outputs/` - Where generated images, videos, and music will go 3. `tools/` - For any helper scripts 4. `.agent/` - For agent configuration Inside `.agent/`, also create: - `skills/` folder - `workflows/` folder - An empty `.env` file (for API keys later) Just create the folders and confirm what you made.
- 1Analyze video · research · Gemini API · Google AI StudioUploads the reference video to Gemini and breaks it into a SEALCaM-framework YAML scene analysis (subject, environment, action, lighting, camera, duration, music).
Prompt
Create a tool script called analyze_video.py in the tools/ folder. This script should: 1. Use the Gemini API to analyze a video file 2. Upload the video to Gemini, wait for processing, then analyze it 3. Use the SEALCaM framework to break down each scene The analysis should extract for each scene: - Scene description - Subject (main focus) - Environment/setting - Action (what motion occurs) - Lighting style - Camera angle/movement - Approximate duration Also capture any music/sound description. Output the analysis in YAML format so it's easy to parse later. The script should: - Load the GEMINI_API_KEY from the .agent/.env file - Accept a video path as command line argument - Optionally accept an output path to save the analysis - Print progress messages so the user knows what's happening Use the google-genai library and dotenv for environment variables.
- 2Create prompts · Words · not stated (Claude Code/Antigravity agent)Converts the video analysis plus a reference product/character photo into per-scene image and video prompts.
Prompt
Based on the video analysis we just did, create prompts for recreating this video with my new product/character. Look at my reference photo in inputs/ - this is what I want to swap into the video. For each scene from the analysis, create: 1. An IMAGE PROMPT - for generating a static frame with NanoBanana Pro 2. A VIDEO PROMPT - for animating that frame with Kling 2.6 Each prompt should: - Keep the same environment, lighting, and camera angle from the original - Replace the original subject with my new product/character - Describe the action/motion clearly for the video prompt Format the prompts in a clear list so we can use them in the next steps.
- 3Set up Airtable tracker · Words · AirtableCreates a Scenes table (Project Name, scene, start_image_prompt, video_prompt, start_image, scene_video) and logs the prompts into it.
Prompt
Set up Airtable to track our creative cloning project. Create a table called "Scenes" with these columns: [Project Name, scene, start_image_prompt, video_prompt, start_image (attachment), scene_video (attachment)] Then log the prompts we just created into this table. I'll give you my Airtable API key and base ID.
- 4Generate images · Picture · nano-banana-pro · Kie.aiUploads reference images to Kie.ai's file hosting, then generates each scene's start-frame image via Nano Banana Pro, logging the result to Airtable.
Prompt
Create a Python script called generate_images.py in the tools/ folder. This script should generate images using the NanoBanana Pro model via Kie.ai API. ⚠️ IMPORTANT: Each image costs $0.09 to generate. Always show the user the cost and ask for approval BEFORE running the script. Example: "This will generate 3 images at $0.09 each = $0.27 total. Ready to proceed?" The script needs to: 1. Upload reference images to Kie.ai's file hosting to get a public URL - Uses the same KIE_API_KEY (no extra keys needed!) - POST to https://kieai.redpandaai.co/api/file-stream-upload - Send the file as multipart form data - Returns a fileUrl in the response Example: files = {'file': ('image.jpg', open(filepath, 'rb'))} response = requests.post(url, headers={"Authorization": f"Bearer {KIE_API_KEY}"}, files=files) file_url = response.json()["data"]["fileUrl"] 2. Generate images using NanoBanana Pro: - POST to https://api.kie.ai/api/v1/jobs/createTask - Model: "nano-banana-pro" ⚠️ CRITICAL: The "image_input" field MUST contain the uploaded reference image URL! This is how NanoBanana Pro knows what product/character to include. Example payload: { "model": "nano-banana-pro", "input": { "prompt": "A fox holding a red bottle in a forest", "image_input": ["https://kieai.redpandaai.co/files/..."], "aspect_ratio": "9:16", "resolution": "2K" } } - Poll for completion, then get the result URL 3. Log the generated image to Airtable: - Find the matching scene record by Project Name and scene name - Update the "start_image" field with the image URL as an attachment Load API keys from .agent/.env: - KIE_API_KEY (used for BOTH file upload and image generation) - AIRTABLE_TOKEN - AIRTABLE_BASE_ID Print friendly progress messages at each step. - 5Generate videos · Video · kling-2.6/image-to-video · Kie.aiFetches the image URL from Airtable and generates each scene's video via Kling 2.6, logging the result and downloading it locally.
Prompt
Create a Python script called generate_videos.py in the tools/ folder. This script should generate videos using Kling 2.6 (image-to-video) via Kie.ai API. ⚠️ IMPORTANT: Video generation costs money. Always ask the user for approval BEFORE generating any videos. The script needs to: 1. Get the image URL from Airtable: [fetch scene record, get start_image attachment URL] 2. Generate video from that image using Kling 2.6: - POST to https://api.kie.ai/api/v1/jobs/createTask - Model: "kling-2.6/image-to-video" ⚠️ CRITICAL: The "image_urls" field MUST contain the image URL from Airtable! Example payload: { "model": "kling-2.6/image-to-video", "input": { "prompt": "The fox slowly turns its head and smiles at camera", "image_urls": ["<URL from Airtable start_image field>"], "duration": "5", "sound": false } } - Poll for completion (can take 2-4 minutes) - Get the result video URL 3. Log the generated video to Airtable: [set scene_video field] 4. Download the video locally to the outputs/ folder Load API keys from .agent/.env: KIE_API_KEY, AIRTABLE_TOKEN, AIRTABLE_BASE_ID Print friendly progress messages at each step. - 6Combine to final video · edit · FFmpeg (local, no API) · localConcatenates the scene clips via FFmpeg and optionally overlays background music with a 2-second fade-out.
Prompt
Create a Python script called combine_all.py in the tools/ folder. This script should combine multiple video clips into one final video using FFmpeg. The script needs to: 1. Take a list of video files (the clips generated in Step 4) 2. Concatenate them using FFmpeg: - Create a text file listing all the video paths - Use FFmpeg concat demuxer to join them - Example command: ffmpeg -f concat -safe 0 -i videos.txt -c copy output.mp4 3. Optionally add background music: - If a music file is provided, overlay it on the video - Add a 2-second fade-out at the end so the music doesn't cut abruptly - Use -shortest flag so audio stops when video ends 4. Save the final video to the outputs/ folder 5. Clean up temporary files (the videos.txt list file, any intermediate files) Requirements: FFmpeg must be installed on the system; print friendly progress messages; handle errors gracefully. Note: This is a local tool using FFmpeg - no API keys or costs involved.
- 7Package as skill · Words · agentWrites a reusable SKILL.md documenting the whole pipeline.
RoboNuggets · R54 | Antigravity Creative Cloner AI Agent- 0
Cinematic ads agent (storyboard-to-video)Partlya single cinematic product/brand video ad (around 15 seconds in the demo), built from a still product photo via an approved 6-panel storyboard
GPT Image 2Seedance 2.0 Pro (source spells it "C-Dance 2.0"/"Sea Dance 2.0 Pro" — exact spelling uncertain)cinematic adb-roll / generic product videography- 1Set product · research · agent's own vision (Claude Code)Agent analyzes an uploaded product image/link/brief, extracts brand name, product name, category.
- 2Storyboard direction · Script · Claude (agent's own reasoning)Agent proposes 3 distinct ad directions (concept name, vibe, hero moment) for the user to pick.
- 3Generate storyboard image · Picture · GPT Image 2 (edit endpoint) · fal.aiRenders a 6-panel storyboard image with timestamps, camera notes and in-image text describing each beat.
- 4Generate final ad video · Video · Seedance 2.0 Pro (name uncertain) · fal.ai (reference-to-video endpoint)Turns the approved storyboard image plus a clean product photo into the full cinematic video ad.
Prompt
Turn this storyboard, which it references as image one, into a cinematic 15-second sports ad.
Missing: the skill.md system prompt and its exact prompt templates are inside cinematic-ads-agent-kit.zip, never opened/extracted; exact provider name is ambiguous ('file.ai' vs fal.ai) per transcript garbling; exact video model name uncertain.
RoboNuggets · R61 | Cinematic Adverts Agent- 1
Product video9
Start/end-frame scene transitionsPartlya smooth video clip transitioning between two still scenes (start frame and end frame), used to chain a multi-scene narrative ad; style is set by the frame images (photoreal product or Pixar-style 3D)
SeedanceKlingSeedream 4.5photoreal productPixar-style 3D animated- 1Generate the next-scene image prompt · Script · custom GPT ('ZaxFilm Script to Prompt') · ChatGPTFeed a script line (plus brief product context) into the 'Script to Prompt Generator' GPT ('ZaxFilm Script to Prompt'). It returns numbered image-concept options for the start frame, then the exact copy-paste-ready image prompt plus a negative prompt once an option is picked.
Prompt
If you cut open this fruit and drink what's inside, your skin will never look the same again.
- 2Generate the start-frame still · Picture · Seedream 4.5 · Max FusionTake the generated prompt to Max Fusion Images > Create Image, using Seedream (named 'Seadream'/'Seadream 4.5') for speed/lower credit cost over Nano Banana. If there's an issue, describe the fix back to the same GPT thread rather than restarting.
- 3Generate the end-frame still · Picture · Seedream 4.5 · Max FusionClick 'generate variation' on the accepted start-frame image (to keep scene/character consistency), then paste the GPT's end-frame prompt on top of that variation. Can also ask the GPT ad hoc for 'the next shot' continuing the story.
Prompt
make the woman look like she's drinking the juice, with the juice touching her lips and the level empty.
- 4Generate the transition video · Video · Seedance / Kling · Max Fusion / OpenArtUpload both start and end stills to the video model and prompt the transition motion connecting them.
Prompt
smooth science-y transition where the juice from the fruit goes down to fill her glass.
Missing: the ZaxFilm GPT's full copy-paste image prompt text was shown on-screen but not read aloud/quoted in the transcript beyond the concept summary
AI Ads (Ad Creators Lab) · Generating Scenes/Transitions / Script-To-Prompt Generator / Creating engaging animated videos with AI tools (Lesson 27)- 1
Local image-to-video with Wan 2.2 14B (ComfyUI)PartlyA short animated video clip generated locally from one start image plus a motion/camera prompt, using a self-hosted Wan 2.2 14B model in ComfyUI, with an optional local vision-LLM step that writes the motion prompt automatically from the image.
Wan 2.2 14B I2V (Lightning 4-step LoRA)Qwen2.5-VL 7B (via Ollama, for auto-prompting)image-to-video clip- 1Load models · By hand · wan2.2-i2v-14b-low-noise-q4_k_m.gguf + wan2.2-i2v-14b-high-noise-q4_k_m.gguf + Wan2.2-Lightning_I2V-A14B-4steps-lora (HIGH + LOW, weight 1.0) · ComfyUI (local)Loads the Wan 2.2 14B I2V high-noise and low-noise GGUF diffusion models plus their 4-step Lightning LoRAs (one per noise stage), the UMT5-XXL text encoder, and the Wan VAE.
- 2(Optional) auto-write the motion prompt from the start image · Words · qwen2.5vl:7b-q8_0 (via Ollama) · Ollama (local, http://192.168.0.210:11434 in the source)A local vision-LLM node (Ollama, Qwen2.5-VL 7B) looks at the uploaded start image and writes a 2-3 sentence cinematic animation prompt following a fixed formula: lead with the subject's natural action, follow with camera movement, matching the scene's mood - an alternative Ollama node (llama-style, seeded) implements a more detailed Motion Description + Camera Movement + optional Aesthetic/Dynamic tags formula, 60-140 words.
Prompt
You are an expert at creating concise, cinematic animation prompts for AI image-to-video models. When given any image, analyze it quickly and produce a direct animation description following these principles: Core Guidelines: Subject Movement First: Consider what natural action the subject would logically perform in their environment (walking down streets, turning to look, subtle gestures, continuing implied motion) Static Only When Appropriate: Keep subjects still only if the scene clearly suggests contemplation, waiting, or dramatic pause Camera as Support: Use camera movement to enhance subject action, not replace it Context Awareness: Match movement intensity to the scene's mood and setting Output Requirements: Write 2-3 sentences maximum Lead with subject action, follow with camera movement Use cinematic language that creates visual flow Maintain the image's original mood and atmosphere Be specific about movement direction and pace Examples of Natural Actions by Context: Streets/alleys: walking, strolling, turning Portraits: subtle head turns, breathing, eye movement Action scenes: continuing the motion Contemplative scenes: minimal but purposeful movement Produce only the final animation prompt - no analysis breakdown, no section headers, just a clean, cinematic description ready for i2v generation.
- 3Resize start image and set positive/negative prompts · Words · ComfyUI (local)Resizes the uploaded start image to the target resolution and sets the CLIP text-encode positive prompt (either the auto-written motion prompt or a manual one) plus a fixed Chinese-language negative prompt covering common generation defects.
Prompt
POSITIVE (worked example in the workflow): The blue sports car accelerates through the tunnel, its taillights glowing brightly as it speeds towards the exit. The camera follows the car, panning slightly to the right to capture the dynamic motion and the tunnel's curved walls illuminated by orange and white lights. NEGATIVE: 色调艳丽,过曝,静态,细节模糊不清,字幕,风格,作品,画作,画面,静止,整体发灰,最差质量,低质量,JPEG压缩残留,丑陋的,残缺的,多余的手指,画得不好的手部,画得不好的脸部,畸形的,毁容的,形态畸形的肢体,手指融合,静止不动的画面,杂乱的背景,三条腿,背景人很多,倒着走
- 4Sample and generate the video · Video · Wan 2.2 14B I2V (Lightning 4-step) · ComfyUI (local)Runs the Wan 2.2 14B I2V Lightning sampler (4 steps, split step 2, shift 5) over the high/low-noise model pair to produce the video frames from the start image and conditioning.
- 5Save the video · edit · ComfyUI (local)Encodes the generated frames to an MP4 (h264) file.
Missing: Runs entirely on self-hosted ComfyUI + a local GPU (repo warns Wan 2.2 14B I2V is very VRAM-intensive) plus an optional local Ollama vision model - not reachable via any hosted API listed in the brief (Kie.ai/ApiMart/MuAPI/fal.ai/Google/OpenAI/ElevenLabs), so it is not directly runnable through our current provider stack without a self-hosted GPU box.
Free online · GitHub — Jeff-Emmett/ComfyUI_Workflows: video_wan2_2_14B_i2v_lx2v.json- 1
Local text-to-video with Wan 2.2 14B (ComfyUI)PartlyA short cinematic video clip generated locally purely from a text description (no start image), using a self-hosted Wan 2.2 14B text-to-video model in ComfyUI, with an optional local LLM step that expands a short user idea into a full structured T2V prompt.
Wan 2.2 14B T2V (Lightning 4-step LoRA)gemma3:latest (via Ollama, for auto-prompting)text-to-video clip- 1Load models · By hand · wan2.2-t2v-14b-low-noise-q4_k_m.gguf + wan2.2-t2v-14b-high-noise-q4_k_m.gguf + Wan2.2-Lightning_T2V-v1.1-A14B-4steps-lora (HIGH + LOW) · ComfyUI (local)Loads the Wan 2.2 14B T2V high-noise and low-noise GGUF diffusion models plus their T2V 4-step Lightning LoRAs (HIGH + LOW, weight 1.0), the UMT5-XXL text encoder, and the Wan VAE.
- 2(Optional) expand a short idea into a full T2V prompt · Words · gemma3:latest (via Ollama) · Ollama (local)A local LLM node (Ollama, gemma3:latest) transforms a short user idea into a 120-200 word production-ready prompt following an explicit formula: Subject(Description) + Scene(Description) + Motion(Description) + Aesthetic Control + Stylization, with defaults for very brief input (naturalistic lighting, subtle camera glide, realistic motion, medium lens, neutral grade).
Prompt
Transform a short user idea into a polished, production-ready positive prompt for Wan-style text-to-video that adheres to the Advanced Formula: Subject(Description) + Scene(Description) + Motion(Description) + Aesthetic Control + Stylization. Instructions: Extract and structure from the user text: subject(s), action(s), setting, time-of-day, mood/tone, key props, and intended style; expand minimally to ensure cinematic specificity without adding new story beats not implied in the input. Compose one paragraph, 120-200 words, in this implicit flow: Subject and action focus, with a compact subject description. Scene description and atmosphere (lighting, time, weather, era). Motion description: specify amplitude/speed/tempo for subject and scene motion. Camera plan: choose 1-2 beats (e.g., dolly in, pan right, tracking, arc), coherent with the action. Aesthetic Control: weave in concise tags from Light Source, Lighting Type, Time of Day, Shot Size, Composition, Lens, Color Tone, and optionally Character Emotion. Stylization: include a single style touch (e.g., documentary realism, cyberpunk, watercolor) if implied or beneficial; keep subtle and unified. Keep language concrete, filmic, model-agnostic; do not output negatives, sampler names, seeds, steps, aspect ratios, or model names. Prefer strong nouns/verbs over adjective stacks; avoid enumerations or shot lists. Defaults when the input is very brief: Use safe cinematic defaults: naturalistic lighting, subtle camera glide, realistic motion, medium lens, neutral color grade; keep the style grounded unless the input suggests otherwise. Example style (not content): At dawn on a misted lakeshore, a lone rower in a weathered skiff pulls through glassy water, breath steady and focused. Soft, diffused sunrise light rims the oars as ripples expand in concentric rings. Start on a medium-wide as the camera tracks parallel, then ease into a slow dolly to a close three-quarter profile. Motion is unhurried and deliberate; droplets arc from each stroke. Soft lighting, low-contrast palette with cool blues and gentle warms; center composition shifting to short-side; medium lens; restrained, documentary realism.
- 3Encode positive/negative prompts · Words · ComfyUI (local)Sends the expanded (or manual) prompt plus the same fixed Chinese negative prompt used in the I2V workflow to CLIP text-encode nodes.
Prompt
POSITIVE (worked example in the workflow): The hero starship approaches a vast ringed gas giant through sparse debris, anodized alloy and painted composites showing fine micrometeorite pitting and streaking as guidance lights tick along the spine. The planet's bands give a motivated cool rim while deep space stays neutral; ice fragments provide clean parallax. Motion is constant-velocity and weighty: tiny attitude corrections only; brief thruster puffs hold vector with zero overshoot. Use a wide establishing then a slow, tripod-steady forward dolly-no easing. Engine exhaust reads as high-temperature plume with 8-12Hz flicker, subtle heat-shimmer and a soft ionization cone refracting the background; nozzle interiors show heat-tinted Inconel. Lighting is physically plausible with inverse-square falloff: planetary rim as key, soft bounce off the hull as fill. Neutral cinematic grade, cool low-contrast tones, soft highlight roll-off; 24-28mm feel for scale; center-weighted composition. Premium live-action photoreal; 180 degree shutter blur; subtle film grain; no stylization NEGATIVE: 色调艳丽,过曝,静态,细节模糊不清,字幕,风格,作品,画作,画面,静止,整体发灰,最差质量,低质量,JPEG压缩残留,丑陋的,残缺的,多余的手指,画得不好的手部,画得不好的脸部,畸形的,毁容的,形态畸形的肢体,手指融合,静止不动的画面,杂乱的背景,三条腿,背景人很多,倒着走
- 4Sample and generate the video · Video · Wan 2.2 14B T2V (Lightning 4-step) · ComfyUI (local)Runs the Wan 2.2 14B T2V Lightning sampler (4 steps, split step 2, shift 5) over the high/low-noise model pair to generate video frames purely from the text conditioning.
- 5Save the video · edit · ComfyUI (local)Encodes the generated frames to an MP4 (h264) file.
Missing: Runs entirely on self-hosted ComfyUI + a local GPU plus an optional local Ollama LLM - not reachable via any hosted API listed in the brief, so not directly runnable through our current provider stack without a self-hosted GPU box.
Free online · GitHub — Jeff-Emmett/ComfyUI_Workflows: video_wan2_2_14B_t2v_lx2v.json- 1
Product photography stills to short product video (ComfyUI)Runsa set of polished still product-photography images (with targeted inpaint edits) plus a short interpolated product video made from two chosen stills, output as a ComfyUI JSON workflow
GPT Image 1 (OpenAIGPTImage1 node)ByteDance first/last-frame video model (ByteDance2FirstLastFrameNode)product photographyproduct video- 1Generate product photography · Picture · GPT Image 1 (gpt-image-2) · OpenAI (via ComfyUI OpenAIGPTImage1 node)OpenAIGPTImage1 generates polished product photography from a reference product image and an optional background image, guided by a concise prompt. LoadImage nodes import the product and backdrop assets; ImageBatchMulti processes variations; SaveImage captures output.
Prompt
Place the product in the first image to the scene in second image, and change the greenery to mint leaves.
- 2Mask-edit for precision fixes · edit · GPT Image 1 (gpt-image-2) · OpenAI (via ComfyUI OpenAIGPTImage1 node)Open the mask editor and paint where changes should apply (background cleanup, reflections, label tweaks), then re-run the same GPT Image 1 node for targeted inpainting on just the masked area. ImageCompare node gives a side-by-side before/after check.
Prompt
Remove the moss at the red circles. And add a strand of mint flowers around blue circle
- 3Interpolate stills into a video · Video · Seedance 2.0 (ByteDance first/last-frame video model) · ByteDance (via ComfyUI ByteDance2FirstLastFrameNode)Select two chosen stills as first and last frame; ByteDance2FirstLastFrameNode (Seedance 2.0) synthesizes the in-between motion; SaveVideo encodes the sequence to a shareable clip. Source notes consistent resolution/aspect ratio across nodes improves temporal consistency.
Prompt
Camera slowly orbits left to right around the subject, one continuous take, no cuts.
Free online · Comfy for Advertising - Product Photography- 1
Brand video library post/clip (image-to-video from library asset)PartlyA video generated from an existing brand video-library still, animated per an operator's 'what happens next' instruction.
- 1libraryPost / libraryClip · VideoOperator picks a library asset and writes a motion instruction (capped 600 chars); libraryPost()/libraryClip() dispatch the generation.
Missing: Provider not stated in repo.
Ours ·- 1
Kling 3.0 Omni Multi-Shot Product AdPartlyA 12-second, 4-shot product video (hero shot, material/shape orbit, detail close-up, iconic low angle).
Kling 3.0 Omnipseudo-documentary- 1Build an Element · Picture · Nano Banana Pro (or clean photos)Build a Kling 'Element' from 3-4 product reference images.
- 2Configure + generate · Video · Kling 3.0 Omni · Kie.aiKling 3.0 Omni, Multi-Shot on, 4 shots x 3s, select the Element on every shot (flagged #1 mistake if skipped).
Prompt
Hero Shot / Material & Shape orbit / Detail close-up / Iconic low angle (literal 4-shot prompt in knowledge/kling-3-guide.md, not reproduced verbatim here — file exceeds a short quote)
Missing: No mechanics given for creating an 'Element'; no cost/account info.
Ours ·- 1
3D-style product animation (Nano Banana Pro + Veo3.1, Airtable-driven)Runsstudio-quality product video with clean start/end shots and a driven camera/product transition
Nano Banana ProVeo3.1 fastlandscape9x16 portrait- 1Collect inputs · input · AirtableAirtable base row holds product photo(s) (ideally multiple angles), a short creative direction text, and aspect ratio; a button field triggers an n8n webhook carrying the record ID
- 2Generate prompts · agent · OpenRouterAI agent (system prompt covers composition, lighting, environment, action, camera, aesthetic, mood) produces a detailed starting-image prompt, a simpler ending-image prompt (since the ending image uses the starting image as its input), and a transition prompt for the video, routed via a switch node that checks whether prompts/images already exist for that row
- 3Generate starting image · generate-image · Nano Banana Pro · Kie.ai or WavespeedNano Banana Pro generates the starting product image from the product photo + starting-image prompt
- 4Generate ending image · generate-image · Nano Banana Pro · Kie.ai or WavespeedNano Banana Pro generates the ending product image, using the starting image as input plus the simpler ending-image prompt, to keep both frames part of the same scene
- 5Generate transition video · generate-video · Veo3.1 fast · Kie.aiVeo3.1 fast generates the video from the starting and ending frames plus the transition prompt (camera movement + product action, e.g. product remains static, or lifts and separates into an exploded view)
Prompt
e.g. 'the keyboard lifts into the air and separates into a detailed exploded view'
- 6Log to Airtable · log · AirtableGenerated images and video are written back to the Airtable row for review
Missing: full system prompt document is linked externally, not reproduced in the harvested transcript
RoboNuggets · n30 - 3D Product Animations with AI- 1
Food-photo-to-video (Kling image-to-video, start/end frame)PartlyA short video clip animating between two AI-enhanced food photos of the same dish (one as the starting frame, one as the ending frame).
Kling v2.5 Turbo Proshort dish animation clip for social/menu use- 1Set start and end frames · Video · Kling v2.5 Turbo Pro (kwaivgi/kling-v2.5-turbo-pro/image-to-video) · Wavespeed.aiThe two Nano Banana Pro images generated for a dish (hero shot and camera-angle variant) are set as the starting frame and ending frame for a Kling v2.5 Turbo Pro image-to-video generation, run as a third, optional n8n workflow ('purple section') on top of the image pipeline.
Missing: Presenter explicitly calls this 'outside the scope of this specific tutorial' and does not walk through node-level setup, duration, resolution, or prompt fields for the Kling call.
RoboNuggets · n32 - Restaurant Photography with Nanobanana Pro- 1
Product photo to video adRunsone vertical (or custom aspect ratio) product ad video, image-to-video, built from a single product photo plus a creative brief, delivered via Telegram
GPT-image-1 (gpt4o-image)Veo3Veo3-fastproduct ad- 1Receive brief · By hand · TelegramTelegram photo + caption trigger starts the run.
- 2Describe product · Picture · OpenAI (exact model not named) · OpenAIVision call describes the product only, ignoring background.
Prompt
Describe the product and brand in this image in full detail. Fully ignore the background. Focus ONLY on the product.
- 3Generate image + video prompts · Words · not named (n8n AI Agent node) · n8nAI Agent (A-G-E-N-T framework) crafts a structured JSON with image_prompt, video_prompt, caption, creative_summary, aspect_ratio, video_model.
Prompt
This is the initial creative brief: {{ $('Telegram Trigger').item.json.message.caption }} Description of the product: {{ $json.choices[0].message.content }} Use the Think tool to double check your output - 4Human approval · Check · same agent · TelegramSends the proposed prompts to Telegram for yes/no plus free-text comment; on 'no' a Revised Prompt agent re-runs step 3 with the comment added.
- 5Generate image · Picture · GPT-image-1 (gpt4o-image) · Kie.ai (POST /api/v1/gpt4o-image/generate)Places the product into the described scenario.
Prompt
Take the product in the image and place it in this scenario: {{ image_prompt }} - 6Poll image · Check · Kie.ai (GET /api/v1/gpt4o-image/record-info)Polls for the finished image.
- 7Generate video · Video · Veo3 or Veo3-fast (agent-chosen) · Kie.ai (POST /api/v1/veo/generate)Animates the generated image into the finished ad video.
Prompt
{{video_prompt}} - 8Poll and deliver · publish · Kie.ai (GET /api/v1/veo/record-info) + TelegramPolls for the finished video and sends it back on Telegram.
RoboNuggets · R36 | The Ad Creator AI Agent- 1
Faceless video11
Cinematic nature quote video (Flux stills + Kling motion + ElevenLabs ambient soundscape + burned-in quote text)Runsa vertical cinematic nature video with a calm ambient soundscape and an overlaid quote + author caption, auto-uploaded to YouTube
Qubico/flux1-dev (PiAPI)Kling 1.0 (PiAPI)ElevenLabs sound-effects-v1vertical cinematic quote video- 1Get data from Google Sheet · Words · Google SheetsReads a row containing a Background (EN)/Prompt (EN) description, a Thai quote and pen name (author).
- 2Generate Image · Picture · Qubico/flux1-dev · PiAPIGenerates a vertical (540x960) ultra-realistic nature landscape still via PiAPI's Flux endpoint, combining the sheet's Background and Prompt fields into a fixed template prompt.
Prompt
REQUEST BODY (verbatim, node 'Generate Image'): prompt = "Ultra-realistic vertical nature landscape, {{ Background (EN) }}, featuring {{ Prompt (EN) }}, high detail, soft atmospheric lighting, cinematic golden hour glow, vertical composition, photorealistic texture, natural depth of field, calm and serene mood, no people, no buildings, HDR, 8k resolution, masterpiece"; negative_prompt = "taking a photo of a room, recording a video of a room, photos app, video recorder, illegible text, blurry text, low quality text, DSLR, unnatural" - 3Image-to-Video · Video · Kling 1.0 (PiAPI, mode std) · PiAPIAnimates the still into a subtle 5-second cinematic motion clip (soft light, mist, swaying trees, slow zoom) via PiAPI's Kling endpoint.
Prompt
REQUEST BODY (verbatim, node 'Image-to-Video'): prompt = "Cinematic vertical video from image of {{ Background (EN) }}, with {{ Prompt (EN) }}, animated subtly from image, with soft light, mist, swaying trees, slow zoom effect, no people or buildings"; negative_prompt = "blurry motion, distorted faces, unnatural lighting, over produced, bad quality" - 4Generate Audio · music · sound-effects-v1 · ElevenLabsGenerates a 20-second ambient nature soundscape (dawn, ambient nature sounds tied to the Background/Prompt text, soft lofi beats) via ElevenLabs sound generation, explicitly marked 'no voice'.
Prompt
REQUEST BODY text field (verbatim, node 'Generate Audio'): "no voice, A peaceful soundscape unfolds as the sun begins to rise over misty {{ Background (EN) }}, casting warm light across the scene. The crisp morning air is filled with ambient nature sounds like {{ Prompt (EN) }} along with soft lofi beats, blending into a calm and immersive atmosphere." - 5Prepare Overlay Text (Quote & Author) · edit · n8n (Code node)A Code node builds FFmpeg drawtext commands to burn the Thai quote (word-wrapped to fit video width) and the author's pen name onto the video, using the Kanit-Italic font.
- 6Upload to YouTube · publish · YouTube APIPerforms a resumable upload of the final captioned video to YouTube.
Missing: none for image/video/audio generation prompts, all confirmed verbatim; no LLM/quote-writing step exists in this template — quotes come pre-written from the Google Sheet, not generated
Free online · Create cinematic nature videos with AI-generated quotes and auto-upload to YouTube- 1
Faceless YouTube Short from a topic (ElevenLabs voice + Hailuo AI visuals)ManualA faceless YouTube Short: a scripted voiceover read by an ElevenLabs voice over a sequence of AI-generated images animated into video clips by Hailuo AI, combined with captions and logged to a sheet.
ElevenLabs (TTS)Hailuo AI (image-to-video)OpenAI Whisper (captions)faceless YouTube Shortnarrated topic video- 1Set video category / topic · By hand · n8nManual trigger sets the video category/topic that drives script and image-prompt generation.
- 2Generate voiceover script and transcript · Voice · ElevenLabs TTS · ElevenLabsGenerates a script/transcript for the short, then converts it to speech.
- 3Generate per-scene image prompts · WordsBuilds a list of image prompts (one per scene/line of narration) from the script.
- 4Generate scene images · Picture · Google Cloud StorageGenerates a still image per scene prompt, saves each to Google Cloud Storage.
- 5Animate first image with Hailuo · Video · Hailuo AI (MiniMax image-to-video) · Hailuo AISends the (first/each) image to Hailuo AI to generate a short video clip, polls generation status, then downloads the finished clip once ready.
- 6Combine music/video and assemble final short · editMaps the video clip(s) with background music, submits an assembly job, polls its progress, and once complete pulls the final video URL and its captions.
- 7Generate captions and log · Words · OpenAI Whisper · OpenAITranscribes/aligns captions (Whisper) and logs the finished short's details to Google Sheets.
Missing: This is a paid ($10) n8n template; the template metadata API (api.n8n.io) returns real node names, connections and node count (38 nodes) but every node's own 'parameters' object is blanked/redacted for non-owners, so no prompts, API URLs or settings could be captured verbatim for most steps. The step list above is inferred from node names and connection order only.
Free online · n8n — Automatically create YouTube short videos using ElevenLabs, Hailuo AI- 1
Veo3 faceless character video clipPartlyone realistic 8-second Veo3 video clip of an AI-invented character/scene (e.g. a vlogging creature), from a niche topic, logged for later posting.
Veo3GPT-4.18-second clip, any aspect ratio set in the request- 1Generate idea · research · GPT-4.1 (agent chat model) · OpenAIAn 'ideas' AI agent is given a niche (e.g. a Yeti vlogger) and produces a fresh video idea, a caption, and an environment/setting description.
- 2Log idea · Words · Google SheetsLogs the idea, caption, status and environment to a Google Sheet for tracking.
- 3Craft Veo3 prompt · Script · GPT-4.1 (agent chat model) · OpenAIA 'prompts' AI agent takes the idea and environment and writes a detailed, well-formed Veo3 video prompt.
- 4Generate video · Video · Veo3 · Kie.ai (see providers note)HTTP POST sends the crafted prompt to Veo3 via the aggregator; request is queued.
- 5Wait and poll · Check · Kie.ai (see providers note)Waits ~10 minutes then polls the task status by ID until the video URL is returned.
- 6Log final video · Words · Google SheetsMarks the sheet row's production status as done and stores the final video URL.
Missing: ideas/prompts agent system prompts not harvested (course community template only); provider identity for the Veo3 call is ambiguous in the transcript.
RoboNuggets · n6 - Veo3 in n8n- 1
Quote YouTube Shorts / IG Reels with cloned voiceRunsA vertical YouTube Short / Instagram Reel of a spoken quote over a looping background video, with burned-in subtitles and a voiceover in a cloned or premade voice.
ElevenLabs (TTS / instant voice cloning)quote videomotivational short- 1Pull viral quotes from a source video · research · ChatGPT (unspecified version) · OpenAItactiq.io generates a transcript of a chosen YouTube video; ChatGPT extracts quotes from it.
Prompt
Get the most viral quotes from this transcript and show in a table format. Column 1 is index starting from 1 Column 2 is a short and punchy Title Column 3 is the quote. Make it 2-3 sentences per quote Give 20 quotes
- 2Store quotes queue · Words · Google SheetsQuotes copied into a Google Sheet template with a status column (not posted / posted) driving the daily automation.
- 3Generate voiceover · Voice · ElevenLabs TTS / Instant Voice Cloning · ElevenLabsElevenLabs converts the quote text to speech using a premade voice or an instant-cloned voice (e.g. cloned from an uploaded sample of a public figure's voice).
- 4Store audio, get link · edit · DropboxAudio uploaded to Dropbox, a shareable link generated for the next step.
- 5Assemble video · edit · JSON2VideoJSON2Video renders a vertical movie from a template: a pre-made looping background video/image, auto-generated subtitles from the audio, and the ElevenLabs audio track.
- 6Publish · publish · Make.comMake.com posts the finished video to YouTube (title, description from sheet) and Instagram Reels (caption from sheet); sheet status row is marked posted; runs on a daily schedule.
Missing: Exact JSON2Video movie-template JSON and background video asset not captured verbatim in the source lesson.
RoboNuggets · R2 | Create & Schedule 100 YouTube Shorts- 1
Zero-human faceless video generator (Kling direct)RunsA faceless short-form video (TikTok/YouTube Shorts/Instagram Reels) generated directly from a text prompt via Kling, posted automatically once a day.
Kling (text-to-video, model 'kling-v1')faceless video- 1Generate video ideas · Words · ChatGPT · OpenAIChatGPT generates a table of video ideas/prompts for a chosen niche.
Prompt
My video niche is [Batman in different superhero costumes] To describe it in more detail: [...] -- Give me 5 video prompts in this table format these are the columns: id, status, title, caption, prompt id goes from 1 onward status is always "todo" title is a short phrase to describe the video caption is optimised for tiktok, instagram, and youtube prompt is a detailed description which you can provide to an ai video generation tool Never use double quotes for any of the values
- 2Generate video · Video · Kling v1 · Kling AIMake.com POSTs the prompt directly to Kling's text-to-video API (model kling-v1, mode Pro, 9:16, duration 10s, CFG scale 0.6) and polls/receives via webhook callback.
- 3Publish · publish · Ayrshare (TikTok), Make.com (YouTube/Instagram)Make.com posts the video to YouTube and Instagram directly, and to TikTok via Ayrshare's API (JSON: post/mediaUrls/platforms).
Prompt
{"post": "(insert caption here)", "mediaUrls": ["(map webhook url here)"], "platforms": ["tiktok"]}
RoboNuggets · R16 | The Zero Human Faceless AI Toolkit- 1
Faceless video generator with multi-clip image-to-video stitchingRunsA faceless short-form video assembled from three separately generated image-to-video clips (each a different AI 'character'), stitched together with background music, posted daily to TikTok/YouTube/Instagram.
Flux (text-to-image, via PiAPI)Kling (image-to-video)Suno (music, manual)faceless video, multi-scene- 1Generate three prompts + actions per idea · Words · ChatGPT · OpenAIChatGPT generates a table of ideas, each with three distinct image prompts (prompt_a/b/c, different characters) and three short action lines (action_a/b/c).
Prompt
You provide detailed video prompts for AI generation tools like Midjourney and Flux ... Quantity of rows I need: 10 Minimum prompt length: 100 words Maximum action prompt length: 15 words My video niche: female Disney characters in 80s panavision style ... Example of a good video prompt: An ethereal and enchanting version of a real-life version of a 24-year old Tinkerbell... Example of a good action prompt: smiles, looks at the camera, then looks away
- 2Generate image (x3) · Picture · Flux (dev) · PiAPIMake.com POSTs each prompt (a, b, c) to PiAPI's Flux text-to-image service.
- 3Animate image to video (x3) · Video · Kling (image-to-video) · PiAPIMake.com POSTs each generated image + its action line to Kling's image-to-video endpoint.
- 4Source music (manual) · By hand · Suno · Suno.comA music track is generated manually at suno.com and dropped into a Dropbox music folder.
- 5Stitch clips + audio · edit · JSON2VideoJSON2Video assembles the three clips (with fade in/out, 5s each, sequential timing) and the Suno audio track into one movie, rendered vertically (1080x1920).
- 6Publish · publish · Ayrshare (TikTok), Make.com (YouTube/Instagram)Make.com posts the finished video to YouTube and Instagram directly, and to TikTok via Ayrshare.
Missing: JSON2Video movie template JSON not fully reproduced verbatim in the source; Suno music prompt not captured (manual, unscripted in the walkthrough).
RoboNuggets · R17 | The ZeroTouch Faceless Videos AI Agent v2.0- 1
Faceless POV AI video machine (n8n)Partlya ~65-second faceless vertical video built from 13 sequential 5-second AI-generated video clips (each animated from a still image) with a matching ambient sound bed, auto-logged and optionally auto-published
GPT-4/4.1 (chat, for ideas/titles/prompts)Flux (Qubico/flux1-dev, txt2img)Kling (video_generation, image-to-video)Hailuo (i2v-01, optional alternative)ElevenLabs sound-generationfaceless storyPOV documentary-style clip series- 1Generate the video idea · Words · GPT-4/4.1 · OpenAIA chat model (Ideas Agent) generates one viral idea as structured JSON (title, idea, environment, status) on a schedule trigger.
- 2Log idea to Google Sheets · Words · Google SheetsAppend the new idea row (title/caption, environment, production status = 'for production', row-number ID) to a tracking Google Sheet.
- 3Generate Titles (system prompt) · Words · GPT-4/4.1 · OpenAIA chat model turns the idea+environment into a finished viral title with hashtags.
- 4Generate 13 scene image-prompts · Words · GPT-4/4.1 · OpenAIA chat model (Generate Prompts) writes 13 cinematic still-image scene prompts matching the idea and environment, one per 5-second clip.
- 5Unbundle scenes · edit · n8n (Code node)A JS code node splits the single 13-scene JSON object into 13 separate items, one per scene, for parallel processing.
- 6Generate image per scene · Picture · Flux (Qubico/flux1-dev) · PiAPIEach scene description is sent to Flux (txt2img) via PiAPI to generate a still image.
Prompt
{{ $json.description }} - 7Generate video per scene · Video · Kling (video_generation) · PiAPIEach scene's generated image is animated into a 5-second video clip via Kling image-to-video; an optional cheaper Hailuo (i2v-01) alternative is mentioned.
- 8Generate ambient sound per scene · music · ElevenLabs sound-generation · ElevenLabsEach clip's environment/action is turned into an ambient sound-effect track via ElevenLabs sound-generation.
- 9Store and share clip assets · edit · Google DriveGenerated images/videos/sounds are uploaded to Google Drive and shared for downstream assembly.
- 10List elements · edit · n8n (Code node)A JS code node collects the 13 clips' asset URLs (video/audio/text) back into one ordered list ready for the render template.
- 11Build Creatomate render template · edit · CreatomateA Creatomate template is created/populated with Audio-1..5, Video-1..5 and Text-1..5 elements sequencing the 13 clips into one timed composition.
- 12Render final video · edit · CreatomateCreatomate renders the populated template into the final stitched vertical video.
- 13Optional: publish to YouTube · publish · YouTubeThe finished video is optionally uploaded/published to YouTube as the last automated step.
Missing: several system prompts (Generate Titles, Generate Prompts) and the exact Kling / ElevenLabs sound-generation request bodies and the full Creatomate template JSON were viewed in the source lesson during an earlier read but did not survive verbatim into this rewrite pass; re-open knowledge/courses/robonuggets-school/all-lessons-and-templates/r24-the-faceless-pov-ai-machine-n8n-no-code--b11c91cd.md directly to recover the exact prompt/body text before building.
RoboNuggets · R24 | The Faceless POV AI Machine (n8n, no-code)- 1
Monetizable faceless TikToks (fal.ai + ffmpeg)Runsa ~65-second (13 x 5-second scene) faceless vertical video in a chosen niche, built from still images animated into clips with matching ambient sound and stitched together, aimed at TikTok's Creator Program minimum length
GPT-4.1 (chat, idea/prompt/sound-prompt generation)Flux Schnell or Flux Pro (fal-ai/flux, txt2img)Kling Standard or Pro (fal-ai/kling-video, image-to-video)MMAudio v2 (fal-ai/mmaudio-v2, video-conditioned sound generation)ffmpeg-api/compose (fal-ai, video sequencing)faceless nature/documentary-style clip series- 1Create new idea · Words · GPT-4.1 · OpenAIOn a schedule trigger, GPT-4.1 generates one immersive viral sea-animal cleaning/rescue idea as a single-line JSON array with Title (incl. 12 hashtags: 4 niche + 4 all-time-popular + 4 trending), Idea (<13 words), Environment (<20 words), and Status fixed to 'for production'.
Prompt
You are an AI designed to generate 1 immersive, viral sea animal idea. Your output must be formatted as a JSON array (single line) and follow all the rules below exactly. RULES: 1. Only return 1 idea at a time. 2. The Idea must: - Be under 13 words. - Describe a realistic cleaning or rescue action involving sea animals. - Focus on real-life actions (divers cleaning whales, rescuers removing barnacles, machines cleaning whale skin, etc.). - No fantasy elements (no animal-to-animal rescues, no magical events). 3. The Title must be: - Short, punchy, and viral-friendly. - Include one relevant emoji. - Include exactly 12 hashtags in this order: ** 4 hashtags relevant to the sea animal idea ** 4 all-time most popular hashtags ** 4 most trending hashtags today (based on live research) - All hashtags must be lowercase. 4. Set Status to "for production" (always). 5. The Environment must: - Be under 20 words. - Match the action in the Idea exactly (same participants and setting). - Clearly describe: - Where the cleaning or rescue is happening (underwater, on boat deck, open ocean surface, etc.) - Main background details (sunlight beams, overcast sky, misty water, foamy waves, etc.) - Key participants (diver, rescuer team, cleaning machine, etc.) - Style of scene (cinematic realism, gritty documentary, macro close-up, aerial view, etc.) - No fantasy settings. OUTPUT FORMAT (single-line JSON array): [ { "Title": "Summarized viral sea animal title with emoji #4_relevant_hashtags #4_all_time_popular_hashtags #4_trending_hashtags", "Idea": "Short immersive sea animal cleaning or rescue idea under 13 words", "Environment": "Brief vivid setting under 20 words matching the action", "Status": "for production" } ] - 2Log the idea · Words · Google SheetsAppend the idea (Idea, Title/caption, Environment, production status, row-number ID) as a new row in a Google Sheet tracker.
- 3Generate Prompts (13-scene photo sequence) · Words · GPT-4.1 · OpenAIGPT-4.1 expands the Idea and Environment into a 13-scene cinematic third-person still-photo sequence, each scene under 500 characters, describing one frozen cleaning moment with no motion, poetic or camera language.
Prompt
You are a prompt-generation AI specializing in cinematic, third-person, still-image prompts. Your task is to generate a **13-scene photo sequence** showing frozen moments of sea animal cleaning or rescue operations. Your writing must match the following style: - Highly detailed, sharp cinematic realism. - Wide shots and mid shots showing both the divers and large portions of the whale. - Rough, physical cleaning scenes (scraping, spraying, grinding). - Strong environmental texture: mist, barnacle dust, sunlight shafts, water spray, wet decks, rough whale skin. - Cold, wet, gritty documentary feeling. - Absolutely no poetic, emotional, or storytelling language. - No fantasy, no movement, no camera terminology. Each prompt must: - Describe **only one frozen cleaning moment** per scene (no transitions, no relaxing, no preparation). - Be written in **cinematic third-person** style. - Stay **under 500 characters** per scene. - Match the provided **Idea** and **Environment** exactly. --- Core Inputs: **Idea:** "{{ $('Create New Idea').first().json.message.content.Idea }}" **Environment:** "{{ $('Create New Idea').first().json.message.content.Environment }}" --- ### Format: Idea: "..." Environment: "..." Scene 1: "..." Scene 2: "..." Scene 3: "..." Scene 4: "..." Scene 5: "..." Scene 6: "..." Scene 7: "..." Scene 8: "..." Scene 9: "..." Scene 10: "..." Scene 11: "..." Scene 12: "..." Scene 13: "..." --- ### Embedded Example Output (TRUE STYLE): Idea: "Divers cleaning barnacle clusters off a humpback whale" Environment: "Open ocean, clear sunlight piercing the water, whale covered in dense barnacles" Scene 1: A full-body shot of a diver gliding near the whale’s vast gray flank, positioning a scraper against thick barnacle patches as sunbeams streak through blue water. Scene 2: Side view of two divers working together along the whale’s scarred back, water swirling with broken shells and drifting debris around them. Scene 3: Low angle shot beneath a diver aiming a high-pressure hose upward at barnacle masses under the whale’s belly, bubbles rising into brilliant sun shafts. Scene 4: Top-down view of a diver pressing a wide scraper into the whale’s fin joint, barnacle chunks peeling away into cloudy plumes. Scene 5: Mid-range shot of a diver brushing stubborn barnacles from the base of the whale’s tail, small shells scattering across the glistening surface. Scene 6: Cinematic side angle of the whale’s head as a diver scrapes barnacles clustered around its massive eye, dust swirling in the water. Scene 7: A group of divers hovers along the spine, using brushes and scrapers in unison. Broken fragments spin around them under fractured sunlight. Scene 8: Close wide shot of a diver's gloved hand pressing a spatula under barnacles embedded in deep scars, gray dust spreading in the current. Scene 9: A mechanical cleaning arm brushes across the whale’s barnacle-coated back, sending rippling waves of foam into the surrounding sea. Scene 10: Low underwater shot capturing barnacle fragments falling from the cleaned flank like snow, illuminated by shifting sunlight beams. Scene 11: Wide shot showing a massive field of loosened barnacles floating over the whale as divers move steadily along its back. Scene 12: Sharp side view of a diver spraying low-pressure water onto the whale’s pectoral fin, soft mist clouding the bright blue background. Scene 13: Distant full-body view of the whale gliding through a debris cloud, its freshly cleaned surface gleaming under the scattered sunlight. - 4Unbundle Prompts · edit · n8n (Code node)A JS code node extracts the 13 Scene fields from the single scene object into 13 separate {description} items for parallel processing.
Prompt
let story; try { const sceneObject = items[0].json.message.content; if (sceneObject && typeof sceneObject === "object") { story = Object.entries(sceneObject) .filter(([key, value]) => key.startsWith("Scene")) .map(([key, value]) => value); // <<<< ADD THIS .map() TO GET ONLY SCENE TEXTS } else { throw new Error("No valid scene object found."); } } catch (e2) { throw new Error("Could not extract scenes properly."); } let output = []; for (const scene of story) { output.push({ description: scene }); } return output; - 5Create Image (per scene) · Picture · Flux Schnell or Flux Pro · fal.ai (queue.fal.run/fal-ai/flux/schnell or /fal-ai/flux-pro)Each scene description is sent to Flux (Schnell, cheap, or Pro, better) on fal.ai to generate a portrait still image; response polled from the fal.ai queue.
Prompt
{{ $json.description }} - 6Video Prompts (image-to-video-prompt) · Words · GPT-4o (chat GPT-4o latest) · OpenAI (Analyze image action)GPT-4o (or latest) analyzes the generated image plus the original photo prompt to write one short third-person cinematic video-motion prompt for that frozen moment (no 'I'/POV phrasing, one action only, <300 characters).
Prompt
You are a prompt-generation AI trained to create short, cinematic, third-person video prompts for underwater whale rescue operations. You will analyze an uploaded image and create one clear, realistic prompt describing a frozen moment of the cleaning process. Each prompt must match the style of professional marine rescue documentation. **Original photo-generation prompt for context:** "{{ $('Unbundle Prompts').item.json.description }}" Use this only to understand the role, setting, or tone. Do not copy or repeat it. --- ### Rules - No "I" or "POV" phrasing. - Short third-person sentences (under 300 characters). - Only one action (scraping, brushing, spraying). - Mention visible environmental textures (debris, barnacle dust, mist, bubbles). - No emotional language, no cinematic words like "camera pans." - One frozen cleaning action per sentence, no chaining actions. - Match cold, rough, realistic underwater marine rescue feeling. --- ### Example Outputs - Diver scraping barnacles off whale’s side under sunlit water. - Water spray clearing barnacle dust from whale’s rough fin. - Brush dislodging barnacle shells near whale’s dorsal ridge. - Tool peeling thick barnacle growth from whale’s underbelly. - Mist rising as diver blasts barnacle patches from whale’s back. - Shell fragments swirling as diver scrapes whale’s pectoral fin. - Gloved hand lifting barnacles off whale’s ridged tail surface. --- Use the uploaded image and matching description to generate one short underwater whale cleaning prompt in this style. - 7Create Video (image-to-video, per scene) · Video · Kling Standard or Kling Pro (v1.6, image-to-video) · fal.ai (queue.fal.run/fal-ai/kling-video/v1.6/standard or /pro/image-to-video)Each scene's image + video prompt is animated via Kling Standard or Pro image-to-video into a 5-second clip.
Prompt
{{ $json.content }} - 8Sound Prompts · Words · GPT-4.1 (text output, not JSON) · OpenAIA chat model writes a concise (<=250 characters) sound description for the scene, drawing on the environment idea, photo idea and video idea, for ElevenLabs-style sound matching.
Prompt
Generate a concise, immersive sound description for Eleven Labs based on the given environment. Remove unnecessary words and limit the output to 250 characters max, including spaces and symbols. The description must be precise, clear, and logically match the environment. No extra details, just essential sound characteristics. Environment Idea: {{ $('Log the Idea').item.json.environment_prompt }} Photo Idea: {{ $('Get Images').item.json.prompt }} Video Idea: {{ $('Video Prompts').item.json.content }} - 9Create Sounds (video-conditioned) · music · MMAudio v2 · fal.ai (queue.fal.run/fal-ai/mmaudio-v2)MMAudio v2 generates a 5-second ambient sound track conditioned on the sound-prompt text and the generated video clip itself.
Prompt
{ "prompt": "ambient musical background sounds for this idea: Underwater, sunbeams filter through blue water, whale and divers framed in cinematic realism, the action is: {{ $json.message.content }}", "duration": 5, "video_url": "{{ $('Get Video').item.json.video.url }}" } - 10List Elements · edit · n8n (Code node)A JS code node collects all 13 scenes' final video URLs (already sound-merged by MMAudio) into one ordered array for sequencing.
Prompt
return [ { video_urls: items.map(item => item.json.video.url) } ]; - 11Sequence Final Video · edit · ffmpeg-api/compose · fal.ai (queue.fal.run/fal-ai/ffmpeg-api/compose)ffmpeg-api/compose on fal.ai stitches the 13 clips end-to-end (5-second keyframes at 0,5,10...60s) into one ~65-second final video.
Prompt
{ "tracks": [ { "id": "1", "type": "video", "keyframes": [ { "url": "{{ $json.video_urls[0] }}", "timestamp": 0, "duration": 5 }, { "url": "{{ $json.video_urls[1] }}", "timestamp": 5, "duration": 5 }, { "url": "{{ $json.video_urls[2] }}", "timestamp": 10, "duration": 5 }, { "url": "{{ $json.video_urls[3] }}", "timestamp": 15, "duration": 5 }, { "url": "{{ $json.video_urls[4] }}", "timestamp": 20, "duration": 5 }, { "url": "{{ $json.video_urls[5] }}", "timestamp": 25, "duration": 5 }, { "url": "{{ $json.video_urls[6] }}", "timestamp": 30, "duration": 5 }, { "url": "{{ $json.video_urls[7] }}", "timestamp": 35, "duration": 5 }, { "url": "{{ $json.video_urls[8] }}", "timestamp": 40, "duration": 5 }, { "url": "{{ $json.video_urls[9] }}", "timestamp": 45, "duration": 5 }, { "url": "{{ $json.video_urls[10] }}", "timestamp": 50, "duration": 5 }, { "url": "{{ $json.video_urls[11] }}", "timestamp": 55, "duration": 5 }, { "url": "{{ $json.video_urls[12] }}", "timestamp": 60, "duration": 5 } ] } ] } - 12Log final video and mark done · edit · Google SheetsUpdate the Google Sheet row (matched by ID) with production status 'done' and the final video URL.
- 13Optional: publish to social · publish · BlotatoThe lesson points to a separate 'ultimate publishing agent' (built with Blotato, per Sabrina Ramonov) that reads the video URL from the sheet and cross-posts to TikTok/Instagram/YouTube.
RoboNuggets · R29 | The Monetizable Tiktoks AI Machine- 1
VeoRobo: Veo3-fast ASMR factory for any nichePartlya short ASMR-style cutting/slicing video in any chosen niche, built from randomized scene elements, generated via Veo3-fast and sequenced into one final clip
OpenRouter chat model (Elements + Prompt AI Agents)Veo3-fast (Kie.ai)ffmpeg-api/compose (fal.ai)ASMR cutting/slicing loop- 1Elements AI Agent · Words · OpenRouter chat model (exact model not specified in the extracted portion) · OpenRouterAn OpenRouter chat model, driven by the Elements AI Agent system prompt, randomly picks scene elements for an ASMR cutting scene (instrument, colour, material, object, surface, etc. from defined pools) and returns them as structured JSON.
- 2Unbundle Elements · edit · n8n (Code node)A JS code node unpacks the Elements Agent's structured JSON into individual fields for the next prompt step.
- 3Prompt AI Agent · Words · OpenRouter chat model (exact model not specified in the extracted portion) · OpenRouterA second OpenRouter chat model assembles the randomized elements into a final video_prompt using a fixed template, ready for Veo3.
- 4Create Veo3 video · Video · Veo3-fast (model: veo3_fast) · Kie.ai (/api/v1/veo/generate)The assembled video_prompt is sent to Kie.ai's Veo3-fast endpoint to generate the video clip.
- 5Poll for Veo3 completion · Check · Kie.aiWait/poll the Kie.ai Veo3 job until ready, then GET the finished clip URL.
- 6Group Clips · edit · n8n (Code node)A JS code node collects generated clip URLs (for multi-clip runs) into one ordered list ready for sequencing.
- 7Sequence Video · edit · ffmpeg-api/compose · fal.ai (/fal-ai/ffmpeg-api/compose)ffmpeg-api/compose on fal.ai sequences the clip(s) with tracks/keyframes into the final composed video.
- 8Get final video · Check · fal.aiPoll and fetch the final composed video URL from fal.ai.
- 9Log video to Google Sheets · edit · Google SheetsLog the final video URL and run metadata to a Google Sheet tracker.
- 10Optional: auto-publish · publish · BlotatoThe lesson references the R25 lesson's Blotato auto-publish step to cross-post the finished video.
Missing: the Elements AI Agent and Prompt AI Agent system prompts, the Unbundle Elements / Group Clips JS code, and the exact Kie.ai Veo3 / fal.ai compose request bodies were all viewed in the source lesson during an earlier read but did not survive verbatim into this rewrite pass; re-open knowledge/courses/robonuggets-school/all-lessons-and-templates/r34-veorobo-the-veo3-factory-for-any-niche--4432d9a1.md directly to recover the exact text before building.
RoboNuggets · R34 | VeoRobo: The Veo3 Factory for Any Niche- 1
Multi-scene narrated story videoRunsa multi-scene narrated video ad or story with AI images, voiceover, background music and burned-in captions, assembled scene-by-scene into one finished video
ElevenLabs turbo-v2.5NanobananaWan2.5SeedanceFFMPEG composeFFMPEG auto-captionnarrated story ad16:99:16- 1Read input row · By hand · Google SheetsReads special request, scene count, aspect ratio, ElevenLabs voice ID, image reference (source of truth for Nanobanana), background music (supports a RANDOM formula across a music sheet).
- 2Generate per-scene content · Words · not named · n8nAI Agent produces voice-over text, image prompt and video prompt per scene, following a narrative-arc guideline.
- 3Create voice · Voice · ElevenLabs turbo-v2.5 · fal.ai (fal-ai/elevenlabs/tts/turbo-v2.5)Generates narration audio for the scene.
- 4Create image · Picture · Nanobanana · Kie.aiGenerates the scene image from the reference image and image prompt.
- 5Create video · Video · Wan2.5 (6c/clip) or Seedance (15c/clip) · Kie.aiAnimates the scene image.
- 6Combine · edit · FFMPEG compose · fal.aiStitches video, voice and music tracks per an auto-generated FFMPEG command; uses a silence placeholder track when no music is set (FFMPEG doesn't auto-trim/lower volume).
- 7Caption · edit · FFMPEG auto-caption · fal.aiAuto-captions the video in white font using a custom TTF hosted on Cloudinary.
- 8Publish (optional) · publish · YouTubeAuto-publishes the finished video.
Missing: no verbatim per-scene generation prompt text anywhere in source; custom ElevenLabs voices are not supported via fal.ai (619 unpublished default voices only).
RoboNuggets · R40 | The Hyperconsistent AI Storymaker (with 🍌)- 1
Multi-model batch video generatorPartlya batch of short AI videos per run, generated in parallel through three different models from one auto-generated idea, for side-by-side comparison (both text-to-video and image-to-video variants)
Sora2Wan2.5Veo3short video batchtext-to-videoimage-to-video- 1Set inputs · By hand · n8nSchedule trigger and Edit Fields collect idea/topic, number of videos per run, aspect ratio, sample prompt structure (plus an image reference URL for the image-to-video variant).
- 2Analyse reference image (image-to-video variant) · Picture · OpenAI (not named) · OpenAIDescribes the reference image (n8n's AI Agent can't parse images natively).
- 3Generate idea + prompt · Words · not named · n8nAI Agent (A-G-E-N-T framework, has a Think tool and a Google Sheet lookup tool to avoid repeating prior ideas) produces title, caption, prompt, aspect_ratio JSON.
- 4Log idea · Check · Google SheetsAppends the generated idea as a new row.
- 5Generate video (Sora2 branch) · Video · Sora2 · Kie.ai (kie.ai/sora-2)Generates the video via Sora2 (text-to-video or image-to-video).
- 6Generate video (Wan2.5 branch) · Video · Wan2.5 · Kie.aiGenerates the same idea via Wan2.5 in parallel.
- 7Generate video (Veo3 branch) · Video · Veo3 · Kie.aiGenerates the same idea via Veo3 in parallel.
- 8Merge and save · Check · Google SheetsCombines all three models' output URLs into one row, side by side.
Missing: no verbatim AI Agent prompt in source. Sora2 output via Kie.ai carries a visible OpenAI watermark (Kie.ai taps the Sora2 web UI, not an official watermark-free API); Magic Eraser (manual) or Wavespeed are mentioned fixes but neither is wired into this workflow.
RoboNuggets · R43 | Sora2 on Autopilot- 1
Long video7
Long-form AI talking-head video (script + voice + photo-to-video)PartlyA long-form talking-head video of a still photo animated to speak a scripted, voiced narration.
HeyGen Photo to VideoVEED Fabric 1.0ElevenLabslong-form talking headexplainer- 1Write the script · Script · Claude · AnthropicWrite the narration script.
- 2Generate the voice · Voice · ElevenLabs V3 (recommended for realism) · ElevenLabsPaste the script into ElevenLabs Text to Speech with a chosen/built voice and generate.
- 3Have a starting character photo ready · PictureSource or generate the still photo that will be animated.
- 4Generate the talking video - route A (HeyGen) · Video · HeyGen Photo to Video · HeyGen (app.heygen.com)In HeyGen's Photo to Video tool, upload the audio file (never paste the script text directly - said to give a "shitty voice"), attach the character photo, turn on "more expressive motion" (no extra cost), generate.
- 5Generate the talking video - route B (VEED Fabric 1.0) · Video · VEED Fabric 1.0 · VEED (veed.io)In VEED (AI tools > AI video > Generate videos > Speech to Video > Veed Fabric 1.0), upload the reference character image and the reference audio, generate.
Missing: Sourcing/generating the starting character photo is not covered by either lesson (assumed done via the JSON image method used elsewhere in the course). Neither HeyGen nor VEED Fabric is shown via an API, only their web UIs.
The AI Content System · HeyGen / VEED Fabric 1.0- 1
Coding-agent explainer/documentary video projectPartlyan editable video project (60-second explainer, documentary, or talking-head shape) rendered to final.mp4, from a one-line brief
Piper TTS (free)ElevenLabs (paid, optional)documentaryexplainertalking head- 1Kick off · By hand · Claude Code / CursorOpen the OpenMontage project folder in Claude Code or Cursor and type a one-line brief, e.g. 'make a 60-second explainer about X'.
- 2Research · research · not stated · OpenMontage step 1Agent researches the topic.
- 3Proposal · Words · not stated · OpenMontage step 2Agent proposes the video's shape.
- 4Script · Script · not stated · OpenMontage step 3Agent writes the script.
- 5Scene plan · Words · not stated · OpenMontage step 4Agent breaks the script into a scene plan.
- 6Assets · Voice · Piper TTS (free) or ElevenLabs (paid, optional) · Pexels (free footage), ElevenLabs (optional)Free path: voice from a free offline TTS, footage from free stock libraries. Paid path: optional paid keys behind a spend guard that estimates cost first and can hard-stop.
- 7Edit · edit · Remotion / FFmpegAssemble the raw assets into a timeline.
- 8Render · Check · FFmpegRender, with a self-check after each step.
Missing: the seven per-step instruction-file prompts themselves are not reproduced in the source, only the step order and tools
AIOps · OpenMontage: A Coding Agent as a Video Production Line- 1
End-to-end video production inside Claude Code (script to render)Runsfull video productions — product demos, launch ads, 9:16 vertical short-form explainers, sprint-review videos — built and rendered entirely through Claude Code slash commands using Remotion
Qwen3-TTSElevenLabsFLUX.2Qwen-Image-EditIdeogram 4ACE-StepLTX-2.3SoulX-FlashHeadSadTalkerproduct demosprint-review9:16 short explainerconcept explainer- 1Create project · By hand · Claude CodeRun the /video command inside Claude Code, choosing a template (sprint-review, sprint-review-v2, product-demo, concept-explainer-short) and a brand profile.
- 2Write the script · ScriptEdit the generated VOICEOVER-SCRIPT.md file directly to plan narration/timing.
- 3Gather assets · Picture · FLUX.2 / Qwen-Image-Edit / Ideogram 4Use /record-demo to capture screen recordings, or generate stills/edits via FLUX.2 (image gen) or Qwen-Image-Edit (edits), with Ideogram 4 for title-card text-in-image work.
Prompt
Ideogram 4 title-card caption (verbatim worked example, .claude/skills/ideogram4/examples.md): {"high_level_description": "A dark modern tech title card with a large bold headline reading 'AI ENGINEERING REVIEW' centered in the upper-middle.", "style_description": {"aesthetics": "bold, modern, high-contrast corporate tech, generous negative space", "lighting": "soft cyan rim glow from below, deep shadows", "medium": "flat vector graphic design with subtle gradient backdrop", "art_style": "minimalist tech poster, clean geometric", "color_palette": ["#0B0B1A", "#00E0C6", "#1B2A4A", "#FFFFFF"]}, "compositional_deconstruction": {"background": "Deep near-black navy gradient with a faint geometric grid and a soft cyan glow rising from the lower edge.", "elements": [{"type": "text", "bbox": [340, 120, 540, 880], "text": "AI ENGINEERING REVIEW", "desc": "Bold heavy sans-serif headline, all caps, crisp white, tight letter spacing, centered.", "color_palette": ["#FFFFFF"]}, {"type": "obj", "bbox": [580, 380, 620, 620], "desc": "A thin glowing cyan horizontal accent rule beneath the headline.", "color_palette": ["#00E0C6"]}]}} Qwen-Image-Edit identity-preserving composite (verbatim worked example, .claude/skills/qwen-edit/examples.md), prompt: "The Irish woman with shoulder-length wavy dark auburn hair from first image is now standing confidently in the open plan office from second image. Same warm friendly smile, same facial features, black blazer, professional." — negative: "different ethnicity, different hair color, straight hair, different face shape, generic stock photo" - 4Scene review · Check · Remotion StudioRun /scene-review inside Remotion Studio to check scene composition before finishing.
- 5Visual refinement · edit · Claude CodeRun /design command to refine visual styling using Claude Code's frontend-design capability.
- 6Generate voiceover audio · Voice · Qwen3-TTS or ElevenLabsRun /generate-voiceover to produce AI narration from the script, using Qwen3-TTS (9 voices + cloning) or ElevenLabs as the voice provider, per the brand's voice.json.
Prompt
ElevenLabs text_to_speech.convert() call (verbatim from .claude/skills/elevenlabs/SKILL.md): text="Welcome to my video!", voice_id="JBFqnCBsd6RMkjVDRZzb" (George, warm narrator), model_id="eleven_multilingual_v2", voice_settings=VoiceSettings(stability=0.5, similarity_boost=0.75, style=0.5, speed=1.0)
- 7Add music · music · ACE-StepGenerate background music via ACE-Step (free cloud API, or ~$0.05 self-hosted), using a caption for overall style and structure-tagged lyrics for vocal sections.
Prompt
Caption (verbatim from .claude/skills/acestep/SKILL.md worked example): "Upbeat indie rock anthem, male vocal, driving drums, electric guitar, studio polish" Lyrics (verbatim, structure-tagged): [Verse 1] Walking through the morning light Coffee in my hand feels right Another day to build and dream Nothing's ever what it seems [Chorus - anthemic] WE KEEP MOVING FORWARD Through the noise and doubt We keep moving forward That's what it's about [Verse 2] Screens are glowing late at night Shipping code until it's right The deadline's close but so are we Almost there, just wait and see [Chorus - bigger] WE KEEP MOVING FORWARD Through the noise and doubt We keep moving forward That's what it's about [Outro - fade] (Moving forward...)
- 8Video generation (where needed) · Video · LTX-2.3 / SoulX-FlashHead / SadTalkerGenerate any needed text-to-video or image-to-video clips via LTX-2.3, and talking-head shots via SoulX-FlashHead (or SadTalker for cheaper drafts).
Prompt
LTX-2.3 good-prompt examples (verbatim, .claude/skills/ltx2/SKILL.md): "Aerial drone shot slowly flying over turquoise ocean waves breaking on white sand, golden hour sunlight, cinematic" / "Close-up of hands typing on a mechanical keyboard, shallow depth of field, soft desk lamp lighting, cozy atmosphere" / "Dark moody abstract background with flowing blue light streaks, subtle geometric grid, bokeh particles floating, cinematic tech atmosphere"
- 9Preview, iterate, render, publish · publish · Modal or RunPod (GPU); Cloudflare R2 (storage)Preview live via npm run studio, iterate on timing/styling/content, then npm run render to produce the final MP4, deployed to Cloudflare R2 via /setup.
Missing: README itself shows no literal generation prompts; steps 3, 6, 7, 8 above now carry verbatim worked examples pulled directly from the toolkit's own bundled per-model skill files (.claude/skills/ideogram4, qwen-edit, elevenlabs, ltx2, acestep) via raw.githubusercontent.com. Other skills in the tree (.claude/skills/ffmpeg, frontend-design, moviepy, playwright-recording, remotion-official/*) and _internal/toolkit-registry.json were not individually opened.
Free online · claude-code-video-toolkit- 1
Seedance longform UGC production pipeline (30s+)PartlyA 30s+ longform UGC video built from a 6-section script architecture and a multi-angle character reference system.
SeedanceGemini OmniHeyGenMuAPI- 1Build persona reference set · PictureMulti-Angle Character Reference System (MAS): 9+ reference images per persona, stored under 02-Brands/vero/creative/avatars/<name>/.
- 2Write the script · Script · Claude (as 'Seedance Prompt Architect')6-Section Longform Script Architecture: Hook / Persona Signal / Pain Amplification / Mechanism Bridge / Product Intro & Proof / CTA; audio-read gate.
- 3Generate shot-by-shot · Video · Seedance / Gemini Omni (model-selection flowchart by cost)Shot-by-shot Seedance generation with a mandatory imperfection_spec field.
- 4Assemble + optional lipsync · edit · HeyGen / MuAPIAssembly, optional HeyGen/MuAPI lipsync.
Missing: Architecture is literal; per-shot prompts are dynamic, not pre-written.
Ours ·- 1
Narrative-chained video extension (Veo3)Runsextended-length video by chaining short Veo3 clips end-to-end
Veo3 fastlandscapeany length via chaining- 1Generate first clip · generate-video · Veo3 fast · Kie.aiGenerate an 8s Veo3 clip from a text/image prompt
- 2Extract last frame · extract-frame · fal.aiAnalyze the generated clip and extract its final frame as an image
- 3Generate next clip from last frame · generate-video · Veo3 fast · Kie.aiUse the extracted last frame as the starting image for a new Veo3 image-to-video generation, continuing the narrative
- 4Repeat and merge · merge-video · fal.aiRepeat steps 2-3 for each additional scene, then merge all clips via ffmpeg into one continuous video
Missing: exact n8n node structure and merge-tool name not fully specified
RoboNuggets · n16 - Narrative Chaining with Veo3- 1
Longform YouTube creator AI agent (JSON2Video)Partlya longform narrated YouTube video (10-scene narrative structure) assembled from AI voice, AI images, music and auto-subtitles via a JSON2Video template, with optional free-tier image/voice swap
GPT-4/4.1 (chat, ideas + scene prompts)ElevenLabs or Azure voice (TTS)Flux-Pro or Flux-Schnell (image)Suno (music)Pexels (stock intro video)narrated longform story/explainer video- 1Generate video ideas (Ideas Agent) · Words · GPT-4/4.1 · OpenAIA chat model produces longform video ideas in a 9-column table format (covering things like title, hook, topic, angle) for the channel niche.
- 2Generate scene-by-scene script (Prompts Generator) · Script · GPT-4/4.1 · OpenAIA chat model expands the chosen idea into a 10-scene narrative structure: per-scene narration text plus a matching image-generation prompt for each scene.
- 3Assemble full video via JSON2Video · Video · JSON2Video (template ID 2oW2DZtCPfAbhck2tdXt) wrapping ElevenLabs/Azure voice, Flux-Pro/Flux-Schnell images, Suno music, Pexels intro video, auto-subtitlesA single JSON2Video 'create movie' POST request assembles the whole video from one template: text-to-speech narration, AI images per scene, background music, an intro stock clip, and automatic subtitles, in one render call.
- 4Poll render status with retry loop · Check · JSON2VideoA wait node + status GET + switch/IF node polls the JSON2Video job until done, running, or error; on error it logs to an Error Log and retries.
- 5Publish to YouTube · publish · YouTubeThe completed video is uploaded to YouTube as the final step.
Missing: the Ideas Agent Prompt, the Prompts Generator system prompt (10-scene structure), the full JSON2Video POST body and the Appendix A full template JSON were all viewed in the source lesson during an earlier read but did not survive verbatim into this rewrite pass; re-open knowledge/courses/robonuggets-school/all-lessons-and-templates/r28-the-longform-youtube-creator-ai-agent--69b7b9b2.md directly to recover the exact text before building.
RoboNuggets · R28 | The Longform YouTube Creator AI Agent- 1
Long-form sleep story videoPartlya long-form (up to ~2 hour) narrated 'sleep story' video: chapters of AI images plus narrated voice plus ambient background music or silence, stitched into one long video
SeedreamChatterboxChatterbox multilingualFFMPEG composeFFMPEG merge-videossleep storylong-form narrated video- 1Read input / chapters · By hand · Google SheetsChapter/snippet rows are pre-populated via a companion external 'Ideator GPT'.
- 2Loop chapters · By hand · n8nLoops over chapters, default batch size 1 (raise to run chapters in parallel pairs).
- 3Create images · Picture · Seedream · Kie.aiGenerates the chapter's images.
- 4Create voice (narration) · Voice · Chatterbox or Chatterbox multilingual · fal.ai (fal-ai/chatterbox/text-to-speech, .../multilingual)Narrates the chapter text; keep input to roughly 70-75 words per generation or output runs long/ignores the script (hard 30-second output cap).
- 5Combine chapters · edit · FFMPEG compose · fal.aiCombines chapter video/audio with background music or a silence/rain ambience placeholder; roughly 30 min render for a 2-hour video.
- 6Merge · edit · FFMPEG merge-videos · fal.aiMerges all chapter videos into the final long-form video.
- 7Save output · Check · Google SheetsStores the final video URL.
Missing: the source lesson's ~155,000-character video transcript could not be fully retrieved by the researching agent (single-line, exceeds read-tool size cap); all other sections (frontmatter, timestamps, Resources, Tools, Tips) were read in full. No verbatim generation prompts were found in the readable sections. Exact node sequence is inferred from timestamps/Tips, not transcript-confirmed line by line. Ideator GPT is external/manual.
RoboNuggets · R42 | The Sleep Story AI Generator- 1
Music video3
K-pop AI cover music video (voice clone + character bible + segmented video)Partlya 9:16 short-form AI cover music video (~3 minutes source, cut into segments) with a cloned idol voice singing new lyrics, a visually consistent AI character across stage and behind-the-scenes shots, and word-level pop-punch subtitles
Suno V5.5 (voice clone)Kling 2Hailuo 02Wan 2.7SunoMV (visualizer, beat alignment, subtitle export)DomoAI (anime/V-tuber variant only)AI cover music videoK-pop stage/BTS split- 1Clone the singing voice · Voice · Suno V5.5 · SunoTake a clean 30-second a-cappella sample (official release, ballad chorus, or radio live), train a voice profile in Suno V5.5, then generate a full cover with new lyrics/language. Blind-test the result with five listeners to confirm voice recognition.
- 2Build the character bible · PictureDefine two parallel reference sets for the same character sharing identical identity — a 'stage performance' set and a 'behind-the-scenes' set — varying only wardrobe, hair, makeup, lighting and camera distance between them.
- 3Segmented video generation · Video · Kling 2 / Hailuo 02 / Wan 2.7 / SunoMV visualizer · Kling / Hailuo / SunoMVSplit the 3-minute cover into song sections and generate each with a different engine by cost/quality need: Kling 2 with up to 4 reference images for expensive 'killing part' chorus shots, Hailuo 02 with 1 reference for cheap verse/BTS shots, Wan 2.7 for bridge (chained via first-frame-to-video), SunoMV visualizer for free transition segments.
- 4Beat-align the cuts · edit · SunoMV · SunoMVUse SunoMV to auto-extract word-level timestamps from the vocal track, tag the 'killing part' start frame (first syllable of chorus's first line), and align every chorus clip's cut-in to that frame. Cut every 5 seconds within the chorus (high density) and every 10 seconds in the bridge.
- 5Export as 9:16 viral short · publish · SunoMV viral-shorts generator · SunoMVExport via SunoMV's viral-shorts generator in 9:16 with word-level pop-punch subtitles color-matched to the artist's palette, positioned in the bottom third, never overlapping the face.
Missing: no literal image or video prompts are given anywhere in the source — every step is described at the method/settings level only; would need to write the actual character-bible and video-generation prompts from scratch using the field structure given
Free online · K-pop AI Cover Music Video Workflow: Suno V5.5 Voice Clone + Kling Character Lock + SunoMV- 1
Lo-fi ambient loop video (manual HaiLuo + Suno + DaVinci)ManualA 1-hour ambient/lo-fi looping video (a short steady-shot cozy scene looped, over a generated lo-fi music track) for a faceless YouTube channel.
HaiLuo AI (video)Suno AI (music)ambient lofi loop- 1Write the video prompt · Script · ChatGPT · OpenAIChatGPT expands bullet-point scene ideas into a detailed, steady-shot video prompt.
Prompt
NO CAMERA MOVEMENT. STEADY SHOT. FIXED SHOT. ... a cozy living room... fireplace crackling... camera steady... weather is autumn
- 2Generate the loop clip · Video · HaiLuo AI · HaiLuo AIHaiLuo AI generates a 5-second looping ambient video clip from the prompt (manual, web UI, no sign-up).
- 3Write the music prompt · Script · ChatGPT · OpenAIChatGPT, in the same chat, writes a matching lo-fi audio prompt.
Prompt
Create a soothing lo-fi track with gentle crackling fire sounds, soft piano melodies, and light autumn ambiance...
- 4Generate music · music · Suno AI · Suno AISuno AI generates a ~2-minute lo-fi music track from the prompt (manual, free tier, web UI).
- 5Loop and assemble · By hand · DaVinci ResolveDaVinci Resolve, using a provided .drp template project, loops/crossfades the 5s video and 2min audio into a full 1-hour video, with pre-configured auto-cropping to hide the HaiLuo watermark.
- 6Publish (manual) · By hand · YouTubeManual export and YouTube upload.
Missing: Fully manual/UI-driven pipeline; the lesson explicitly states this is not yet automated (no APIs used for HaiLuo, Suno or DaVinci in this version).
RoboNuggets · R10 | AI Automated Lofi Videos - Faceless YouTube- 1
Auto music album video creatorRunsa single long-form video (up to roughly an hour) combining N AI-generated Suno songs (instrumental or with lyrics) stitched to one static/looping cover image, optionally auto-published to YouTube with timestamps
Suno V5FFmpeg (via ffmpeg-api.com)GPT-4.1music album video- 1Create FFmpeg working folder · edit · ffmpegapi.comCreates a directory for uploads on the FFmpeg service.
- 2Intake form · By hand · n8nCollects theme, music description, number of songs, instrumental true/false.
- 3Upload cover image · Picture · ffmpegapi.comCreates a file path in the FFmpeg service and uploads the user's cover image there.
- 4Generate song prompts · Words · GPT-4.1 · OpenAIGiven theme, description and song count, produces N song objects (prompt, style, title).
- 5Split songs · edit · n8nUngroups the song array into individual items.
- 6Create songs · music · Suno V5 · Kie.aiPosts each song's prompt/style/title/instrumental flag to Suno.
- 7Poll songs · Check · Kie.aiPolls per-song status, branching on success vs in-progress.
- 8Log songs · Words · Google SheetsAppends song URL, title, duration, theme.
- 9Upload songs to FFmpeg service · edit · ffmpegapi.comCreates a per-song file path, fetches the MP3 from the Suno URL, uploads it to FFmpeg storage.
- 10Build FFmpeg command · edit · n8nA code node constructs the FFmpeg command string that concatenates all songs and loops the cover image.
- 11Combine (render) final video · edit · FFmpeg · ffmpegapi.comStitches songs and looped cover image into one long MP4.
- 12Log final video · Words · Google SheetsAppends final video URL, theme, timestamp.
- 13Auto-publish to YouTube (optional) · publish · YouTubeUploads the final video.
- 14Timestamp generation (optional) · Words · custom GPT ('Ideator and Timestamp GPT') · OpenAIPastes song titles/durations into a custom GPT to auto-generate YouTube timestamps.
Missing: agent system prompt for song-prompt generation, and the FFmpeg command-building code, are described but not quoted verbatim.
RoboNuggets · R45 | Auto Music Creator with Suno v5- 1
Static ad16
Static ad creation (research to Canva)Partlya static image ad built from customer research through to a finished Canva template, using AI for image-replace and copy generation
ChatGPTPoppyCanva template ad- 1Customer research · researchCollect psychographics, problems/desires, moments, and objections from Reddit, TikTok comments, Amazon reviews and forums; compile into one research document.
- 2Pick an avatar · Words · an unnamed custom GPT · ChatGPTFeed the research document into a named GPT to get 3 avatar options; pick one specific persona.
- 3Concept/industry research · research · ChatGPT · 4Print / Atria / adnova.aiApply a 70/20/10 sourcing ratio (steal from proven ads / reuse own winners / new swings). Use AdSpy tools 4Print and Atria (running-length based) and adnova.ai (beta, hook/angle data from competitor traffic). Also ask ChatGPT for adjacent-niche ideas. Time-box to 1 hour; end with 10 competitor + 10 adjacent-niche ideas.
- 4Build an angle bank · ScriptRun the research document through the course's linked 'angles prompt': analyzes research, finds mass desires, asks you to choose one, builds a bank of angles organized by awareness level.
- 5Pick a template and replace the image · Picture · ChatGPT / Poppy · ChatGPTSelect from 10 proven Canva templates. In ChatGPT or Poppy, prompt an image replace to swap in your product while keeping the template's style/composition.
Prompt
create an image, replace the image of the can, that's this can here, with the product that we have here and the others photo. So this other photo with a generated pill bottle that says Terkestron, keep everything else the same.
- 6Generate the copy · Words · ChatGPT / Poppy · ChatGPTPrompt for benefits/negatives per awareness level and five headlines, based on the concept and angle bank.
Prompt
Based on this concept and the angle bank, give me the benefits and negatives I should list out. Make one set for each awareness level. Also give five headlines for this ad.
- 7Assemble in Canva · edit · CanvaPaste the replaced image and generated copy into the Canva template, adjust resizing/recoloring.
Missing: the 'angles prompt' (step 4) is a linked course asset not captured verbatim; the customer-research process (step 1) is described only
AI Ads (Ad Creators Lab) · Our exact process for creating our static ads / Creating Effective Static Ads- 1
Winning-ad pattern to new static ad variationPartlynew static ad creative variations, generated from what is already winning on a Meta ad account, launched as paused ads only
Anthropic Claude (vision tagging)Gemini (image generation)static Meta ad variation- 1Scaffold · By hand · Next.js/SQLiteScaffold a local Next.js app with a SQLite database (Drizzle) holding 7 tables: creatives, metrics, analysis, insights, generations, launches, reference_ads.
- 2Set up Meta access · By hand · Meta Marketing APISet up Meta access via a system user token (never personal login) with ads_management/ads_read/business_management/page permissions; verify a local Meta CLI can list campaigns/ad sets/pages/datasets before building UI.
- 3Sync creatives and metrics · By hand · Meta Marketing APIPull ad images, map ads to creatives, download images, store spend/CTR/conversions/CPA/ROAS.
- 4Vision-tag every ad · Check · Anthropic (Claude, vision) · Anthropic SDKTag every ad image into strict JSON: composition, palette, person present, text overlay, mood, hook type, target avatar, pain point, proof elements, objections, variation ideas. Prompted as a senior performance marketer, forbidden from inventing brand facts.
- 5Pattern discovery · Check · the app (rule-based grouping)Group creatives by attribute, compare to account averages, surface only patterns with >=3 samples and >=1.3x lift as 'test recipes'.
- 6Generate a variation · Picture · Gemini image generation · Gemini APIBuild a prompt from a seed creative, its analysis, a chosen insight, user direction, and optionally a reference ad from a swipe board. Seed ad is the identity source, reference is inspiration only, never copy another brand's claims or packaging.
- 7Launch paused · publish · Meta Marketing APILaunch as a paused ad only, with dry-run preview and validation; a human activates it in Ads Manager.
Missing: the 12-phase build order, schema, and guardrails are specified but the exact prompt text for the vision-tagging and generation calls is not captured in the source
AIOps · AdPulse: A Local AI Creative Testing Cockpit for Meta Ads- 1
Ad creative variant generator (image analysis + GPT-5 recipe-based rewriting + Nano Banana edit)Runsmultiple ad creative image variants from one existing ad image, rewritten according to a selected creative methodology (metaphor, feature-pointer callouts, or before/after comparison) and regenerated as new images
gpt-4o (vision)gpt-5fal-ai/nano-banana/editstatic ad image set- 1Describe ad selected · Words · gpt-4o · OpenAI (vision, via n8n langchain node)Analyzes the uploaded reference ad image and produces an exhaustive text description of it.
Prompt
describe the image extremely comprehensively. don't leave anything behind
- 2Create variants of the prompt · Words · gpt-5 · OpenAI (via n8n langchain node)Rewrites the original image description into N new variant descriptions, applying a selected creative methodology (The metaphor bake / The pointer platter / The comparison combo, or none) plus user instructions and copywriting inspiration, keeping the company name consistent across all variants. Returns strict JSON.
Prompt
SYSTEM: You are a helpful, intelligent prompt rewriting assistant. You help rewrite prompts USER: You are tasked with generating {{ number_ads }} image description variants for ad creatives based on a specific creative methodology. INPUT PARAMETERS: - Company Name: {{ company_name }} - Number of Ads to Generate: {{ number_ads }} - Creative Recipe: {{ ad_creative_recipe }} - User Instructions: {{ user_comments }} - Copy writting inspiration: {{ copywriting inspiration field }} ORIGINAL IMAGE DESCRIPTION: {{ image_description }} CREATIVE METHODOLOGY GUIDE: **IF "The metaphor bake" is selected:** - Turn the key message into a visual metaphor or analogy - Use figurative imagery (objects, scenes, symbols, idioms, cultural references) - The visual should communicate the message clearly without needing extra copy - Make it instantly understandable and human, not overly conceptual - Ensure the product/brand connects logically back to the metaphor - Examples: Running after money on a fishing rod (chasing gains), board game representing guesswork **IF "The pointer platter" is selected:** - Show the product UI or mockup clearly - Feature 3-4 visual pointers/callouts highlighting key features or capabilities - Keep labels short (1-3 words) and position them around the UI with arrows or pills - Make the value proposition unmissable at a glance - Ensure all text elements stay readable and not cluttered - Focus on product education and clarity **IF "The comparison combo" is selected:** - Show a clear Before/After, Us vs. Them, With vs. Without, or Old Way vs. New Way comparison - Split the visual into two contrasting sides - Use visual or textual elements to show the "hard way" vs. the "better way" - Make the contrast unmistakable even without reading text - Use color, structure, or layout to emphasize the difference - Show transformation or improvement that the product creates **IF NO recipe is selected or field is empty:** - Keep the style and approach similar to the original image description - Maintain the same visual language and messaging approach - Focus on clarity and brand consistency TASK: 1. First, determine which creative recipe applies (or if none selected) 2. Modify the original image description according to the recipe, user instructions, and original description 3. Generate the requested number of additional variants with modifications to color schemes/backgrounds, visual styling/composition, copy/text positioning/wording, character poses/UI layouts/comparison elements, creative execution details. RULES: keep company name consistent; copywriting must relate to the company and inspiration provided; apply BOTH the recipe AND user instructions to ALL variants; ensure descriptions are detailed enough for image generation; variants should feel cohesive as one campaign. OUTPUT FORMAT (strict JSON): {"variants":["First variant...", "Second variant...", ...]} (truncated by template length — full field confirmed present in the real template JSON at api.n8n.io/api/templates/workflows/11163) - 3Submit Request to generate image · Picture · fal-ai/nano-banana/edit · Fal.aiSends each variant description plus the original base64 image to fal.ai's nano-banana edit endpoint to regenerate the ad as a new image.
Prompt
REQUEST BODY (verbatim): {"prompt": "{{ variant text }}", "image_urls": ["data:image/png;base64,{{ image }}"]} - 4Poll + download · Check · Fal.aiPolls the fal.ai queue for job completion, then downloads the generated image variants.
Missing: GPT-5 variants JSON output schema example is truncated in the source template's parameter string itself (not a fetch failure — the template's own field is cut short in the same place both times it was read)
Free online · Generate ad creative variants using AI image analysis and editing- 1
Influencer-holding-product composite image (Gemini 2.5 Flash Image edit)Runsa single composite photo of a supplied influencer/person photo appearing to hold and drink from a supplied product image, styled as a casual candid cafe photo
gemini-2.5-flash-image-previewproduct/UGC composite image- 1Convert images to base64 · By hand · n8nConverts the supplied product image and influencer/person image to base64 for the API call.
- 2generate_image · Picture · gemini-2.5-flash-image-preview · Google Gemini API (generativelanguage.googleapis.com)Sends a fixed scene-directing prompt plus both images to Gemini 2.5 Flash Image (preview) to composite the influencer holding/about to drink from the product cup/tumbler, at a table in a cafe, smiling at camera, candid friend-photo styling, slight side angle (varied).
Prompt
Create an image where the cup/tumbler in image 1 is being held by the person in the 2nd image (like they are about to take a drink from the cup). The person should be sitting at a table at a cafe or coffee shop and is smiling warmly while looking at the camera. This is not a professional photo, it should feel like a friend is taking a picture of the person in the 2nd image. Only return the final generated image. The angle of the image should instead by slightly at an angle from the side (vary this angle).
Missing: none — corrects first-pass error which wrongly said this source had no visible prompt; the full verbatim prompt is confirmed present in the real template JSON
Free online · Create UGC-style product photos by merging product and influencer images- 1
Auto-written product photography prompt + Gemini Nano Banana compositeRunsa styled product photo composited from a product image plus up to two reference images, using an LLM-written photography-directive prompt generated from the product's name/description and the requester's stated requirement
gemini-2.0-flash-litegemini-2.5-flash-imageproduct photo- 1Collect product + requirement via Jotform · By hand · JotformTakes Product Name, Product Description, a Product Image, and a free-text Requirement (the desired shot type) plus up to two reference images from a Jotform submission.
- 2Message a model · Words · gemini-2.0-flash-lite · Google Gemini API (via n8n langchain node)Acts as a 'professional product photographer' and writes a single image-generation prompt from the product name/description and stated requirement, returning only the prompt with no extra commentary.
Prompt
You are a professional photographer who do Product protography. Below are the details of a product and what type of photogaric image he need. Write down a prompt for an AI model to generate that image: Product: {{ Product Name }}, {{ Product Description }} Product image: {{ Product Image }} Requirement: {{ Requirement }} Note: Just provide the prompt directly, no need to write extra text e.g here is the prompt or something - 3Gemini Nano Banana · Picture · gemini-2.5-flash-image · Google Gemini API (generativelanguage.googleapis.com)Sends the generated photography prompt plus up to 3 inline images (product + 2 reference images) to Gemini 2.5 Flash Image to produce the final composited product photo.
Prompt
prompt text = Message a model output verbatim (newlines stripped, trimmed); passed alongside up to 3 inline_data image parts
Missing: none — both generating steps confirmed verbatim from the real template JSON
Free online · Generate product photos with AI using Google Sheets and Jotform- 1
Static ad generated from a brand's website + logo + product photoRunsOne finished static advertisement image (headline, subheadline, CTA, on-brand visual) built entirely from a brand's own website content plus an uploaded logo and optional product photo - no manual copywriting.
Claude Sonnet 4.5 (claude-sonnet-4-5-20250929)Nano Banana / Gemini 3 Pro Image (google/gemini-3-pro-image-preview)value-proposition adB2B SaaS adUI-led ad- 1Collect brand inputs · By hand · n8n form triggerA form collects the brand's Website URL (required), Logo file (required, .jpg/.png/.webp) and an optional Product photo. Uploaded files are split into separate binary items for later steps.
- 2Classify the uploaded product photo · Check · claude-sonnet-4.5 (anthropic/claude-sonnet-4.5) · OpenRouterClaude looks at the product photo and returns a strict-JSON classification (category, confidence, description, and booleans for is_ui/is_person/is_product/is_packaging/is_illustration) so later steps know how to treat the asset (e.g. a UI screenshot must stay readable, not be cropped/blurred).
Prompt
Analyze the provided Product_photo and classify what it is. Return strict JSON only with this structure: {"category":"","confidence":"","description":"","is_ui":false,"is_person":false,"is_product":false,"is_packaging":false,"is_illustration":false}. Choose category from exactly one of: person, product, packaging, ui, dashboard, object, icon, illustration, abstract_graphic, scene, other. Base the answer only on the image. Do not guess the brand or business context. - 3Detect the website's language · Check · claude-sonnet-4-5-20250929 · AnthropicA small agent looks only at the submitted domain (not the page content) and infers the working language from its TLD, so later copy-writing steps know what language to write in.
Prompt
Identify Language of the dommain: Input domain: {{ $('Base inputs').item.json.Website }} - Identify the top-level domain (TLD), including multi-part endings like .com.pl or .co.uk. - If the TLD corresponds to a specific country, output that country’s dominant language. - If the TLD is generic or not tied to a country (e.g., .com, .co, .io, .ai, .net, .org, .app, .dev, .tech), DEFAULT to English. - If unsure or ambiguous, DEFAULT to English. Return only the language as a plain string. Example: "English" No JSON. No arrays. No extra text. - 4Fetch homepage and pick useful internal pages · research · claude-sonnet-4-5-20250929 · AnthropicFetches the homepage, extracts nav links, then an agent selects up to 4 internal page handles likely to carry the strongest marketing copy (features/pricing/solutions/etc.), excluding careers/legal/login/localized pages; those pages are then fetched and converted to text.
Prompt
From the list of website handles, select up to 4 handles that will likely contain the most valuable marketing insights, brand tone, and product messaging. Exclude the main homepage. List: {{ $json.navLinks }} PRIORITIZE HANDLES LIKE: /product/, /features/, /solutions/, /platform/, /pricing/, /customers/, /use-cases/, /blog/, /resources/, /why-us/, /about/ IGNORE: careers, legal, login, language variants (/de/, /fr/), support pages. OUTPUT FORMAT (strict): [ "/first-handle/", "/second-handle/", "/third-handle/", "/fourth-handle/" ] - 5Build structured brand insights · Words · claude-sonnet-4-5-20250929 · AnthropicCleans and joins all fetched page text into one block, then an agent extracts strictly-sourced brand facts (product name, summary, customer group, problems, solution, key features, key benefits, brand voice, business type) plus a ~250-word marketing brief - explicitly forbidden from inventing anything not in the text.
Prompt
You are a B2B SaaS brand analyst. Your task is to extract structured information strictly from the provided website text. You must maximize the amount of information extracted, but: Do not invent facts. Do not guess or rely on prior knowledge. Only return elements that are directly supported by explicit text or strongly implied within the same paragraph or section. If any element cannot be found, return "Unknown" or []. Use short, clear phrasing. Never include marketing fluff - only what the text clearly states. WEBSITE TEXT {{ $json.json_safe }} TASKS Product / Brand Name Extract only if clearly visible in the text. Prefer headings, titles, product descriptions, hero sections, or navigation labels. Remove domain suffixes (.com, .io, etc.). Capitalize first letter. If multiple appear, pick the most prominent. If unsure -> "Unknown". Look at the website url to determine. And how many times it is used. Product Summary 1-2 neutral, factual sentences describing what the product does, based ONLY on the text. Customer Group Who the product is for (ICP). Extract exactly as worded. If multiple groups appear, just write them afer,. Problems List the core pain points described or clearly implied. Solution Short, factual explanation of how the product solves the problems. Key Features Extract main functional capabilities. Do not turn benefits into features. Key Benefits / Differentiators Extract value outcomes explicitly stated. Brand Voice & Tone Describe the tone (e.g., friendly, technical, formal). If unclear -> "Unknown". Business type Short classification (max 3 words) of the business type helping to identify the positioning like (B2B SaaS, Ecommerce, Services etc.) --- After extracting all fields, write a **~250-word marketing brief** synthesizing ONLY the extracted information. Rules for the brief: - ~250 words. - No fluff. - Very dense and insight-heavy ("raw meat"). - Explain the ICP's world, their frustrations, and motivations. - Show how the product fits the problem -> solution -> outcome arc. - Maintain the extracted tone. - No made-up facts. --- OUTPUT FORMAT (Valid JSON Only) { "product_name": "", "product_summary": "", "customer_group": "", "problems": [], "solution": "", "key_features": [], "key_benefits": [], "brand_voice": "", "marketing_brief": "", "business_type":"" } Return ONLY this JSON object. No explanations, no markdown, no extra text. - 6Write the ad concept (copy + visual direction) · Script · claude-sonnet-4-5-20250929 · AnthropicAn agent turns the brand insights + the product-photo classification into 2 headlines, 2 subheadlines, 2 CTAs, plus a described visual twist, main visual, layout and style direction, with hard rules per element (e.g. headline max 6 words, one phrased as a question; if the asset is a UI it must stay the dominant, undistorted focal element).
Prompt
You are a senior {{ $json.output.business_type }} marketing strategist, creative director, and conceptual advertising designer known for bold, intelligent, and memorable campaigns. Your task is to create a complete value-proposition advertisement concept that combines clear product value with playful, unexpected creative expression while remaining credible for a professional {{ $json.output.business_type }} audience. All reasoning must rely strictly on the provided inputs. Do not introduce external assumptions, examples, or predefined metaphors. -------------------------------- LANGUAGE Create all text in: -------------------------------- PRIMARY CONTEXT (USE ONLY THIS) Brand Name: {{ $json.output.product_name }} Brand Summary: {{ $json.output.product_summary }} Customer Group: {{ $json.output.customer_group }} Key Benefits: {{ $json.output.key_benefits }} Key Features: {{ $json.output.key_features }} Desired Outcomes: {{ $json.output.key_benefits }} MARKETING BRIEF {{ $json.output.marketing_brief }} -------------------------------- ASSET INTERPRETATION (MANDATORY) {{ $('Merge').item.json.Product_description }} Determine: - what type of visual it is (character, UI/dashboard, object, icon, abstract graphic, or other) - what real product meaning or function it represents - how it should structurally influence the advertisement composition The final visual concept must always be built around this asset. The asset must never be decorative or optional. UI-SPECIFIC RULE: If the asset is a UI/dashboard: - the UI must be clearly visible, readable, and undistorted - it must occupy the primary focal area of the advertisement - the layout must be designed to showcase the UI as the main proof of value - no heavy perspective, blur, cropping, or decorative obstruction is allowed - supporting visuals may exist only to reinforce the UI, never compete with it -------------------------------- CORE CREATIVE REQUIREMENT The advertisement must: -> communicate real product value clearly and instantly -> use an unexpected, bold, or playful visual idea -> remain intelligent, professional, and credible for {{ $json.output.business_type }} Playfulness must serve clarity, not replace it. -------------------------------- PART 1 - COPY CREATION HEADLINES Generate exactly 2 headlines. Maximum 6 words each. Requirements: - clear connection to real product value - confident, bold, or witty tone - still understandable for professional audience - do not mention the customer group - one headline must be phrased as a question - no unusual spelling SUBHEADLINES Generate exactly 2 subheadlines. Maximum 14 words each. Requirements: - explain how the product delivers the value - grounded in real features or benefits - concrete and believable - support the bold tone without exaggeration CTA TEXT Generate exactly 2 CTAs. Maximum 3 words each. Requirements: - short, confident, action-oriented - natural for {{ $json.output.business_type }} interaction - no unusual spelling - plural 's' must be lowercase -------------------------------- PART 2 - VISUAL STRUCTURE CREATIVE CONTRAST LOGIC Define: - the unexpected or playful visual twist - how the twist reveals real product value - why the idea remains credible for {{ $json.output.business_type }} MAIN VISUAL Describe: - a bold, memorable central scene or object - strong focal hierarchy and visual clarity - balance between fun and professional tone LAYOUT Define: - clear relationship between visual idea, headline, and CTA - readable composition with confident whitespace STYLE DIRECTION Specify: - modern, high-end, creatively daring {{ $json.output.business_type }} aesthetic - controlled color usage with intentional contrast - polished lighting, depth, or graphic precision - campaign-level visual quality suitable for paid media -------------------------------- OUTPUT FORMAT (STRICT JSON) { "headline1": "", "headline2": "", "subheadline1": "", "subheadline2": "", "cta_text1": "", "cta_text2": "", "main_visual1": "", "layout_description1": "", "style_direction1": "" } - 7Render the final ad image · Picture · google/gemini-3-pro-image-preview (Nano Banana Pro) · OpenRouterSends the ad concept text plus the logo and product photo (as base64 image_url content blocks) to a Gemini image model, with explicit anti-AI-look and text-first composition rules baked into the prompt.
Prompt
Create a premium {{ $('Build Brand Insights').item.json.output.business_type }} value-proposition advertisement based strictly on the structured concept and text below. AD CONCEPT SOURCE: Main visual: {{ $json.output.main_visual1 }} Layout: {{ $json.output.layout_description1 }} Style direction: {{ $json.output.style_direction1 }} DISPLAYED TEXT CONTENT (use these exact words, but do not print labels like "headline" or "CTA"): Primary headline text: {{ $json.output.headline1 }} Supporting subheadline text: {{ $json.output.subheadline1 }} Call-to-action text: {{ $json.output.cta_text1 }} TEXT-FIRST COMPOSITION: - the layout must be designed around the headline as the primary anchor - headline must be dominant and clearly readable at first glance - subheadline must be smaller and support the headline - CTA must be a distinct button or label - maintain generous whitespace and clean spacing - never hide, crop, distort, or omit any text ASSET USAGE: - always include the provided asset and integrate it according to the concept - if the asset is a UI/dashboard, it must be clearly visible, readable, and occupy the primary focal area - do not blur, heavily crop, tilt, or obscure the UI ANTI-AI LOOK STYLE RULES: - avoid glossy 3D, plastic materials, neon glow, sci-fi lighting, excessive reflections, or overdone gradients - prefer flat or softly shaded illustration, editorial product-marketing visuals - restrained lighting, calm contrast, minimal visual noise Keep the background bright, clean, and uncluttered.
Free online · n8n — Generate AI ads from website + images with Claude and NanoBanana- 1
Product creative image via a form + Defapi Nano Banana APIRunsOne product advertising/creative scene image, generated from a text prompt plus a product image URL submitted through a simple form.
Nano Banana (google/nano-banana)product creative scene- 1Collect prompt, image and API key · By hand · n8n form triggerA form asks the user for a text prompt describing the desired creative scene, a public product image URL, and their Defapi API key.
- 2Submit the generation request · Picture · google/nano-banana · Defapi.org (api.defapi.org/api/image/gen)POSTs the prompt, model name and product image URL to Defapi's image generation endpoint, returning a task_id.
Prompt
={{$json.prompt}} - 3Poll for completion · Check · Defapi.org (api.defapi.org/api/task/query)Waits 10 seconds, then queries the task status by task_id; repeats until status equals 'success'.
- 4Return the image URL · edit · n8nFormats the completed result into a single image_url field for display/download.
Missing: No worked example creative prompt is given beyond the generic form placeholder; the source's 'Customization Tips' note only says to include scene, lighting, style and product placement in the prompt without giving a template.
Free online · n8n — Generate product creative images with Nano-Banana via Defapi- 1
Static ad concept + image (money-gated)RunsA single on-brand static ad image (image-to-image with a real product seed, or text-to-image), with an operator-editable concept/copy/prompt before spend.
gpt-image-2gpt-imagerecraftv3soul-v2 (Higgsfield, unconfigured)static ad- 1Operator brief submitted · WordsOperator types an art-direction brief in the Static composer.
- 2Concept agent builds prompt + copy · Words · Claude (CLI)Claude CLI concept agent runs buildConceptPrompt, which loads the brand strategy pack, house style, global rules and the selected model's guide, then returns strict JSON (concept_type, angle, copy, concept, prompt).
Prompt
You are the Static (image creative) agent for the brand ${brand}. The operator's art-direction brief: "${instruction}" ${strategy} ${houseStyle} ${rules} CRAFT BIBLE (read BEFORE concepting — it is the method): Read ${ANDROMEDA_REF} and follow its 5 steps in order: 1. STRATEGY FIRST — fix the persona, the ONE belief to install, the awareness stage, and the emotional driver from the strategy pack. Honour the operator's brief above all. 2. COPY FIRST — write the EXACT on-image copy BEFORE the visual, in the customer's own voice (voice-of-customer). Specific beats vague. The copy serves the belief + awareness stage. 3. PICK ONE concept TYPE from the menu that best executes that strategic job (use the awareness-stage table; if the operator named a type, use it). 4. WRITE THE PROMPT with the 4-part structure for the selected model. 5. CONGRUENCE CHECK — the visual, headline and copy must all sit in the SAME awareness stage. If they don't, rework the off-stage element before returning. TARGET MODEL: ${model || "the configured image engine"}. ${guide} ${hasSeed ? "A real ${brand} product photo is supplied as an image-to-image SEED — preserve the exact product (shape, material, lid colour, label text, typography, brand mark, proportions) per the model guide's i2i rules; only the scene around it changes." : "No product seed — this is a text-to-image generation."} Return STRICT JSON only, no prose, no code fences, shaped exactly: {"concept_type": "<the Andromeda concept type you chose, e.g. UGC Simulation>", "angle": "<one line: awareness stage · persona · the ONE belief this installs>", "copy": "<the exact on-image copy — headline, plus subhead/CTA/badge only if the concept needs it; empty string only if deliberately product-only>", "concept": "<2-3 sentence plain-English description of the image>", "prompt": "<the exact image prompt, written for ${model || "the selected model"} per its guide>"} The "prompt" MUST follow the selected model's guide, obey the brand global rules (props, palette, product facts), render the on-image copy + its placement (unless the concept is deliberately product-only), include "No watermarks", and end with "Aspect 16:9, high resolution." Output ONLY the JSON object. - 3Money gate · CheckOperator approves spend before the real generation call fires.
- 4Generate · Picture · gpt-image / gpt-image-2 / recraftv3 / soul-v2 · Codex CLI / muAPI / Recraft / HiggsfieldresolveProvider tries providers in order: Codex CLI gpt-image (free), muAPI gpt-image-2 (i2i or t2i), Recraft V3, Higgsfield Soul V2 (always unconfigured by design).
Ours ·- 1
Static/social-post image (t2i and i2i, model-picker)PartlyA static social-post image, operator-editable prompt seeded by an externally-generated draft.image_prompt.
gpt-image-2nano-banananano-banana-pronano-banana-2- 1Pick model + edit prompt · Picture · gpt-image-2 / nano-banana / nano-banana-pro / nano-banana-2Model picker, then an editable textarea seeded by draft.image_prompt.
- 2Generate · PicturesendStudioMessage() → actions.ts insert-only row; no direct provider call in this repo.
Missing: No direct provider call in this repo; types.ts comment claims gpt-image-2 runs 'via muAPI' but that call code is not in this repo.
Ours ·- 1
Static-ad structural guides (Recraft v3 / gpt-image-2 / Nano Banana)PartlyA static ad image, structured by a 4-part prompt (format+brand / scene / overlay text / style) per model.
Recraft v3gpt-image-2Nano Banana- 1Fill 4-part structure · Picture · Recraft v3 / gpt-image-2 / Nano Banana · MuAPIEach model file gives a 4-part prompt structure; gpt-image-2 additionally has a literal PRESERVE-clause i2i pattern for keeping a real product identical while changing the scene.
Prompt
Place this exact product in the following scene...shape, material, lid colour, label text, layout, typography, brand mark and proportions must remain identical... Aspect 16:9, high resolution.
Missing: Structure given, no filled worked example; binds to skills/muapi/SKILL.md for execution.
Ours ·- 1
Scroll-stop prompt trio (product photography/video templates)RunsThree model-agnostic static/video templates: an Assembled Shot (studio macro), a Deconstructed/Exploded Shot (per-component breakdown), and a Video Transition (cascading deconstruction).
assembled shotexploded shotvideo transition- 1Assembled Shot · PicturePhase One IQ4/120mm macro/f8 studio-spec literal template.
- 2Deconstructed/Exploded Shot · PicturePer-object component lists given verbatim for laptop/phone/shoe/food.
- 3Video Transition · VideoStart→end frame, 2-3s cascading deconstruction.
Ours ·- 1
Fantastic Posters bulk client poster/brochure generatorPartlyClient-ready posters/brochures (vertical print, landscape, square social) generated from a written brief via GPT Image 2, editable layer-by-layer in Canva Magic Layers, and batch-adapted across a folder of many photos for a consistent design (e.g. real-estate listings).
GPT Image 2vertical A2 print posterlandscape postersquare social postbulk real-estate listing poster (batch of 10+)- 1Install Fantastic Posters skill · By handInstall the robonuggets/fantastic-posters GitHub skill, which bundles 30+ starter poster styles and prompt logic for GPT Image 2 via fal.ai.
Prompt
I'm giving you a skill called Fantastic Posters. Get the files with: git clone https://github.com/robonuggets/fantastic-posters Follow the README to set it up. Recommend how it would best apply to our setup in plain language, and ask me a few questions to clarify my intent before you install it.
- 2Propose styles from brief · researchPaste a client brief and ask Claude to propose 3 poster styles: two picked from the skill's template library and one novel style not yet in the skill.
Prompt
this is a brief from the city of Melbourne. Can you propose to me three different styles for their posters? Use two from the database from the quality poster skill, and maybe suggest one that is out of left field that is not yet in the design templates that are available in the skill that you just installed, but which you think would be a good fit for this brief.
- 3Generate low-res draft posters · Picture · GPT Image 2 · fal.aiGenerate low-resolution drafts of the chosen styles via GPT Image 2, respecting the client's requested aspect ratio and incorporating the client's logo.
- 4Parallel sub-agent drafts for other briefs · Picture · GPT Image 2 · fal.aiFork the session and spin up two sub-agents to each draft two poster designs (one from the skill's styles, one novel) for two other client briefs concurrently, using their respective logos.
Prompt
I have two more briefs for poster designs here. Can you spin up two sub agents that will create a draft poster for each of these two? I want two designs each where one design is selected from our styles in the quality poster skill, and the second design is a completely new take based this end-to-end so that by the end of it, I want to see those high-res images that I can review.
- 5Upscale + multi-aspect regenerate · Upscale · GPT Image 2 · fal.aiRegenerate a selected poster at the highest possible resolution and in additional aspect ratios (A2 print, landscape, square); Claude also suggested ESRGAN or Topaz (also on fal.ai) as separate upscaler models to hit GPT Image 2's per-side pixel ceiling.
Prompt
I think we can regenerate this, but in the highest resolution possible, please, and also in different aspect ratios. So, I think we need one that is in landscape and one that is also square for a social media post. Can you summarize the different resolutions for me that you're about to create before generating them?
- 6Layer-edit in Canva Magic Layers · edit · CanvaUpload the chosen poster to canva.com/magiclayers, which auto-separates it into editable layers (text, price, photo elements) for manual tweaks, then export as image or PDF.
- 7Batch-adapt design across a photo folder · Picture · GPT Image 2 · fal.aiFeed Claude a reference design plus a folder of photos and ask it to generate a version of the design for each photo via GPT Image 2, keeping font, logo and color palette consistent, after Claude summarizes its plan.
Prompt
looking at the reference design, that turned out really well. Can you feed that as a reference to GPT image two and just tell GPT image two to create versions of it for each of the photos here in our photos folder. So, I want the same font, I want the same compass logo, and I also want the same green color palette. Can you give me an outline of the steps that you're about to take so that I can inspect it?
Missing: Needs the Fantastic Posters skill installed from GitHub, a fal.ai account for GPT Image 2 (and optionally ESRGAN/Topaz), and a Canva account with Magic Layers access for layer-editing.
RoboNuggets · n50 - Fantastic Posters with GPT-image-2- 1
Higgsfield Supercomputer batch product ad imagesPartlyA batch of roughly 10 static product ad images in varied aspect ratios, generated from just a product link, with an option to animate the best one into a short video.
static product ad image setanimated ad video (optional follow-up)- 1Give the product link · research · HiggsfieldSend Higgsfield Supercomputer a single simple command with a link to the product page, no other context supplied.
Prompt
Make 10 image ads for this product.
- 2Autonomous research and skill load · research · HiggsfieldSupercomputer analyses the product page, loads its internal product-image creation skill(s), reads the web page, and loads an 'ad creative pack' reference.
- 3Generate the ad batch · Picture · HiggsfieldThe underlying brain model (Opus 4.6 in the demo) reasons through hooks and scenes, and an image model (not named in the source) batch-generates roughly 10 product ad images across varied aspect ratios, viewable in a gallery.
- 4(Optional) Animate the winner · Video · Kling 3.0 (attempted, failed) then Seedance 2.0 (succeeded) · HiggsfieldAdd a chosen still to the composer and ask to animate it with a named video model; Supercomputer shows a checkpoint (model, aspect ratio, resolution, duration, sound, prompt-enhance toggle, credit cost) before generating, and retries with a fallback model on failure.
Prompt
Animate it with Kling 3.0, please. If it fails, try again up to maybe three times up until we get an output. ... Use Seed 2.0 instead.
Missing: The specific image model used for the initial 10-image batch is never named in the source; Supercomputer gave no failure-reason diagnostics when Kling 3.0's generation failed.
RoboNuggets · b17 - Higgsfield Supercomputer- 1
Blog/news summary to social posts with an AI-generated Instagram imageRunsA short caption plus one AI-generated photorealistic image, posted to Instagram, summarizing a linked news article or blog post; matching plain-text posts (no image) are also created for Twitter/LinkedIn/Facebook.
DALL-E (via ChatGPT image tool)single social image post- 1Source the article · research · BitlyA link (news article in r3-4/r5, or the user's own blog post in r8) is put in a Google Sheet row; Bitly shortens the link.
- 2Summarize · research · Perplexity (unspecified) · Perplexity AIPerplexity AI fetches and summarizes the linked article.
- 3Generate Instagram image · Picture · DALL-E (via ChatGPT) · OpenAIA ChatGPT agent writes a photorealistic image prompt from the article summary and passes it to DALL-E to generate one accompanying image for the Instagram post.
- 4Write Instagram caption · Words · ChatGPT · OpenAIA ChatGPT agent writes the Instagram caption to accompany the generated image.
- 5Publish · publish · Make.comMake.com posts the image + caption to Instagram. (Parallel text-only posts for Twitter/LinkedIn/Facebook are also generated in the same scenario but carry no image.)
Missing: Verbatim DALL-E and Perplexity prompts not given in any of the three source lessons, only described in the walkthrough narration.
RoboNuggets · R3-4 Social Media Channels AI Agent / R5 AI Twitter System / R8 Repurpose Your Blog for Socials via AI Agents- 1
Batch image ad splitterRunsa batch of unique static image ads from one product/character photo — each variant bakes in a different ad-copy/product/character combination and watermark placement
google/gemini-2.5-flash-image-preview:free (Nanobanana)batch static ad- 1Receive reference photo · By hand · TelegramTelegram photo + caption trigger.
- 2Get image path · By hand · Telegram Bot APIFetches the Telegram file path.
Prompt
https://api.telegram.org/bot{{ $('Set Bot ID').item.json['bot id'] }}/getFile?file_id={{ $('Telegram Trigger').first(0,0).json.message.photo[2].file_id }} - 3Describe image · Picture · OpenAI (not named) · OpenAIVision call distinguishes product vs character vs both, producing a YAML-style field list (brand_name/color_scheme/font_style/visual_description for products; character_name/color_scheme/outfit_style/visual_description for characters).
- 4Get production rows · By hand · Google SheetsReads Google Sheet rows with status = create, pre-populated via a companion external 'Ideator GPT' with product/character/ad_copy/visual_guide/text_watermark/colours.
- 5Build one image ad prompt · Words · not named · n8nAI Agent ('Image Ad Prompt Generator Agent', A-G-E-N-T framework) builds one image ad prompt per row, batched 1 per 3000ms, with smart defaults (watermark bottom-left by default, ad_copy max 7 words, omit null fields).
Prompt
Your task: Create 1 image prompt as guided by your system guidelines. *** The user's special request: {{ $('Telegram Trigger').first().json.message.caption }} *** Description of the reference image: {{ $('Describe Img').first().json.content }} ELEMENTS FOR THIS IMAGE: product: {{ $json.product }} character: {{ $json.character }} ad copy: {{ $json.ad_copy }} visual_guide: {{ $json.visual_guide }} text_watermark: {{ $json.text_watermark }} Primary color: {{ $json["1_color"] }} Secondary color: {{ $json["2_color"] }} Tertiary color: {{ $json["3_color"] }} *** Use the Think tool to double check your output for this run - 6Generate ad image · Picture · google/gemini-2.5-flash-image-preview:free (Nanobanana) · OpenRouter (POST https://openrouter.ai/api/v1/chat/completions)Calls the free Nanobanana model with the reference image and generated image_prompt.
- 7Extract and save image · edit · n8nBase64 extraction and conversion to a file.
- 8Deliver / log · publish · Box.com + Google SheetsOn success, uploads to Box and marks the row done; on error (e.g. content-policy stop), logs the error reason in the sheet.
Missing: the upstream 'Ideator GPT' that pre-populates the sheet rows is external and not part of this n8n workflow. Referenced n8n template JSON not present locally.
RoboNuggets · R39 | 100 Ads in 1 Minute - Split AI System- 1
Brand-grounded ChatGPT image agent (via n8n MCP)Partlyon-brand marketing images (e.g. a holiday sale ad, close-up product shot) generated from a brand style guide and product reference images, produced conversationally inside ChatGPT and optionally auto-posted to Instagram with a suggested caption
Nano Banana ProGPT (ChatGPT, version not stated)brand-book mood boardon-brand product ad- 1Build brand book (one-off asset) · Picture · Nano Banana Pro · Gemini (Thinking mode) or Kie.aiGenerates a brand mood board (logo, style grid, fonts, colour palette, tone of voice) from a single inspiration image.
Prompt
Create a clean, modern brand book mood board using the provided image as the core inspiration. Place the brand logo at the top center in a style that matches the mood of the image. On the left side, add a section titled "Style Aesthetic:" with the subtitle "What we aim our style to look like". Present the provided image as a neat grid or collage that reflects its natural style and tone. On the right side, create structured brand book sections. Let all visual choices be inferred from the uploaded image: Font: Show two font examples labeled "Headlines" and "Subheadlines". Choose font styles that match the character, mood, and feeling of the provided image. Main color palette: Generate a set of color swatches directly based on the dominant and supporting colors found in the provided image. Label each swatch with simple names that match the tones. Tone of voice: Create a short list of brand voice traits that reflect the emotions, attitude, and personality suggested by the provided image. Keep the overall layout minimal, balanced, and clean with gentle spacing and simple black text. The final result should look like a polished brand book page that adapts itself entirely to the style of the uploaded image.
- 2Connect ChatGPT to n8n via MCP · By hand · n8n + OpenAIEnables n8n's instance-level MCP server and pastes the server URL into ChatGPT's custom connector settings (developer mode), authorized via SSO.
- 3Enable specific n8n workflows for MCP · By hand · n8nSets the target image-gen workflow active and exposes it to MCP; trigger must be form/webhook/chat.
- 4Set ChatGPT project custom instructions · Words · ChatGPT (project custom instructions) · OpenAIGives ChatGPT persistent brand context, asset URLs, and rules for prompting/confirmation/preview.
Prompt
SYSTEM GUIDELINES: - Whenever you're asked to create images - ALWAYS use the n8n tool by executing the workflow: "MCP - Nanobanana Pro (by RoboNuggets)". - When you send data to any n8n workflow, have the inputs be of the type: 'form', and input a JSON body that includes the values that the workflow expects - Unless the user specifies otherwise, always confirm with the user the prompt that you will send to create the image - When creating a prompt, ALWAYS append this at the end exactly, word for word, do not change it: "Refer to the brand style guide provided for the intended mood, style, font, and brand logo - but DO NOT use any of the text or human images in the brand book as it is just intended to guide you on what the brand's style should look like. Do not blend or copy ANY image from the brand book into the final image APART from the brand logo." - Whenever you create an image, ALWAYS return the final image URL as an inline image preview in the chat. Never skip this. *** You have access to the following assets. Use these URLs as input for creating images using n8n. Aje Brand Book: [URL] Manon Mini: [URL] Splendour Dress: [URL] *** When crafting captions, consider the description of the brand here: Aje is a contemporary Australian fashion brand that is known for a signature blend of "raw beauty, tough femininity, and effortless cool," combining luxury and artisanal craftsmanship with modern designs and sculptural silhouettes. The brand is recognized for its handcrafted heritage, featuring hand-painted prints, natural fabrications, and artisanal techniques, while also balancing statement pieces with everyday essentials
- 5User request to agent-drafted image prompt · Words · ChatGPT (agent) · OpenAIUser asks for an image; ChatGPT drafts the exact prompt referencing the brand book and product, confirming with the user before running.
- 6Generate image via n8n MCP tool call · Picture · Nano Banana Pro · Wavespeed (14c) or Kie.ai (9c)ChatGPT invokes the n8n workflow (form input), which runs Nano Banana Pro with product and brand book reference images.
- 7Publish to Instagram · publish · n8n workflow → Instagram (likely Blotato, per prior lessons)On approval, calls a second MCP-exposed n8n workflow to post the image and an AI-suggested caption.
Missing: exact per-request image prompt text is agent-generated live, not fixed; the Instagram-publish node's underlying tool isn't named explicitly in this file.
RoboNuggets · R49 | Agentic Creative Systems (beta)- 1
Pictures28
Claymation frame prompts (ChatGPT method)Partlyscene-by-scene claymation-style still image prompts for a script, generated via a custom ChatGPT GPT, used to feed a video-animation step
Nano Banana 2claymation- 1Turn script into a storyboard · Script · custom ChatGPT GPT (not named) · ChatGPTDrag the script into the custom GPT, click 'turn this script into claymation storyboard' (or use the alternate idea-generation button with no script).
- 2Generate the scene images · Picture · Nano Banana 2 (not Pro — judged better/newer for this) · Max Fusion / Fal / HiggsfieldCopy each scene prompt into an image generator, text-to-image, 9:16 aspect ratio; batch through scenes.
- 3Fix consistency issues · Picture · Nano Banana 2 · Max FusionDownload a generated image, switch to image editing mode (not generation), drag it in and add an edit instruction to refine.
Prompt
remove the scent and brain
- 4Reuse a character across scenes · Picture · Nano Banana 2 · Max FusionSave the character image, use image editing mode with the character dragged in as reference, and prompt the new scene action — otherwise the character changes each generation.
Prompt
create an image of the woman standing up in her living room
Missing: the custom GPT's own prompt is referenced as 'available below' but not included in the transcript
AI Ads (Ad Creators Lab) · Prompts + Frames - ChatGPT (Lesson 18)- 1
Claymation frame prompts (Poppy method)Partlyscene-by-scene claymation-style image prompts generated at scale inside Poppy's project system (style reference group + house-style knowledge base + product info), for multiple products/avatars
Nano Banana 2Claude 5 Sonnetclaymation- 1Set up project references · By hand · PoppyInside a Poppy project, set up a claymation style reference group (images), a project prompt (house style/rules knowledge base), a transition-tip note, and product info (research doc, site screenshot, product link).
- 2Extract scene breakdown · Script · PoppyLink the style group, project prompt, and product info into a working prompt that extracts a scene breakdown (script line, visual intention, emotion to evoke) per scene.
Prompt
You are a direct response strategist and creative director. I provided you with a script as well as resources for my product... extract the following information
- 3Generate an image prompt per scene · Script · Claude 5 Sonnet (optional) · PoppyIn Poppy's 'Easy Prompt' chat, ask for an image prompt for a given scene number; the AI reads the project knowledge base and outputs a full claymation prompt matching house style. Can switch to Claude 5 Sonnet to batch-generate prompts for all scenes at once.
Prompt
Create an image prompt for the following scene, scene one
- 4Generate the image · Picture · Nano Banana 2 · Poppy / Fal / HiggsfieldPaste the generated prompt into Nano Banana 2 inside Poppy (Nano Banana Pro judged worse for this than Nano Banana 2); can also route to Fal or Higgsfield.
- 5Vary and keep characters consistent · Picture · Nano Banana 2 · PoppyUse Poppy's 'AI Edit' mode with an instruction for variation. For character consistency, connect an existing character image as reference and prepend 'Keeping the character the same, execute this prompt: [original prompt]'.
Prompt
zoom out so we see the full body of the person
- 6Storyboard the sequence · By hand · PoppyOrganize/reorder the generated scene images visually inside Poppy against the script before moving to video generation.
Missing: the 'Prompt 1' strategist/creative-director instruction and per-scene image prompts are paraphrased in the transcript, not given as full literal text
AI Ads (Ad Creators Lab) · Prompts + Frames - Poppy Method (Lesson 19)- 1
Realistic reference-JSON images with Nano Banana ProRunsA realistic product/UGC/influencer still photo that keeps a real reference photo's lighting, coloring and composition while swapping in a new subject or product.
Nano Banana ProGemini 3 Proproduct photoAI influencer photopodcast-set still- 1Find a reference photo · research · PinterestSearch Pinterest for an image matching the wanted aesthetic.
- 2Extract a detailed JSON description · Picture · Gemini 3 Pro (high thinking) or ChatGPT (thinking mode) · Google AI Studio / ChatGPTConvert the reference photo into a very detailed JSON description covering age, ethnicity, hair, styling etc.
Prompt
create me a very detailed JSON that can recreate this image exactly as it is. Make sure it is very detailed.
- 3Generate the new image · Picture · Nano Banana Pro · Google AI Studio ("used like an API") or Higgsfield.aiFeed the JSON back in plus the new subject/product to insert, keeping the original's lighting/coloring/composition.
Prompt
[paste the JSON], generate me this photo, but put her in the hoodie
- 4Iterate · Picture · Nano Banana Pro · Google AI Studio / Higgsfield.aiFeed the same JSON back with a follow-up description of the wanted change; the tool returns an adjusted JSON and a new image built on the same base style.
Prompt
hey, I don't like this, keep this same podcast style, then make it a man who hits the gym and make the podcast studio more appropriate for men over 40 who hit the gym.
The AI Content System · How To Make Realistic Images With NBP- 1
5-slot static image, then motion handoffPartlyone plain-English-briefed still image (product shot, hero, brand asset) up to 2560x1440, batched up to 8 on-model variants, optionally then animated as a short motion clip in a second tool
GPT Images 2.0product stillbrand herobatch variant setlocked-keyframe-for-motion- 1Write the 5-slot brief · Words · GPT Images 2.0 · OpenAIStructure the prompt into five slots: scene, subject, important details, use case, constraints. Plain specific English, not JSON.
- 2Apply anti-slop rules · By hand · manualEdit the draft prompt: visual facts not mood words (kill 'cinematic, 8k, masterpiece'), name objects not moods, quote in-image text literally and spell hard words letter by letter, separate change/preserve lists on any edit turn, one revision per turn.
- 3Generate · Picture · GPT Images 2.0 (Instant or Thinking mode) · OpenAIRender at quality high, 2560x1440, aspect as needed (multiples of 16, max edge 3840). Turn Thinking Mode on when consistency or counts matter (it plans composition, web-grounds real-world facts, holds 8-frame batch consistency, self-verifies against the brief); use Instant for practice/cheap iteration.
- 4Hand off to motion · Video · not stated (Higgsfield's routed video model) · HiggsfieldGive the locked still to Higgsfield for camera moves, parallax or character action; GPT Images 2.0 owns identity, Higgsfield owns movement.
Missing: no worked example prompt/output is quoted in the source; the 5-slot template and settings are exact but untested against a real brief
AIOps · GPT Images 2.0: 5-Slot Prompting, Thinking Mode, Codex Integration, and the Higgsfield Motion Handoff- 1
Reference-DNA style extraction and transferPartlya new still image that carries the extracted style (or full clone) of a reference image, via a JSON 'DNA profile' pasted into the generator with a new scene brief
Nano Banana 2 (course names it gemini-3.1-flash-image)Nano Banana ProGemini 3.1 Pro (vision, extraction only)style-transferred product/scene imagecloned-and-modified image- 1Build a swipe file · By hand · manualSave liked reference images from cosmos.so, Pinterest, Instagram, meigen.ai.
- 2Full Profile Extraction · Check · Gemini 3.1 Pro (vision) · Google AI StudioRun the 'Full Profile Extraction' prompt against one reference image to get forensic JSON: emotion, composition, hex colours, lighting, camera/lens estimate, artifacts, genre. Used to clone the image with controlled modifications.
- 3Style-Only Extraction (alternative to step 2) · Check · Gemini 3.1 Pro (vision) · Google AI StudioRun the 'Style-Only Extraction' prompt to strip content and keep only the recipe: lighting blueprint, colour system, composition framework, camera language, post-processing signature, plus a 'style_dna' paragraph usable as a standalone brief.
- 4Generate from the profile · Picture · Nano Banana 2 (volume/graphics) or Nano Banana Pro (realism/UGC/precision ad creative) · Google AI StudioPaste the JSON plus a new scene description into the generator with Thinking Level High and Grounding with Google Search on.
Missing: the two extraction prompts' literal text is not quoted in the source, only their required output structure
AIOps · Nano Banana 2: Reference-Based Image Generation with JSON Style Extraction- 1
Mood-board to product-on-model shotPartlya product-on-model photo (hero shot plus a matching second angle), generated from a Pinterest-style mood board and a real product photo
Gemini 2.5 FlashMidjourney V7Topaz (upscale)Nano Banana Pro (named, exact compositing model not stated)product-on-model herosecond-angle variant- 1Feed the mood board in · By hand · manual upload to Weavy canvasUpload Pinterest reference images (in the walkthrough: ski images) to the Weavy canvas as a node.
- 2Describe the board · Words · Gemini 2.5 Flash · WeavyPrompt a vision model to describe the mood board's colours, textures, lighting, mood and vibe.
- 3Render a hero candidate · Picture · Midjourney V7 · WeavyTurn the text brief into an image prompt and render candidates; pick one.
- 4Realism pass · edit · not stated · WeavyRun a realism-correction node on the chosen hero.
- 5Upscale · Upscale · Topaz · Topaz (Weavy node)Upscale the realism-passed image.
- 6Composite the real product · edit · not stated (course lists Nano Banana Pro among tools) · WeavyComposite the actual product photo onto the model, prompted 'put this on the model' (walkthrough used a real North Face jacket photo). Use the real product photo, never a text description of it.
Prompt
put this on the model
- 7Second angle · Picture · same as step 6 · WeavyGenerate a matching alternate angle of the same composited shot with a 'different angle' prompt.
Prompt
different angle
Missing: the text-guide half of this pair is almost pure video-wrapper with no written steps; the walkthrough is machine YouTube captions (names may be misspelt) and exact prompt wording beyond the two short phrases quoted is not preserved
AIOps · Advanced AI Image Generation with Nodes (Weavy / Figma Weave)- 1
Fill-in-the-blank brand image templatesPartlya text prompt (not an image itself) for a product shot, thumbnail or infographic, built from a pre-boiled template with brand colours and exact on-image text filled in; pasted into an image tool to render
GPT-Image-2 (renders the prompt the skill writes)product shotthumbnail4:5 infographic- 1Install the skill · By hand · manual (terminal install command)Install the free awesome-gpt-image-2 skill, a folder of pre-boiled prompt templates, via a terminal command (Node.js 18+ required).
- 2Ask for a fill-in prompt · Words · the skill (picks/fills a template)Ask the agent for a product shot with your hex colours, exact on-image text, and a size ratio (e.g. 16:9); the skill picks a template and fills it. Drawing happens elsewhere.
- 3Lock as a house template · Words · manualWhen an image comes back right, ask the agent to save that working prompt as a house template at brand/image-templates.md.
- 4Repeat per format · Words · the skillRepeat for other formats (thumbnail, 4:5 infographic).
Missing: the actual template library is an external Chinese-language dependency not captured locally; half the linked project is the author's paid shop
AIOps · GPT-Image-2 Prompt Engine: Reusable Image Templates the Agent Fills In- 1
50 UGC product images from one photoRuns50 unique still UGC-style marketing images of a single product used in different realistic situations, generated automatically from one product photo with a white background.
Nano Banana (Gemini 2.5 Flash Image, via fal-ai/gemini-25-flash-image/edit)Mistral Large (mistral-large-latest)UGC product photoe-commerce lifestyle shot- 1Watch for a new product photo · By hand · Google DriveGoogle Drive Trigger polls a specific 'UGC' folder every minute; a new file (product image, white background) starts the run.
- 2Generate 50 UGC prompts · Words · mistral-large-latest · Mistral CloudAn LLM chain writes 50 diverse, realistic image-to-image prompts describing the product in different settings, people, times of day and moods, returned as strict JSON via a structured output parser.
Prompt
SYSTEM: You are an expert content creator specializing in UGC (User-Generated Content) style marketing images. HUMAN: The product the Promote is a {{ $('Google Drive Trigger').item.json.name.match(/^[^.]+/)[0] }} Your task is to generate 50 creative, diverse, and realistic situations prompts for an image-to-image generation model. Each prompt should describe a different situation or scenario where the product is being used. Requirements for each prompt: 1. Specify the environment or setting (e.g., home, office, kitchen). 2. Include a human or realistic context interacting with the product (e.g., holding it, using it, displaying it). 3. Highlight at least one functionality or feature of the product in the scene. 4. Keep the prompts concise (1-2 sentences), vivid, and visually descriptive. 5. Maintain diversity: do not repeat the same type of scene. Include different times of day, moods, activities, or locations. Output format (JSON array): { "ugc_prompts": [ {"prompt": "First prompt here..."}, {"prompt": "Second prompt here..."}, ... {"prompt": "Fiftieth prompt here..."} ] } Example input: Product: "Smart Water Bottle" Example output: { "ugc_prompts": [ {"prompt": "A young woman jogging in a park, drinking from the Smart Water Bottle, enjoying cold water on a sunny morning."}, {"prompt": "A man sitting at his office desk, checking water intake on his Smart Water Bottle synced with his phone."}, {"prompt": "A family hiking on a trail, each carrying a Smart Water Bottle to stay hydrated."},... ] } DON'T RETURN ANYTHING ELSE THAN THE JSON! - 3Split and loop the 50 prompts · edit · n8nSplit Out breaks the ugc_prompts array into 50 items; Loop Over Items (splitInBatches) processes them one at a time.
- 4Generate the UGC image · Picture · Nano Banana (Gemini 2.5 Flash Image) · Fal.ai (fal.run/fal-ai/gemini-25-flash-image/edit)Each prompt is sent with the original product photo (as a public Google Drive URL) to Nano Banana's edit endpoint, producing one new image per prompt.
Prompt
={{ $json.prompt }} - 5Retrieve and save the image · edit · Google DriveDownloads the generated image from the returned URL and uploads it back into the same Google Drive folder, named from the product name plus a prompt number.
Free online · n8n — Create high-quality UGC images with Nano Banana (cheaper & faster)- 1
One base image expanded into a full ad campaignRunsA set of on-brand campaign assets (ad, blog graphic, thumbnail, landing hero, email banner, social post, poster, video source frame) all derived from one strong base image and a small set of controlled variants.
Nano Banana Pro (Gemini 3 Pro Image)product heropostersocial slideUI mockup- 1Create a base asset · Picture · Nano Banana Pro (Gemini 3 Pro Image)Generate one strong image using a structured prompt formula covering format/purpose, subject, composition, environment, lighting, style, exact text (if any), reference-image roles, factual details to include, restrictions to avoid, and the intended output use.
Prompt
Create a [format/aspect ratio] image for [purpose]. Subject: [main subject with specific details]. Composition: [framing, layout, crop, negative space]. Environment: [scene or background]. Lighting: [direction, mood, color]. Style: [photographic, editorial, poster, infographic, UI, illustration, etc.]. Text if needed: [exact words, placement, typography]. Reference roles if using images: [what each reference controls]. Details to include: [materials, textures, factual details, brand cues]. Restrictions: [what to avoid]. Output intent: [ad, product page, poster, source frame, infographic, social post]. ---EXAMPLE--- Create a square e-commerce product hero image for a premium coffee brand. Subject: matte black coffee bag with a blank copper label area, standing upright. Composition: centered product hero frame with clean negative space around the bag. Environment: dark walnut table, blurred coffee shop background. Lighting: warm side light from the left, subtle rim highlight on the bag edge. Style: realistic premium product photography. Text: no generated label text, keep the label area blank. Details: matte packaging texture, soft shadow, clean reflection. Restrictions: no extra bags, no fake brand text, no distorted packaging, no watermark. Output intent: product page image and source frame for later image-to-video.
- 2Create controlled variants · Picture · Nano Banana Pro (Gemini 3 Pro Image)Re-run generation changing exactly one axis at a time (aspect ratio, background, crop, colour palette, lighting, text placement, platform format, CTA space, or product environment) so each variant is a clean A/B test of one variable, using an explicit reference-image role system when source photos are involved.
Prompt
Use Image 1 as the exact product reference. Use Image 2 as the background environment reference. Use Image 3 as the lighting and color grade reference. Create a vertical 9:16 product campaign image. Preserve the product shape, packaging, color and label placement from Image 1. Place the product naturally into the environment from Image 2. Apply the warm golden lighting mood from Image 3. Keep the product centered in the lower half with clean space above for headline text. Do not change the product label, do not invent extra text, do not add extra products.
- 3Polish the winner · edit · Cliprise Pro Image Editor / AI Background Remover / Universal UpscalerOnce a winning variant is picked, refine it with a dedicated image editor, background remover and upscaler rather than re-prompting from scratch.
- 4Repurpose across formats · editTakes the polished asset and reformats it (crop/resize/re-layout, not regenerate) into a paid ad, blog graphic, thumbnail, landing-page hero, email banner, social post, poster, and video source frame.
Missing: The Cliprise tool names (Pro Image Editor, Background Remover, Universal Upscaler) are the author's own product features being promoted alongside the method; the underlying models behind those specific tools are not named, only that they run on Nano Banana Pro / Gemini 3 Pro Image for generation.
Free online · GitHub — cliprise/awesome-nano-banana-pro-prompts: 'Workflow: one prompt to campaign system'- 1
Fact-grounded infographic imageRunsA single infographic-style still image (e.g. a vertical 4-step explainer) with exact, pre-supplied labels and icons, built to avoid the model inventing facts.
Nano Banana Pro (Gemini 3 Pro Image)infographicexplainer graphic- 1Supply facts and exact labels only · WordsThe operator writes out the real facts and the exact short labels to use (e.g. step names) before prompting, and the prompt explicitly instructs the model to use ONLY the provided facts - never invent any.
Prompt
Use only the provided facts below.
- 2Generate the infographic · Picture · Nano Banana Pro (Gemini 3 Pro Image)Generates a simply-structured infographic (3-5 sections, clear icons, large readable text, strong hierarchy) using the exact labels supplied, in a defined visual style.
Prompt
Create a vertical infographic about how image-to-video workflows work. Audience: beginner creators. Structure: 4 stacked steps. Text: use these exact labels: "1. Source Image", "2. Motion Prompt", "3. Generate Video", "4. Review & Export". Visuals: simple clean icons for image, motion arrow, video frame and checkmark. Style: modern dark-mode SaaS infographic with cyan accents. Composition: clear hierarchy, readable spacing, mobile-friendly layout. Restrictions: no invented steps, no extra text, no misspellings, no clutter, no fake logos.
- 3QA the output manually · CheckA human checks the generated infographic for spelling errors, invented facts, label accuracy, icon relevance and overall visual clarity before publishing.
Free online · GitHub — cliprise/awesome-nano-banana-pro-prompts: 'Workflow: infographic generation'- 1
Consistent character across many images/scenes (Midjourney + Nano Banana)Runsa reusable character (and later a shared world of multiple characters) rendered consistently across many angles, close-ups and composited scenes — an image set usable as a UGC/animation base
MidjourneyNano Banana (via Krea.ai)consistent character image set- 1Create the base character in Midjourney · Picture · Midjourney · MidjourneyGenerate the character with high chaos/exploration settings to get a strong, distinctive look, then run permutation batches to pick the best variant.
Prompt
fashion shot of a modern woman [ADD DETAILS], casual, minimalist --chaos 100 --ar 9:16 --exp 100 --sref 2216056865 4129353691 1742826451 3368914030 689668977 1639812559 2825562739 --profile de2v2uv gy97k8r gtbhewc gj5325t
- 2Generate new full-body angles · Picture · Nano Banana · Krea.aiUpload the chosen Midjourney character image into Krea.ai, select the Nano Banana model, and generate new angles with a simple prompt.
Prompt
woman, full body shot
- 3Refine detail/macro shots · Picture · Nano Banana · Krea.aiReplace the reference image with the latest output and generate focused macro variations of specific body details.
Prompt
woman, different angle, macro, hands / woman, different angle, macro, shoes / woman, different angle, macro, earrings
- 4Generate a scene element · Picture · Nano Banana · Krea.aiGenerate a reusable environment/prop asset (e.g. a chair in a room) in the same visual style, independent of the character, so it can be reused across multiple character composites.
Prompt
a chair in an empty room in the style of the image.
- 5Composite character into scene · Picture · Nano Banana · Krea.aiCombine the character reference and the scene reference with a simple composite prompt to place the character into the generated environment.
Prompt
a woman sits on a chair.
- 6Build a multi-character shared world · Picture · Midjourney + Nano Banana · Krea.aiRepeat steps 1-5 for additional characters, reusing the same environmental/scene references so all characters share a consistent visual world.
Free online · Step-by-Step Guide: Consistent Characters & Scenes with Midjourney + Nano Banana- 1
Start-frame person swap (i2i, for i2v handoff)RunsA single edited still image: the reference frame's person swapped for a new subject, everything else (framing, background, lighting) preserved — built as the first frame for a downstream image-to-video generation.
muapi-gpt-image-2UGC start frame- 1Extract reference frame · Videoffmpeg extractFrameAt pulls a still from the sourced reference clip.
- 2Build edit prompt · PicturebuildStartFramePrompt composes the edit instruction: swap only the on-camera person (from the script's angle, or an operator steer), preserve everything else.
Prompt
Edit the supplied reference image. CHANGE — only this: ${change} PRESERVE — everything else exactly as in the reference: the framing, crop, composition, camera angle and distance, the background/setting, and the lighting. One person only, talking to camera. No on-screen text, no captions, no watermark, no logos. - 3Money gate · CheckAtomic compare-and-set approval before spend.
- 4Generate · Picture · gpt-image-2 · muAPIresolveProvider explicit default muapi-gpt-image-2.
Ours ·- 1
New picture (Studio/Make chat-driven image edit)PartlyA single static image, generated or edited from an operator's chat instruction.
gpt-image (family)nano-banana (family)- 1regenerate_image verb · Picturesay() dispatches the regenerate_image verb, writing a row into social_studio_messages; the external engine (named only via a text-stripping regex in words.ts: gpt-image, nano-banana, muapi, kie.ai, gemini, claude, elevenlabs) does the actual call outside this repo.
Missing: Provider call happens outside this repo; provider names only inferable from a reply-text-stripping regex.
Ours ·- 1
Nano Banana Pro — end-to-end persona/product imageRunsA static persona or product image, built by turning a Pinterest reference into a JSON prompt via Claude, then generating in Google Flow.
Nano Banana Pro- 1Source + describe · research · ClaudePinterest reference, then Claude: 'Give me a detailed JSON prompt for this image.'
Prompt
Give me a detailed JSON prompt for this image.
- 2Refine · Words · ClaudePaste the JSON back to Claude with a product/clothing image + instruction; Claude returns updated JSON.
- 3Generate · Picture · Nano Banana Pro · Google FlowGenerate in Google Flow, Images tab, 9:16, 4 outputs, Nano Banana Pro (alternatives Keyai/Higgsfield/AI Studio); optional in-Flow edit.
Ours ·- 1
Nano Banana Pro — Action TransferRunsA static image of a locked character performing a specific action taken from a reference image, holding a product.
Nano Banana Pro- 1Generate · Picture · Nano Banana ProUpload locked character image, then reference action image (e.g. Getty stock), then product as an 'ingredient'; regenerate 2-3x if imperfect.
Prompt
make the model in the first image perform the action in the second image
Ours ·- 1
Nano Banana Pro — 20 Curated PromptsRuns20 distinct static-image style looks (street fashion, cyberpunk, Y2K, anime, etc.), copy-paste ready.
Nano Banana Prostreet fashioncyberpunkY2Kanime+16 more- 1Pick + paste · Picture · Nano Banana ProPick one of 20 full literal style prompts from the lookup table, paste into Nano Banana Pro.
Ours ·- 1
Nano Banana Pro/2.0 — Ads & Persona Formula LibraryRunsPersona and ad images across 5 categories (UGC selfie, lifestyle, studio portrait, product/brand, night/urban), with a face-lock protocol for identity consistency across a shoot.
Nano Banana ProNano Banana 2.0UGC selfielifestylestudio portraitproduct/brandnight/urban- 1Fill category template + face-lock · Picture · Nano Banana Pro / 2.05 literal face-lock phrases + matching negative prompts; plain-text and JSON templates per category; lighting-formula table; a 5-rule Persona Consistency Protocol; a vertical-aesthetic table (VERO→editorial luxury, cannabis→relaxed lifestyle).
Ours ·- 1
Ultra-Realism Character GenerationPartlyA close/medium-close static character portrait with deliberately imperfect, realistic skin.
Nano Banana ProChatGPT Image 1.5- 1Generate + upscale · Picture · Nano Banana Pro / ChatGPT Image 1.5Pinterest reference → Nano Banana Pro or ChatGPT Image 1.5, close/medium-close framing with a literal imperfection keyword fragment, then upscale to 4K.
Prompt
natural skin texture, small imperfections, realistic pores, slight blemishes
Missing: Keyword fragment literal, no full worked prompt.
Ours ·- 1
Freepik Spaces Fashion/Ecommerce Photography & VideoRunsA full ecommerce fashion shoot: a consistent AI model dressed in numbered product nodes, posed via a Pinterest pose-hack, covered from 360°, then animated between poses.
ClaudeNano Banana ProKling- 1Create model · Picture · Claude, Nano Banana Pro · FreepikLiteral Claude prompt for a hyper-realistic human being from a reference, then Nano Banana Pro generate + Freepik upscale.
Prompt
hyper-realistic human being (from a reference — full literal 'Master Character Prompt' JSON template with Hasselblad H6D/Kodak Portra 160 camera-spec detail is in the source file, not re-quoted here)
- 2Dress via node graph · Picture · Freepik SpacesNode-based dressing: product images as numbered nodes; a scene node carries the brand palette.
- 3Pose-hack · Picture · Nano Banana ProA Pinterest pose reference is converted to a stick-figure via Nano Banana, then used as a pose reference.
- 4360° coverage · Picture · Freepik Spaces45°-increment sequence nodes give full 360° coverage.
- 5Video sequences · Video · KlingStart/end-frame pairs from the 360° set fed to Kling, then After Effects timing.
Ours ·- 1
AI news infographicPartlyone Instagram-ready infographic image with rendered text and a source citation, built from a same-day researched news article.
GPT-image (OpenAI 4o image generation model)GPT-4.1infographicinstagram post- 1Research topic · research · GPT-4.1 · OpenAI (agent) / Tavily (search tool)A research AI agent uses a Tavily web-search tool to find today's latest news article on a given topic.
Prompt
Find today's latest news about [topic of your choice]
- 2Write caption · Words · GPT-4.1 · OpenAIA captions agent turns the research into an organized Instagram caption.
- 3Craft image prompt · Words · GPT-4.1 · OpenAIAn image-prompt agent combines the research, caption and a predefined infographic 'image style' text into a prompt for the image model, with guidelines on layout so the text renders correctly.
- 4Generate infographic · Picture · GPT-image (OpenAI 4o image gen) · OpenAIThe 4o image generation API renders the infographic with high text adherence, including the article's source line at the bottom.
- 5Convert binary to file · edit · n8nn8n converts OpenAI's binary image response into a file n8n can preview/use.
- 6Upload and publish · publish · imgur / BlotatoImage is uploaded to imgur to get a public URL (Blotato only accepts URLs, not files), then Blotato posts the image with the caption to Instagram.
Missing: the 'image style' infographic template text and the research/caption agents' system prompts are referenced but not printed in the harvested lesson; Tavily and Blotato credential setup shown elsewhere, not in this file.
RoboNuggets · n1 - Experimenting with ChatGPT's new image API!- 1
Image model comparison tester (Seedream 4.0 vs Nano Banana)Runsproduct/creative images across multiple use cases, optionally animated
Seedream 4.0Nano BananaUGC imageproduct try-on/dress-upstyle transfercharacter turnaround boardbrand merch mockupstudio product ad- 1Create image · generate-image · Seedream 4.0 or Nano Banana · Kie.aiHTTP POST to Kie.ai create-image endpoint per model, with model-specific body (prompt, aspect ratio, reference image where applicable)
- 2Poll for result · poll · Kie.aiWait node + GET request polling job status, switch node branches on success/in-progress/error
- 3Optional animate · generate-video · Veo3 fast · Kie.aiOptionally animate the resulting image with Veo3 fast
Missing: not a single fixed pipeline — a tester comparing two models across 6 use cases; use-case-specific prompts not fully quoted
RoboNuggets · n20 - Seedream 4.0 in n8n- 1
Digital product image generator (Nano Banana Pro, Telegram agent)Runsprint-ready digital product images: wall art/posters, print-on-demand shirt designs, coloring book pages
Nano Banana ProNano Banana (base, comparison)2K4K- 1Request via Telegram · input · TelegramSend a request (e.g. replicate this wall art style in a new niche, generate N shirt design variations, generate coloring page variations) to a Telegram-connected creative AI agent
- 2Generate image(s) · generate-image · Nano Banana Pro · Kie.aiCreate-image HTTP request node calls Nano Banana Pro with prompt, aspect ratio, resolution, output format, and image reference (fields taken from Kie.ai's sample API request)
- 3Return via Telegram · deliver · TelegramAgent sends generated image(s) back to the requester
Missing: no native transparent-background support at time of recording — squares shown for 'transparency' are baked into the image, not a real alpha channel; a separate background-removal step would need to be added
RoboNuggets · n28 - Nanobanana Pro for 3 Digital Product ideas- 1
Restaurant food-photo enhancement (Nano Banana Pro)PartlyA set of professionally styled food photos (hero shot plus an alternate camera-angle variant) generated from one ordinary phone/menu photo of a dish.
Nano Banana Prosingle enhanced hero food photocamera-angle variant (overhead / close-up / low angle)batched menu image set (e.g. 30 dishes x 2 images)- 1Source restaurant leads and reference photos · research · ApifyAn Apify Google Maps scraper pulls a batch of local restaurants (name, address, website, Instagram handle) plus existing dish photos, logged into an AirTable base.
- 2Filter for single-dish images · Check · n8nAn 'analyze image' node checks each scraped photo and keeps only images of a single food dish, skipping restaurant/interior shots.
Prompt
identify if the image is a single dish of food
- 3Generate premium hero shot · Picture · Nano Banana Pro · Wavespeed.ai / Kie.ai (automated) or Google Gemini (manual, free, watermarked)Nano Banana Pro edits the reference dish photo into a premium, professionally styled food shot, retaining the same ingredients while making it look more appetizing. Manual version: paste the photo into Google Gemini (thinking mode, Nano Banana Pro, 'create images' option) for a free, watermarked result; automated version runs the same edit via Wavespeed.ai or Kie.ai for a clean, watermark-free image at scale.
- 4Generate alternate camera-angle variant · Picture · Nano Banana Pro · Wavespeed.ai / Kie.aiThe generated hero image is fed back into Nano Banana Pro as a reference, asking for a different camera angle of the same dish (choosing between a direct overhead shot, a close-up detail shot, or a low-angle shot), giving the client more options per dish.
Missing: Exact verbatim prompt text lives in the course's linked Google Doc ('Prompts used'), not harvested into this lesson file; the n8n template JSON (n32 | Restaurant Photography) is an attached file pack, not read in full here — node-level configuration (credentials, exact API params) is not confirmed beyond what's narrated.
RoboNuggets · n32 - Restaurant Photography with Nanobanana Pro- 1
GPT Image 2 brand asset batch (Claude + GPT Image 2 skill)PartlyA one-shot batch of on-brand marketing images (social posts, magazine spreads, pricing cards, charts, merch mockups, UI mockups) generated by GPT Image 2 through Claude Code, using a brand book PDF and logo as references, with an optional handoff to Kling 3.0 for animating the winning frame.
GPT Image 2Kling 3.0social media post (grid variations)magazine spreadpricing cardschart/infographicmerch mockupUI component mockupanimated social post (via Kling handoff)- 1Install GPT Image 2 skill · By handGive Claude Code the install prompt to clone the robonuggets/gpt-image-2-skill repo, which unlocks text-to-image, image-edit and a 700+ JSON prompt library via a fal.ai connection.
Prompt
I'm giving you a skill called GPT Image 2 Skill. Get the files with: git clone https://github.com/robonuggets/gpt-image-2-skill Follow the README to set it up. Recommend how it would best apply to our setup in plain language, and ask me a few questions to clarify my intent before you install it.
- 2Prepare brand references · By handConvert the one-page brand book PDF to JPEG (GPT Image 2 does not accept PDFs) and gather the brand logo file path as reference images for Claude.
- 3Batch-generate a numbered grid · Picture · GPT Image 2 · fal.aiAsk Claude to prompt GPT Image 2 for a 5x5 grid of variations in a single generation, each cell numbered so a favourite can be isolated later.
Prompt
use GPT image 2 to make a 5 by 5 grid of cinematic RoboLab social media ads. Number each cell 1 to 25 in a small white circle at the top left so that I can isolate them later.
- 4Upscale the chosen cell · Upscale · GPT Image 2 · fal.aiAsk Claude to identify the prompt behind a chosen grid cell and regenerate just that one at the highest quality/resolution as a standalone image.
Prompt
so, I think number eight turned out really good. Can you just upscale that and generate a separate image for that at the highest quality settings, please?
- 5Animate the winning frame via Kling · Video · Kling 3.0 · fal.aiEnrich the skill with Kling 3.0 documentation copied from fal.ai's site, then pass the upscaled image as a first frame and ask Claude to animate it, keeping text fixed and aspect ratio the same.
Prompt
this image turned out really well. Can you now animate this via Kling? Just make sure that the text is fixed and that the aspect ratio stays the same.
- 6Generate a UI component mockup · Picture · GPT Image 2 · fal.aiUse GPT Image 2 to generate an image of a website component (e.g. a 'liquid glass' style UI element) as a design reference, which can then be fed to Claude Code to approximate in real HTML.
Missing: Needs the GPT Image 2 skill installed from the robonuggets GitHub repo, a fal.ai account/API connection for image and Kling generations, and a brand book PDF converted to JPEG.
RoboNuggets · n48 - Claude + GPT-Image-2- 1
Claude /generate skill for product image adsPartlya set of product image ad variants (and optionally an animated hero image/video-background website) generated across multiple image models and routed to the cheapest available provider.
GPT Image 2Nano Banana 2Nano Banana 2 LiteNano Banana ProKlingproduct image adsimage-to-video background animationwebsite hero section- 1Brief the /generate skill · WordsInvoke the /generate skill in Claude with a prompt naming the product, reference brand images, a total generation budget, an instruction to use a variety of image models, and a rule to always route to the cheapest available model provider.
- 2Route to cheapest provider · Check · Kie.ai, fal.ai, WavespeedThe skill picks the model and tries providers in cost order: Kie.ai first, then fal.ai, then Wavespeed, falling through if one is unavailable.
- 3Generate image ad variants · Picture · GPT Image 2, Nano Banana 2 Lite, Nano Banana Pro · Kie.ai / fal.ai / Wavespeed (whichever is cheapest/available)Generate multiple distinct image ads for the product across GPT Image 2, Nano Banana 2 (Lite) and Nano Banana Pro so the user can compare styles/models side by side.
- 4Animate chosen image to video · Video · KlingTake a favoured generated still and animate it to video with Kling to use as an animated background (e.g. for a follow-on website build).
- 5Log and review generations · CheckThe skill logs every prompt used and auto-loads all generated images/videos into a local 'generations' gallery tab (masonry grid) for visual review, plus a reusable 'styles' shelf of saved prompt+reference combos.
Missing: the /generate skill file itself (exact routing logic, prompt templates, cost-check rules) is a separate community resource not included in this lesson file, so exact implementation can't be reproduced from this source alone.
RoboNuggets · n68 - Replace Higgsfield with this Claude Skill- 1
AI character art factory (TikTok image carousel)RunsA 3-image TikTok carousel of a generated anime-style character in a curated art style, auto-posted with music.
gpt-image-1 (OpenAI)image carousel- 1Pick a style · Words · GPT-4.1 · OpenAI (via n8n)A GPT-4.1 'Style Agent' with a structured output parser selects one style from a curated list of 50 styles.
- 2Pick a character · Words · GPT-4.1 · OpenAI (via n8n)A GPT-4.1 'Character Agent' selects a character from a curated 50-character anime inclusion/exclusion list.
- 3Build final image prompt · Words · GPT-4.1 · OpenAI (via n8n)A GPT-4.1 'Final Prompt Agent' with a Think tool constructs the final text-to-image prompt from style + character.
- 4Generate image · Picture · gpt-image-1 · OpenAIThe final prompt is POSTed to OpenAI's image generation endpoint.
Prompt
{"model":"gpt-image-1","prompt":"Generate a high-quality image with this prompt: {{ $json.textToImagePrompt }}","size":"1024x1536","quality":"high","moderation":"low","output_format":"jpeg","background":"opaque","output_compression":50} - 5Host image · edit · ImgurBase64 image converted to a file and uploaded to Imgur for a public URL.
- 6Publish carousel · publish · BlotatoImages uploaded to Blotato media, then a 3-image TikTok carousel post created via Blotato with auto-added music.
RoboNuggets · R31 | The AI Digital Art Factory- 1
shirtRobo: A.G.E.N.T.-framework shirt design generatorManuala transparent-background PNG shirt-design image, generated by an AI agent from a shirt-design brief and saved to Dropbox
an AI agent built via the A.G.E.N.T. Framework (chat model unspecified)an image-generation model referred to in the transcript only as 'chatpt' (likely a ChatGPT/GPT image model, not confirmed)print-on-demand shirt design- 1Build the shirtRobo system prompt via the A.G.E.N.T. Framework GPT · By hand · a companion custom GPT ('A.G.E.N.T. Framework GPT'), used interactively, not called via API in this workflowUsing a separate, companion 'A.G.E.N.T. Framework GPT' (an interactive custom ChatGPT agent), the lesson author works through the A.G.E.N.T. framework's questions to generate a system prompt for a shirt-design agent, then pastes that generated system prompt into the n8n AI Agent node.
- 2Generate shirt design prompt (shirtRobo agent) · Words · chat model, exact provider/model not specified in the harvested transcriptThe configured AI agent (running the system prompt from step 1) takes a shirt-design brief/topic as input and produces a detailed image-generation prompt describing the shirt design.
- 3Generate the shirt design image · PictureThe generated image prompt is sent to an image-generation model (referred to in the transcript only as 'chatpt') to produce a transparent-background PNG of the shirt design.
- 4Upload to Dropbox · edit · DropboxThe finished PNG is uploaded to Dropbox for storage/hand-off.
- 5Log to Google Sheets · edit · Google SheetsLog the run (brief, generated prompt, Dropbox link) to a Google Sheet tracker.
Missing: the shirtRobo agent's own system prompt is only demonstrated interactively via a separate companion GPT and is never printed verbatim in this lesson; the image-generation step's exact model/endpoint is named only as 'chatpt' in the transcript with no API details shown. A separate downloadable n8n template file exists in the course resources ('__AGENT_framework___shirtRobo.json') which was not opened this session and could resolve both gaps if fetched and read.
RoboNuggets · R35 | The A.G.E.N.T. Framework for AI Agents (shirtRobo example, timestamp 15:57)- 1
30-day branded image-ad batch (making step inside a posting-focused lesson)Partly30 unique branded product images per campaign (mix of UGC-style and studio shots) plus matching captions, generated from one brand-guideline file and one product reference photo
Nano Banana Pro30-day image ad batch- 1Brand discovery · Words · agentAgent interviews the user or analyzes an Instagram feed screenshot to write brand.md (~100 lines: foundation, origin story, audience, positioning, pricing).
- 2Generate 30 unique image prompts + captions · Words · agentDrafts prompts referencing a product reference photo and prompt best practices.md, logged to Airtable.
- 3Generate images · Picture · Nano Banana Pro · Google AI Studio / Kie.aiGenerates the 30 images.
Missing: brand.md and prompt best practices.md content not captured verbatim (external template files not opened). Posting step (via Blotato to Instagram over 30 days) deliberately excluded from this entry per the making-vs-posting rule.
RoboNuggets · R57 | The Antigravity x Blotato Marketing Agent- 1
Carousel3
HTML-rendered social carousel and cover cardsRunsfeed cards, carousels (up to 28 layouts, 10 colour themes, 2 design systems), and matching WeChat cover pairs, rendered as PNGs from agent-written HTML
Xiaohongshu/Rednote carouselWeChat cover pair (21:9 hero + 1:1 share card)- 1Install · By hand · Guizang skillInstall via npx skills add (known bug: sometimes installs into ~/.agents/skills/ without linking; fall back to git clone into ~/.claude/skills/).
Prompt
npx skills add https://github.com/op7418/guizang-social-card-skill --skill guizang-social-card-skill
- 2Trigger · Words · Claude Code / CodexTrigger with a request naming a platform, 'social cards', or 'carousel'.
Prompt
Make me a Swiss-style Xiaohongshu carousel from this article, 5 cards, IKB blue. Here is the article: [paste your article text here]
- 3Pick design system and layout · Words · the skillAgent asks intent, picks a design system (Editorial for storytelling or Swiss for data/reviews), picks one of 28 layouts, sources images.
- 4Write cards · edit · the skill (HTML/CSS)Agent writes one HTML file per card.
- 5Render to PNG · edit · PlaywrightRender each card to PNG.
Prompt
node render.mjs
- 6Validate · Check · validate-social-deck.mjsLoad cards in a real browser and flag 6 failure modes: text overflow, oversized fonts, footer collisions, empty bands, image spill, too-heavy Swiss weights. Must be called by name, does not run automatically.
Prompt
node validate-social-deck.mjs path/to/task-dir
- 7Iterate · edit · the skillAsk for feedback / iterate.
- 8Optional WeChat cover pair · edit · the skillGenerate a matching cover pair (21:9 hero + 1:1 share card) in the same visual style.
Prompt
Now turn the same content into a WeChat cover pair: a 21:9 hero image plus a 1:1 share card, keep the visual style consistent with the carousel.
AIOps · Guizang Social Card Skill- 1
Carousel slide deck (Studio/Make)PartlyA social carousel deck built from draft content.
- 1render_deck verb · Pictureslides() dispatches makeSlides(), verb render_deck, gated by deck.ts.
Missing: Rendering logic not fully inspected (deck.ts).
Ours ·- 1
Render Studio brand-to-carousel/merch generatorPartlya set of branded, on-brand Instagram carousel ad slides (or merch/product mockup designs, when reskinned) with matching captions, generated from any input website's brand identity
Claude Sonnet 4.6GPT Image 2Instagram carousel ad slidesmerch/mockup design (reskin variant)- 1Brand scrape · researchScrapes a given website URL for colour palette, brand voice, fonts, images, product photos.
- 2Generate carousel ideas · Script · Claude Sonnet 4.6 · OpenRouterGiven a creative direction and idea count, generates N distinct carousel concepts each with a caption.
- 3Render carousel slides · Picture · GPT Image 2 · fal.aiGenerates the final branded carousel slide images matching the approved concept, with product correctly composited in.
- 4Schedule/publish · publish · BlotatoSchedules the approved carousel post with caption to the connected social channel.
Missing: the real prompt engineering lives inside the zipped app's server code/README, never opened here; brand-scrape method (Firecrawl? raw fetch?) not stated.
RoboNuggets · R64 | Render Studio- 1
Voice10
ElevenLabs voice design from a text descriptorRunsa calm, non-animated designed voice (male or female) generated purely from a text description, used to narrate an ad script
ElevenLabsvoiceover- 1Design the voice · Voice · ElevenLabs (Voice Design) · ElevenLabsIn ElevenLabs, click 'Create a Voice' > Voice Design, and type a text descriptor of the voice (accent, age, tone, pacing, mic/room quality).
Prompt
a soft, older British male speaking in a calm, assured way
- 2Select, name, and save · Voice · ElevenLabs · ElevenLabsPick a generated voice, name it, save with a label. Rule: keep it calm and slightly serious, storytelling tone — 'if it sounds like a YouTuber shouting, it's wrong.'
- 3Generate the narration · Voice · ElevenLabs · ElevenLabsPaste the ad script into the voice, click Generate Speech; listen before enhancing (raw output sounds slightly robotic); add enhancements and regenerate for a smoother, more natural take; download the audio.
Prompt
Alcohol is actually not that toxic compared to what it turns into in your body, and that's what makes your next day feel like hell.
AI Ads (Ad Creators Lab) · Generating Voice Over (Lesson 21) / Improving Your Audio Skills- 1
Coherent voice clone across clipsRunsa consistent cloned voice reused across an entire AI-generated video, matching a voice that first appeared in one generated clip
ElevenLabs (Instant Clone)voiceoverlipsync- 1Export the source audio · editStart from a Sora 2 Pro-generated clip that already has an AI voice in it. Export the video as audio only (ElevenLabs caps upload at 10MB, so extracting audio is needed).
- 2Clone the voice · Voice · ElevenLabs (Instant Clone) · ElevenLabsIn ElevenLabs, Clone Voice > Instant Clone, drag in the exported audio. Remove background music from the source audio first if present (leaving music in ruins the clone); if there's no music, leave as is.
- 3Use the cloned voice · Voice · ElevenLabs / RIS · ElevenLabs / Max FusionEither write a script and generate speech directly in ElevenLabs, or go back into Max Fusion and use the RIS feature to generate speech from other images/footage using the cloned voice.
Prompt
There's a reason hair growth companies don't use this ingredient.
AI Ads (Ad Creators Lab) · Creating coherent AI voices for your videos- 1
Real-creator voice clone with matched backing trackRunsa cloned narrator voice matching a specific real YouTube creator's tonality, paired with a genre-matched backing track, used as ad voiceover
ElevenLabs (Instant Voice Clone, V3 model)voiceover- 1Source and download the voice · research · Y2MateFind the target creator's videos on YouTube (example: Zach D Films), download the audio via a YouTube-to-mp3 tool.
- 2Clone the voice · Voice · ElevenLabs Instant Voice Clone (V3 model) · ElevenLabsElevenLabs Voices > Add > Instant Voice Clone, drag in the audio. Explicitly do NOT check 'remove background noise' — this specific clone degrades if you do. May need two source videos to get it right. Use the ElevenLabs V3 model for tonality.
- 3Generate the script · Voice · ElevenLabs · ElevenLabsGenerate Speech with the new script; minor residual background noise in the clone is acceptable, gets masked once music is added.
Prompt
If you're sleeping 8 hours but still can't drag yourself out of bed, your body...
- 4Add a backing track · music · Epidemic SoundSearch Epidemic Sound for a genre-matched track (used: genre 'sneak') and download it to sit behind the voiceover.
AI Ads (Ad Creators Lab) · What we are making - Zack D Films Style- 1
Custom AI UGC voice design (ElevenLabs + Gemini Gem)RunsA brand-new, room-appropriate character voice (not a catalogue voice) generating spoken lines as audio, built via a Gemini assistant fed ElevenLabs' own prompting best-practices.
ElevenLabs Voice DesignElevenLabs V3 (Alpha)ElevenLabs Multilingual V2ElevenLabs Instant Voice CloneUGC character voicevoice clone- 1Build a voice-prompt-engineer Gem · Words · Gemini (custom Gem) · GoogleCopy ElevenLabs' voice-prompting best-practices page into a Google Doc, download as PDF, create a Gemini Gem with that PDF as its knowledge file.
Prompt
you are an ElevenLabs voice prompt engineer, follow the resources in your knowledge and create me high quality voices.
- 2Generate a voice-design prompt for the character · Words · Gemini (custom Gem) · GoogleUpload the character's reference photo to the Gem plus an instruction that the voice should sound like it belongs in the character's room.
Prompt
create me a high quality ElevenLabs voice prompt.
- 3Generate the voice · Voice · ElevenLabs Voice Design · ElevenLabsPaste the Gem's prompt into ElevenLabs Voice Design and generate; returns 3 options to choose from.
- 4Generate speech · Voice · ElevenLabs V3 or Multilingual V2 · ElevenLabsPick ElevenLabs V3 (Alpha, more expressive, supports audio tags like laughing/whispering, 2 output variants per generation) or Multilingual V2 (more consistent, better for automation) and generate the script as speech with the chosen voice.
- 5Alternative: clone an existing voice · Voice · ElevenLabs Instant Voice Clone (or Professional Voice Clone) · ElevenLabsInstead of designing a voice, drag in 10+ seconds of a real reference recording (e.g. from TikTok/YouTube) into Instant Voice Clone, then use the same text-to-speech step.
The AI Content System · ElevenLabs Full Audio Guide- 1
Local voice clone, dub, and transcriptRunscloned narration audio, a dubbed video (translated, re-synthesised, original timing preserved), and diarized transcripts, all run locally with no cloud upload
OmniVoice (voice clone + synthesis, CC-BY-NC weights)Whisper Tiny (transcription)Argos / NLLB-200 (local translation)cloned voiceoverdubbed videotranscript- 1Install · By hand · VoiceStudio (local)Install via one curl command or the releases page (Apple Silicon M1+/16GB RAM required on Mac; Intel Mac unsupported). Downloads OmniVoice weights (~2.4GB) on first run.
- 2Record a reference clip · Voice · VoiceStudio voice-cloning workspaceRecord a 3-10 second clean reference clip and fill in its transcript field (skipping this forces a full speech-recognition pass and slows every later generation).
- 3Clone and generate · Voice · OmniVoice (CC-BY-NC weights, non-commercial only) · VoiceStudio (local)Clone the voice and generate speech from text.
- 4Dub · Voice · Whisper Tiny (transcription), Argos/NLLB-200 (translation), OmniVoice (re-synthesis) · VoiceStudio dubbing workspaceChain transcription, then translation (Argos or NLLB-200, kept on-device; Google Translate/DeepL send data off-machine), then re-synthesis, preserving original timing.
- 5Connect from an agent · By hand · MCP (mirrors OpenAI's audio API shape at local port 3900)Connect via claude mcp add; set OMNIVOICE_MCP_OUTPUT_MODE=files and OMNIVOICE_MCP_BASE_PATH=<folder> first, otherwise raw audio bytes overflow the agent's context.
AIOps · VoiceStudio: Local Voice Cloning, Video Dubbing, and Transcription via MCP- 1
Agent-driven ElevenLabs audio productionRunsvoiceovers, cloned-voice reads, cleaned/diarized transcripts, ambient soundscapes, and podcast-to-clip hook intros, all via an MCP-connected agent
ElevenLabs voice cloningeleven_multilingual_v2ElevenLabs isolate_audioElevenLabs speech_to_textElevenLabs generate_soundscapeElevenLabs text_to_speechvoiceover / VSL narrationcloned-voice readisolated + diarized transcriptambient soundscapepodcast-clip hook intro- 1Install and connect · By hand · ElevenLabs MCPSign up (10K free credits/month), get an API key, install elevenlabs-mcp via pip or uvx, register it in Claude Code with the API key plus output-path and output-mode env vars.
Prompt
claude mcp add elevenlabs uvx elevenlabs-mcp --env ELEVENLABS_API_KEY=sk_your_real_key_here --env ELEVENLABS_MCP_BASE_PATH=$HOME/Desktop/elevenlabs-output --env ELEVENLABS_MCP_OUTPUT_MODE=files
- 2Smoke-test TTS · Voice · default voice (e.g. Rachel) · ElevenLabsRender a short test line with the default voice.
Prompt
Use the ElevenLabs MCP to generate a 15-second voiceover that says "This is a smoke test from Claude Code, the audio pipeline is live." Use the default Rachel voice and save it to ~/Desktop/elevenlabs-output/smoke-test.mp3. Then tell me the file path and the credit cost.
- 3Clone a voice · Voice · ElevenLabs voice cloning · ElevenLabsClone from a 60-second sample, name and describe it, run a 10-second test read.
Prompt
Use the ElevenLabs MCP to clone a voice from the sample at ~/Desktop/elevenlabs-output/founder-sample.wav. Name the voice "Founder - {YourName}" and add a description: "Conversational, mid-tempo, North American English." After the clone succeeds, return the voice_id and run a 10-second text_to_speech test with the new voice saying "Hey, this is {YourName}, testing the clone." Save the test to ~/Desktop/elevenlabs-output/clone-test.mp3. - 4Render a VSL/script voiceover · Voice · eleven_multilingual_v2 with cloned voice_id · ElevenLabsRender voiceover from a script file with a cloned voice_id, splitting scripts over 2500 chars at paragraph breaks into numbered files.
Prompt
Read the script at ~/Desktop/elevenlabs-output/vsl-script.md. Use the ElevenLabs MCP to generate a single text_to_speech file using voice_id {YOUR_CLONED_VOICE_ID} with the eleven_multilingual_v2 model, stability 0.45, similarity 0.85, style 0.30, speaker_boost on. Save the output to ~/Desktop/elevenlabs-output/vsl-final.mp3. Report the duration and the credit cost. If the script is longer than 2500 characters, split it into chunks at paragraph breaks, render each chunk to a numbered file, and tell me the order. - 5Isolate and diarize · Check · ElevenLabs isolate_audio + speech_to_text · ElevenLabsIsolate vocals from a noisy recording, then transcribe with speaker diarization.
Prompt
Use the ElevenLabs MCP isolate_audio tool on ~/Desktop/elevenlabs-output/podcast-raw.mp3. Save the cleaned vocals-only file to ~/Desktop/elevenlabs-output/podcast-clean.mp3. Then run speech_to_text on the cleaned file with speaker diarization on, and save the transcript with speaker labels to ~/Desktop/elevenlabs-output/podcast-transcript.txt.
- 6Generate a soundscape · music · ElevenLabs generate_soundscape · ElevenLabsGenerate an ambient soundscape from a scene description.
Prompt
Use the ElevenLabs MCP generate_soundscape tool to create a 12-second ambient audio bed for a B2B SaaS ad opener. Prompt: "Quiet open-plan office at 7 AM, distant typing, soft HVAC hum, single faraway phone ring, no music, no voices." Save the file to ~/Desktop/elevenlabs-output/ad-opener-bed.mp3. Report the cost.
- 7Podcast-to-clip routine · Voice · isolate_audio, speech_to_text, text_to_speech (cloned voice) · ElevenLabs, response_timeout_secs 180, trust_context lowWatched-folder routine: isolate, diarize/transcribe, identify 3 highest-emotion 60-90s segments as JSON (start/end/title/hook_caption), render a 5-second cloned-voice hook intro per segment.
Prompt
You are a podcast-to-clip pipeline. Watch ~/Dropbox/podcast/incoming/ for new .mp3 or .wav files. For each new file: (1) run ElevenLabs isolate_audio to clean the recording and save to ~/Dropbox/podcast/cleaned/. (2) run speech_to_text with diarization on the cleaned file and save the transcript to ~/Dropbox/podcast/transcripts/. (3) read the transcript, identify the three highest-emotion 60-to-90 second segments (look for laughter markers, voice raises, story arcs), and write a JSON file at ~/Dropbox/podcast/clips/{episode_name}.json listing the segments as {start_seconds, end_seconds, title, hook_caption}. (4) for each segment, use text_to_speech with my cloned voice {YOUR_CLONED_VOICE_ID} to render a 5-second hook intro from the hook_caption and save to ~/Dropbox/podcast/hooks/{episode_name}-{N}.mp3. Set response_timeout_secs to 180 on every MCP call because long renders can run past the default 30. Run trust_context low because this is an unattended automation. Report a summary table when done.
AIOps · ElevenLabs MCP (Official): Agent-Driven Marketing Audio End to End- 1
Voice reference asset (face-hidden)RunsA short black-screen video carrying an extracted real voice clip — not a generation, a packaging step so a voice can be fed into Google Omni for cloning without exposing the source person's face.
- 1Build voice reference · editPure ffmpeg buildVoiceRef extracts the audio track and renders it against a black frame — no AI/API call.
Ours ·- 1
AI Voice Generation for UGC (MiniMax primary / ElevenLabs custom-only)RunsA voice track for a UGC script, matched to a character brief via 3 generated candidate voice designs.
MiniMaxElevenLabsGemini- 1Design a voice-design prompt · Voice · GeminiAsk Gemini for an ElevenLabs voice design prompt for the character, ensuring the voice sounds like it's in an actual room.
Prompt
Give me an ElevenLabs voice design prompt for this character. Make sure the voice sounds like it's in an actual room.
- 2Generate + pick · Voice · MiniMax / ElevenLabsMiniMax (pre-made voice, or Voice Design/Voice Clone from 10s audio) OR ElevenLabs custom-only (never pre-made/speech-to-speech). Generate 3 options, pick the best.
Ours ·- 1
AI podcast episode generator (script + voice)PartlyA 2-3 minute spoken podcast episode (script written by GPT, voiced by ElevenLabs TTS in a chosen host persona), delivered as an MP3 ready for manual upload to Spotify for Podcasters.
ElevenLabs (TTS)podcast episode- 1Generate episode ideas · Words · ChatGPT · OpenAIChatGPT generates a list of episode titles and topic descriptions for the chosen podcast topic.
Prompt
Give 10 ideas for topics for a podcast about [topic]... Col 1: Title, Col 2: Topic description...
- 2Write episode script · Script · ChatGPT · OpenAIA ChatGPT agent writes a 2-3 minute podcast script per episode idea, in a named host persona.
- 3Generate voice audio · Voice · ElevenLabs TTS · ElevenLabsElevenLabs TTS reads the script aloud in a chosen voice (e.g. 'Brian' or 'Charlotte' from the ElevenLabs voice library).
- 4Store episode · edit · DropboxNumbered/titled MP3 uploaded to Dropbox; Google Sheets row marked done.
- 5Publish (manual) · By hand · Spotify for PodcastersEpisode manually uploaded/scheduled on Spotify for Podcasters, since no publishing API is available.
Missing: Spotify publish step has no API and must be done by hand each time.
RoboNuggets · R6 | AI Agent Automation for Spotify Podcasts- 1
Cloned voice via ElevenLabs voice sharingManuala personal AI voice clone uploaded to ElevenLabs, made speakable by any ElevenLabs-compatible model and (optionally) shared to the public Voice Library for per-use payouts
ElevenLabs Professional Voice Clonevoice clone- 1Sign up and start a Professional Voice Clone · By hand · ElevenLabsIn the ElevenLabs web app go to Voices, click Add a new voice, choose Professional Voice Clone, name the voice and choose its language.
- 2Upload training audio · By hand · ElevenLabsUpload an audio file of the speaker's own voice: at least 30 minutes of speech required, ElevenLabs states hours is optimal.
- 3Label the voice · By hand · ElevenLabsAdd five searchable labels using ElevenLabs' own Voice Library categories: accent, gender, age, style (e.g. 'calm'), and category (e.g. 'educational'), so the voice is discoverable in search.
- 4Write the voice description · Words · ChatGPT (model unspecified) · OpenAI (ChatGPT web app)Take a screenshot of the label area, upload it to ChatGPT with a prompt (kept in the RoboNuggets community, not reproduced in this lesson) to draft descriptive sentences for the voice, then edit the output as the voice description.
- 5Create the professional voice and verify · By hand · ElevenLabsClick Create professional voice to upload the sample, complete voice verification by reading and recording a supplied verification passage, then finish. Once processing completes (about an hour) the cloned voice is ready to use.
- 6Enable all recommended models · By hand · ElevenLabsOpen the finished voice and, under Recommended models, tick every model so the cloned voice is usable across ElevenLabs' full model range.
- 7Share to Voice Library and enable payouts (optional monetization) · By hand · ElevenLabsToggle on voice sharing and Discovery in Voice Library, connect a Stripe account for payouts, enable financial rewards, set a removal notice period, optionally enable live moderation and a custom preview line, then submit for ElevenLabs review (approval takes about a day). Once approved the voice earns passive income whenever other users generate speech with it.
Missing: entirely manual ElevenLabs web-app workflow — no API calls or n8n automation are shown anywhere in this lesson, so it cannot be run as an unattended pipeline. The ChatGPT prompt used to write the voice description is referenced but never printed verbatim in the lesson (only 'find it in our community'). Requires an ElevenLabs Creator-plan-or-above paid subscription to do Professional Voice Cloning.
RoboNuggets · R15 | Clone your Voice with AI (Passive Income)- 1
Music3
Suno ad jingle/songRunsa singing/jingle-style song used as ad audio, with lyrics built around product pain points and USPs
Sunojinglehip hopR&B/pop- 1Write lyrics · Words · ChatGPT / Claude · ChatGPT / ClaudeGenerate lyrics via ChatGPT/Claude using product pain points/USPs as inputs.
- 2Write the Style prompt · music · Claude / ChatGPT · Claude / ChatGPTAsk Claude/ChatGPT for a Style-field description matching a desired genre feel; Suno rejects prompts naming real artists directly, so the naming happens in the chat step, not in Suno's own field.
Prompt
Hey, I wanna have a hip hop song. I want it to sound like, I don't know, Tupac, Biggie, whatever.
- 3Generate in Suno · music · Suno · SunoSuno Pro plan, Create > Advanced tab, paste lyrics into the Lyrics section and the Style prompt into the Style section. Match genre to target audience (e.g. avoid hip hop/dubstep for 50-year-old moms; use old-school pop/R&B for an older generation) and test multiple genres per ad.
Prompt
See cologne on his song just evaporates fast. Alcohol hits the air and the scent doesn't last. But oil bonds to skin, that's the move that they miss. Lay your Penrose body butters, Baccarat time for—
- 4QC, iterate, download · Check · Suno · SunoCheck pronunciation of product/ingredient names and that lyrics make sense. Downvote disliked songs, upvote and Remix liked ones to bring a favorite in as a base for a new take. Download as MP3.
AI Ads (Ad Creators Lab) · How To Generate AI Songs for Your Brand (Lesson 11)- 1
Suno V5 song remixerPartlyone remixed/finished song (audio file) produced from a source track through an n8n automation.
Suno V5song remix- 1Connect Kie.ai credentials · Check · Kie.aiLink a Kie.ai account inside the n8n template so the workflow can call Suno V5 through Kie.ai's cheaper instance.
- 2Connect output storage · Check · n8nLink file storage (Box.com in the source example) in the workflow's Output section so finished remixes land somewhere durable.
- 3Remix song · music · Suno V5 · Kie.aiRun the n8n template, which sends the source track/settings to Suno V5 via Kie.ai to generate a remixed song and writes it to the linked storage.
Missing: the actual n8n workflow JSON (node config, exact Suno remix parameters) was not opened/read, only the lesson text describing it.
RoboNuggets · b3 - Suno V5 Song Remixer- 1
Lo-fi music album generator (Suno via PiAPI, Make.com automated)PartlyDozens of unique lo-fi MP3 tracks (30 from 15 prompts, since Suno returns two takes per prompt) around a chosen theme/album, auto-generated and delivered to Dropbox, then manually assembled with a generated thumbnail image into a 1-hour+ YouTube video in DaVinci Resolve.
Suno (music, via PiAPI)lofi music album1-hour ambient music video- 1Generate song titles + prompts · Words · ChatGPT · OpenAIChatGPT generates a table of song titles and Suno prompts for an album/theme (r18 adds a suggested YouTube title, description and thumbnail image prompt in the same response).
Prompt
I am generating music for -my lo-fi music channel ... Your task: output a table with 30 rows and two columns: -Title - a one word title in lower case for the song to generate... -Prompt - a detailed prompt for Suno AI to generate the lo-fi music. the prompt should be a maximum of 200 characters and should include the word "lo-fi"
- 2Request music generation · music · Suno (music-s) · PiAPIMake.com HTTP module POSTs the prompt to Suno's music-generation endpoint via PiAPI (model 'music-s' in r18), with a webhook callback configured to receive the finished track.
Prompt
{"model": "music-s", "task_type": "generate_music", "input": {"gpt_description_prompt": "{{1.`4`}}", "make_instrumental": true}, "config": {"service_mode": "public", "webhook_config": {"endpoint": "[insert-webhook-here]", "secret": ""}}} - 3Receive and store track(s) · edit · DropboxA receiver webhook scenario extracts the MP3 URL(s) from Suno's callback (regex pattern) and uploads each to Dropbox, named by index and title.
- 4Generate thumbnail image · Picture · ChatGPT image / Microsoft Designer · OpenAI / MicrosoftChatGPT (or Microsoft Designer as an alternative) generates a background/thumbnail image from a supplied style prompt.
Prompt
Create a retro, vintage lo-fi scene with a stylish whiskey glass floating in mid-air. The background features a deep, rich gradient from black to navy, capturing the 1980s aesthetic - the background is minimalistic and without any other elements...
- 5Assemble video (manual) · By hand · DaVinci ResolveDaVinci Resolve: all generated MP3s dropped onto the timeline in sequence, thumbnail image stretched to cover the full duration, rendered as a single long video.
Missing: Music generation is fully API-automated; thumbnail generation and final video assembly are manual/UI steps. Suno's direct API was deprecated April 2025 per the lesson's own warning, requiring the PiAPI workaround.
RoboNuggets · R13 | YouTube Lo-Fi Channel AI Toolkit / R18 | Music Creator AI Agent Toolkit v2.0- 1
Motion copy6
Video restyle (video-to-video, 20 style presets)PartlyAn existing video re-rendered in one of 20 named styles (claymation, lego, pixel-art, etc.) — style + source URL only, no text prompt.
Omni Flash (per code comment only)claymationlegopixel-art+17 more- 1queueRestyle · Video · Omni Flash (comment only, unverified in repo)Operator picks a style + source video; queueRestyle() inserts action_type restyle_video.
Missing: Provider only named in a code comment, not verified call code.
Ours ·- 1
Kling Motion Control (motion transfer onto existing character)PartlyA UGC/talking video where an existing character image is animated using the motion of a separate reference video.
Kling Motion Control- 1Pick sources · By handPick a reference video (motion source) and a character image (identity/scene source).
- 2Generate · Video · Kling Motion Control · HiggsfieldKling Motion Control, mode 'Match Video' (motion-driven, ≤30s) or 'Match Image' (framing-driven, ≤10s, talking).
Prompt
natural lighting, realistic motion
Missing: No account/API/pricing detail in the source file.
Ours ·- 1
Kling Character Swap / Motion CloningPartlyA UGC video that clones a viral reference video's exact motion onto a new identity (a face-swapped first frame).
Nano Banana ProKling Motion Control- 1Source a viral video · researchSource a reference/viral video via Kalo Data, PiPiAds, Facebook Ad Library, TikTok/IG.
- 2Download + anchor · By handDownload clean (SnapTik for TikTok, or the DevTools method for FB Ad Library), screenshot the first frame as the 'anchor'.
- 3Face-swap the anchor · Picture · Nano Banana ProSwap the person in the anchor frame, keeping pose/background identical.
Prompt
Change the person on the image to a slightly different looking [describe who you want]. Keep pose and background the same. Don't change anything except the actual person.
- 4Generate · Video · Kling Motion Control · HiggsfieldKling Motion Control (v2.6 named in one source), ref video + new first frame, mode Quality/Scene Control, background = new first frame.
Missing: No account/cost info given; 'DevTools method' referenced but not detailed in this file.
Ours ·- 1
Product motion-graphics transition videoPartlyone combined, upscaled (up to 4K/60fps) motion-graphics video built from a sequence of image-to-image transition frames animated into clips and merged
Seedream 4.0Nanobanana (text-rendering fallback)Hailuo-02 ProFFMPEG merge-videosTopaz Labs upscaleproduct motion graphics- 1Read input row · By hand · Google SheetsReads title, scene, production status, Image Reference A, Image Reference B, transition prompt.
- 2Generate reference frames · Picture · Seedream 4.0 (fallback Nanobanana) · Kie.ai playgroundManual/interactive step (not wired into n8n): edits/combines reference images to produce start and end frames per scene, falling back to Nanobanana for on-screen text.
Prompt
edit the phone in file one to be in the angle orientation of file 2...
- 3Generate transition video · Video · Hailuo-02 Pro · fal.ai (fal-ai/minimax/hailuo-02/pro/image-to-video)Animates the transition between the start and end frame.
- 4Update row · Check · Google SheetsMarks the row done and stores the video URL.
- 5Aggregate scenes · edit · n8nGroups all scene clips into one item.
- 6Merge · edit · FFMPEG merge-videos · fal.aiMerges all scene videos into one.
- 7Upscale · Upscale · Topaz Labs upscale · fal.ai (fal-ai/topaz/upscale/video)Upscales the merged video.
- 8Final save · Check · Google SheetsStores the final output URL.
Missing: the frame-generation step (Seedream/Nanobanana) is explicitly manual/interactive, not automated in n8n; would need separate automation to make this end-to-end. Automated merge step requires >=2 scenes (single-scene use unsupported).
RoboNuggets · R41 | AI Motion Graphics (with Seedream 4.0)- 1
AI character/motion transfer (motion capture replacement)Runsa video where a person's motion, expressions and (optionally) voice from a driving video are transferred onto a different reference character/image, either replacing them in the same scene or moving them into a new one, with matched voice if the audio path is used
Wan 2.2 Animate (move/replace)Chatterbox speech-to-speechFFmpeg merge-audio-videoextract-audio (Replicate: lucataco/extract-audio)motion transfermotion transfer + voice clone- 1Load input rows · By hand · Google SheetsReads sheet rows where status = create (driving video URL, character image URL, output column, status).
- 2Transform video (motion transfer) · Video · Wan 2.2 Animate (move) · fal.aiTransfers the driving video's motion onto the character image.
- 3Poll/fetch result · Check · fal.aiPolls the request status.
- 4Log output · Words · Google SheetsUpdates the sheet row with the MP4 URL and status = done.
- 5Load input rows (voice variant) · By hand · Google SheetsReads a second sheet including audio reference URL and type (replace/move).
- 6Extract audio from driving video · edit · extract-audio · ReplicatePulls the MP3 from the driving MP4.
- 7Transform video (move or replace, per row) · Video · Wan 2.2 Animate (move or replace) · fal.aiSame as step 2 but dynamically switches between 'move' and 'replace' endpoints per the row's type field.
- 8Poll/fetch video · Check · fal.aiPolls the status.
- 9Transform audio (voice clone) · Voice · Chatterbox speech-to-speech · fal.aiClones the source audio onto the target reference voice.
- 10Poll/fetch audio · Check · fal.aiPolls the status.
- 11Merge video + transformed audio · edit · FFmpeg · fal.aiCombines the video and voice branches into the final clip.
- 12Poll/fetch final combined video · Check · fal.aiPolls the status.
- 13Log output · Words · Google SheetsUpdates the row with the final MP4 and status = done.
RoboNuggets · R47 | The AI Motion Capture System- 1
Claude Code + Remotion promo-video generatorPartlya fully animated, branded promo/marketing video (motion graphics, text, logo, sourced photos/video from a website) in landscape or vertical, one-shot from a single text prompt, editable via further chat
Claude (Claude Code agent, writes Remotion/React code — no separate image or video generation model is called)promo video, horizontalpromo video, vertical (9:16)- 1Install · By hand · Anthropic + RemotionInstalls VS Code, Claude Code extension, Node.js, Git; runs npx skills add remotion-dev/skills.
- 2Brief the agent (one-shot) · Video · Claude (Claude Code), Remotion renderer · AnthropicClaude Code autonomously scrapes site assets (photos/videos), writes copy, and builds an animated Remotion composition.
Prompt
Make a promo video for the Sweat app: https://sweat.com/ Make sure to use photos available in that website, and include videos as well. Use the sweat app logo available in sweat app white.png in the assets. Horizontal video. Ensure brand and font is aligned to website. Call to action is 50% off this February 2026. Make it look like a $100,000 video. Sleek and well-animated, professionally designed.
- 3Iterate · edit · Claude · AnthropicConversational edits (e.g. change CTA text/date, change colour tone) applied live to the Remotion project.
- 4Render · edit · Remotion · Remotion (local)Click Render in Remotion Studio to export the final MP4.
Missing: not a hosted-API workflow — requires Claude Code (agentic coding tool) plus local Remotion rendering, not simple Kie.ai/fal.ai/MuAPI calls. No named generative image/video model: the video is 100% code-generated motion graphics plus scraped real assets, not AI-generated imagery. Requires Claude Pro and a local dev environment.
RoboNuggets · R53 | Vibecreate Videos with Claude Code- 1
Edit and finish17
Music-video lyric-synced editRunsa music-video-style ad edit, cutting available clips to a previously generated song's lyric phrases, with keyframed camera moves
music video- 1Lock the song, cut in on the first lyric · editLock the song audio to the timeline. Cut the video to start on the first sung lyric line, not any intro, since it's an ad.
- 2Sync clips to lyric phrases · editDrag in available clips matched to lyric phrases in order. When a clip is too short/missing for a phrase, go back to the same chat used to generate video prompts and ask for a new visual idea for that phrase, stating the problem with the current clip.
Prompt
Generate some visual ideas for this part.
- 3Speed and keyframe · editSpeed-adjust clips to match lyric timing/pacing; keyframe camera pan/zoom manually (e.g. start on a close-up, pan/zoom out) across each lyric-matched clip.
Missing: no editor software named
AI Ads (Ad Creators Lab) · Simple video editing from song lyrics- 1
Realistic AI UGC video editRunsa full realistic, non-AI-looking UGC-style ad edit that mixes AI-generated shots with real B-roll, disguising AI generation seams and looking like an authentic creator video
UGC ad edit- 1Timeline from voice first · edit · ElevenLabs · Max FusionGenerate voiceover in ElevenLabs first, cut it down on the timeline to establish pacing before adding visuals; reuse the same cloned voice in Max Fusion to generate any extra needed dialogue matching the timeline's line length.
- 2Add a TikTok comment overlay · edit · TikTok comment generator (unnamed)Use a TikTok comment generator tool to create a random profile image/name and custom comment text as a PNG, layer onto the video as a pattern-interrupt.
- 3Blend AI shots with real B-roll · edit · Premiere ProMix AI-generated shots with real human iPhone footage of matching skin tone. Limit close-up AI-talking shots (they look worst up close); show the AI talker full-frame briefly then push into a small moving green-screen box with real B-roll photos playing behind. Use jump cuts for pacing.
- 4Keep the product visually consistent · Picture · Nano Banana · Max FusionUse Nano Banana to keep the product (e.g. a ring) visually consistent across shots.
- 5Style the captions · edit · Premiere Pro (plugin, name unconfirmed)Use a Premiere Pro social-caption graphics template plugin (named as spoken, unconfirmed spelling: 'MoGuard'), added via the Graphics Templates panel. Pick a native-platform style (e.g. Classic), type the caption, break lines to fit, use Roboto Black font, pick native platform colors. Other named styles: Modern, Neon, Typewriter. Only use for the hook caption — hand-building native-style captions won't look authentic.
Prompt
replace your current toothbrush
- 6Fix the CTA shot and add realism · edit · Premiere ProIf a generated CTA shot looks broken (e.g. hovering mic artifact), green-screen a different clean shot of the same line over it, keeping the voice audio glitch-free. Add Premiere Pro's built-in 'Wiggle' effect to a static selfie clip to simulate a real handheld phone video.
Prompt
Go to this page, pick out seven pieces of waterproof jewelry.
- 7Smooth the hook-to-main transition · edit · Premiere ProPick a deliberately different-vibe song for the intro section, add a 'Fierce Riser' SFX at the transition point (nudged for breathing room using the Gadget tool), a hit/cymbal sound as the new section starts, and a whip-transition visual effect.
- 8Mix and finish · edit · Premiere ProAdd seasonal pop music, mix/duck with Premiere Pro's Essential Sound panel. Philosophy: no flashy transitions/VFX/whoosh sounds — real creators don't use fancy effects.
Missing: caption plugin name unconfirmed
AI Ads (Ad Creators Lab) · Creating a realistic AI video editing process and tips / Enhancing Captions in Premiere Pro / Adding UGC content and smoothing transitions- 1
CapCut UGC reaction-video editRunsa raw UGC reaction/taste-test-style ad edit (e.g. wine taste test), assembled live from multiple reaction takes with resynced audio and sound effects
UGC reaction video- 1Select best reaction takes · edit · CapCutAssemble reaction clips of a UGC creator taste-testing two options, selecting the best reaction takes.
- 2Remix audio to best visuals · edit · CapCutExtract audio from a clip (Select clip > Voice > Extract Audio) to pair the best spoken reaction with the best visual reaction shot; shift clips along the timeline to resync.
- 3Add action SFX · edit · CapCutSearch the CapCut SFX library (e.g. 'audio slurp'), audition several options, pick the best/shortest for an action shot (e.g. a sipping zoom-in).
- 4Captions and hook overlay · edit · CapCutAdd captions/subtitles via CapCut's caption tool, adjusting timing/sync iteratively; add a title/hook text overlay early in the edit.
Prompt
No freaking way, my Italian husband just sent me a hundred dollars. I can't tell the difference between his fancy bottle wine and boxed wine.
- 5Emphasize with zoom · edit · CapCutUse zoom-in shots to emphasize physical reactions (e.g. sipping).
AI Ads (Ad Creators Lab) · 18 mins Uncut Live Editing AI UGC Video (Lesson 16)- 1
General AI-video upscale (Topaz + CapCut)ManualA grainy raw AI video clip cleaned up to a sharper, natural-feeling finished clip.
Topaz (Proteus model)any- 1Upscale in Topaz · Upscale · Topaz Proteus · TopazRun the raw clip through Topaz at fixed settings.
- 2Downscale the frame rate · edit · CapCutBring the upscaled file into CapCut and export at 30fps so it feels natural/camera-like.
Missing: Shown only as desktop-app actions in Topaz and CapCut; no API route mentioned for either.
The AI Content System · Main Upscaling Workflow- 1
Sora 2 upscale and realism grade (Higgsfield + Topaz + CapCut)ManualA Sora 2 clip (Pro or Regular) brought up to a sharper, more camera-like finished grade, with Regular clips reaching quality comparable to Pro through an extra pass.
Higgsfield Sora 2 UpscalerHiggsfield Sora 2 EnhancerTopaz (Proteus or Gaia)any Sora 2 output- 1Enhance (Sora 2 Regular clips only) · Upscale · Higgsfield Sora 2 Enhancer · HiggsfieldRun the raw Sora 2 Regular clip through Higgsfield's Sora 2 Enhancer to remove grain before the Topaz step. Skip this step for Sora 2 Pro clips.
- 2Upscale (Sora 2 Pro clips) / or continue from step 1 · Upscale · Higgsfield Sora 2 Upscaler · HiggsfieldRun the clip through Higgsfield's Sora 2 Upscaler.
- 3Grade in Topaz · Upscale · Topaz Proteus or Gaia · TopazRun the Higgsfield output through Topaz using Proteus or Gaia.
- 4Finish in CapCut · edit · CapCutDownscale to 30fps and apply a fixed realism grade.
Missing: All four tools are used through their web/desktop UIs; no API route is mentioned for Higgsfield's enhancer/upscaler, Topaz, or CapCut in this lesson.
The AI Content System · Sora 2 Upscaling Workflow- 1
Seedance 2.5 video edit (smart edit / Omni Reference person, language and object swap)RunsAn existing reference video re-rendered with the person swapped, the spoken language changed, or on-screen text/objects removed, keeping the rest of the original shot identical.
Seedance 2.5UGC video re-edit- 1Source and trim the reference video · research · TikTokDownload a TikTok reference video and trim it to fit Seedance 2.5's 30s clip cap (demo trims to 15s for faster generation).
- 2Extract the first frame and build a replacement image · Picture · GPT Image 2 or Nano Banana Pro · ChatGPT / Google FlowCapture the video's first frame (via Google Flow or CapCut) and regenerate a replacement-character/edit image from it.
Prompt
remove all text and buttons on the screen. Also change her dress color to red.
- 3Edit in Dreamina smart edit / Omni Reference mode · Video · Seedance 2.5 · Dreamina (or Higgsfield / fal.ai / Kie.ai)Upload the reference video, switch to smart edit mode, select Seedance 2.5, attach the replacement image, and prompt the edit. For localized fixes (e.g. removing a logo) use "edit with marks" to box the area instead.
Prompt
replace the person in the original video with the person attached. Keep everything about the original video the same.
- 4Language + person swap variant · Video · Seedance 2.5 · DreaminaSame pipeline in Omni Reference mode, attaching both the video and a reference photo.
Prompt
switch the person in the video attached and change the language she is speaking to French. Keep everything else in the video exactly the same. Just change the language and the person.
- 5Handle a policy rejection · CheckIf the platform refuses the prompt as a possible guideline violation, try a different reference image, or switch from "smart edit" to "Omni Reference" mode; other providers (Higgsfield/fal.ai/Kie.ai) rarely throw this error.
The AI Content System · How To Do AI Video Editing With Seedance 2.5- 1
Higgsfield Genjutsu motion-transfer / object-swap video recreationPartlyAn existing reference video (real or AI) recreated with a different character (Motion Transfer) or a different held product (Object Swap), keeping all cuts, camera moves and backgrounds identical; can be driven manually or as an automated multi-variant ad recreation via ChatGPT + Higgsfield MCP.
Higgsfield GenjutsuUGC ad recreationproduct-swap ad recreation- 1Motion Transfer (manual) · Video · Higgsfield Genjutsu · HiggsfieldInsert a multi-cut reference video plus a generated character image into Genjutsu's Motion Transfer mode.
Prompt
Replace the man in the original video with the man attached as an image reference. Keep everything else in the video exactly the same.
- 2Object Swap (manual) · Video · Higgsfield Genjutsu · HiggsfieldInsert a reference video showing a held product plus a reference image of a different product, in Genjutsu's Object Swap mode.
Prompt
switch the product she is holding in the video to this new product here. Don't change anything else in the video.
- 3Connect Higgsfield to ChatGPT Desktop · By hand · Higgsfield MCPAdd the Higgsfield MCP plugin (higgsfield.ai/mcp) to ChatGPT Desktop (required over the browser version for this workflow).
- 4Break down a reference end-to-end · research · GPT-6 Astra (ChatGPT Desktop)Attach a breakdown skill file (a Drive-hosted "Miko video breakdown skill file") and a reference ad video, asking the model to analyze the video end-to-end by audio and every frame.
Prompt
read the breakdown skill file and analyze this video end-to-end... ignore the screen recording, ignore all buttons and text on the screen.
- 5Recreate and split-test, automated · Video · Higgsfield Genjutsu (via MCP) · Higgsfield MCP (called from ChatGPT Desktop)Attach a second video-editing skill file, reference the desired character, and ask the model to produce the recreated ad plus split-tested variations; it uses the Higgsfield MCP tools to generate everything from one prompt, no manual editor opened.
Missing: Both named "skill files" (breakdown and video-editing) are referenced as available in the presenter's community/Drive folder but their contents were not captured in this sweep.
The AI Content System · New UGC Video Breakdown (position 2)- 1
Transcript-driven chat edit of raw footageRunsa cut of existing raw footage (final.mp4), edited by describing the desired cut in plain English rather than reviewing frames
any cut from raw source footage- 1Start the agent on the footage · By hand · Claude CodePut raw clips in a folder and start the agent there.
- 2Transcribe · Check · not stated · ElevenLabsTranscribe every clip.
- 3Propose the cut · edit · Claude Code / video-useDescribe the desired edit in plain English; agent proposes a cut plan from the words, not the frames.
- 4Approve · By hand · manualWait for 'go ahead' approval before cutting.
- 5Cut · edit · ffmpegCut the approved plan; originals stay untouched.
- 6Self-check · Check · video-useSelf-check for jumps/pops before showing the result.
- 7Remember style · By hand · video-useSave style preferences to a memory file for next time.
AIOps · Video Use: Edit Raw Footage by Chatting With a Coding Agent- 1
Code-defined repeatable clips with local captionsRunsrepeatable motion-graphics clips and captioned video, rendered locally with no video model, credits or upload
Whisper (local, captions)repeatable branded clipcaptioned video- 1Install · By hand · RemotionInstall the 12-skill set plus a router skill that loads the matching one.
Prompt
npx skills add remotion-dev/skills
- 2Read the router · By hand · manualRead the router skill first, since skills run with the agent's full permissions.
- 3Write the video as code · edit · Claude Code / Remotion skillsInstruct the agent to write the video as code, explicitly saying 'do not render until told' (the router never auto-renders).
Prompt
do not render until told
- 4Captions · Check · Whisper (local) · Remotion captions skillRun local Whisper transcription for captions, free, no upload.
- 5Preview · Check · Remotion dev serverPreview in browser.
- 6Render · Video · RemotionRender an MP4 on your own machine; tokens are the only cost, maps/cloud rendering may need paid keys.
AIOps · Remotion Agent Skills: Let a Coding Agent Write and Render Videos as Code- 1
Upscaling (image + video finishing)RunsAn upscaled final image (2x/4x/8x/16x) or video (2K/4K, 24-120fps), as the last step before delivery.
Topaz- 1Upscale · Upscale · Topaz · HiggsfieldRoute through Higgsfield to Topaz; pick image (2x/4x/8x/16x) or video (2K/4K, 24-120fps) settings.
Ours ·- 1
Video Reference Recreation (ChatCut)RunsA real-footage re-edit (not generative) that rebuilds a reference video's shot rhythm, layout, captions, motion graphics, B-roll and music onto the brand's own raw footage.
UGC testimonialproblem→solutionlisticlekinetic typographyhard-cut compilation- 1Brief + import · By hand · ChatCutBrief (format/length/reference/raw clips/brand rules); import via ChatCut asset-import.
- 2Table the reference · edit · ChatCutFeed reference + raw clips into ChatCut preset a482ad98-63ef-47c0-8153-d18f65a2b2bb; tables the reference's shots/rhythm/layout/captions/motion-graphics/B-roll/transitions/music into reference.md.
- 3Rebuild on brand footage · edit · ChatCutStructure-first build order rebuilds the same rhythm on the brand's own footage.
- 4Verify · CheckVerify with ffprobe/loudness checks; export only when asked.
Ours ·- 1
Recreate a Video (ChatCut sister preset)PartlyA video recreated from a reference in one of three modes: style adaptation/free, limited edit, or multi-reference generation.
- 1Pick mode · edit · ChatCut3 modes: style adaptation/free, limited edit, multi-reference generation (costs Seedance credits).
Missing: Modes named, no worked example.
Ours ·- 1
4K 60fps video finishing passRunstakes an existing AI-generated video clip (e.g. a 720p 30fps Veo3 output) and finishes it into a 4K, 60fps clip with the original audio re-attached.
Topaz (video upscale)RIFE (frame interpolation)FFmpeg (compose)4K video60fps video- 1Set creative direction · By hand · n8nUser sets a resolution multiplier (e.g. 3x for a 720p source, worked out by asking ChatGPT) and a frames-to-interpolate count (e.g. 1, which doubles fps).
- 2Load input video · Words · Google SheetsFinds a Google Sheet row with status 'for production' and reads its input video URL.
- 3Upscale to 4K · Upscale · Topaz (fal-ai/topaz/upscale/video) · fal.aiHTTP POST to fal.ai's Topaz upscale/video model with the input video URL, upscale factor and a target FPS (set to 30 to prep for interpolation).
- 4Wait/check upscale · Check · fal.aiWaits 60s, polls result, loops back to wait if the video URL isn't ready yet.
- 5Interpolate frames · Video · RIFE (fal-ai/rife/video) · fal.aiHTTP POST to fal.ai's RIFE model with the upscaled (silent-capable) video URL and the frames-to-interpolate count, with 'use calculated FPS' enabled; strips audio as a side effect.
- 6Wait/check interpolation · Check · fal.aiWaits then polls result, looping back if not ready.
- 7Re-add audio · edit · FFmpeg (fal-ai/ffmpeg-api/compose) · fal.aiHTTP POST to fal.ai's FFmpeg 'compose' function layers the silent 4K/60fps video with the original clip's audio track, both starting at timestamp 0.
- 8Wait/check compose · Check · fal.aiWaits then polls until the final composed video URL is ready.
- 9Log final output · Words · Google SheetsUpdates the sheet row's status to done and stores the upscaled and final upscaled+interpolated video URLs.
RoboNuggets · n13 - Stop making Low-Res AI Videos, do 4k Instead- 1
Hyperframes video edit + motion-graphics finish (BIT framework)PartlyA polished short promo/intro video: raw talking-head footage auto-cut of silences and mistakes, given animated floating title cards synced to speech, subtitles, and composited 3D motion-graphic assets (a rotating 3D logo, a wireframe globe), all rendered as code (HTML/CSS) via the open-source Hyperframes tool.
Hyperframes (HTML/CSS-based video renderer)FFmpegfast WhisperNano Banana 2Kling 3.0promo/intro videopodcast introwebsite-to-video short (10s, from a URL)title-card motion graphics overlay- 1Install Hyperframes Helper skill · By handGive Claude Code the install prompt to clone robonuggets/hyperframes-helper, which teaches the agent how to install and use Hyperframes (storyboarding, composing, rendering).
Prompt
I'm giving you a skill kit called Hyperframes Helper. Get the files with: git clone https://github.com/robonuggets/hyperframes-helper Follow the README to set it up. Recommend how it would best apply to our setup in plain language, and ask me a few questions to clarify my intent before you install it.
- 2Website-to-video (quick draft) · Video · HyperframesGive Claude a website link and ask it to use Hyperframes' website-to-video feature to scrape content/photos/brand palette and assemble a short video draft; useful as a rough first pass, needs guidance for polish.
Prompt
use Hyperframes website to video to turn this into a 10-second video.
- 3Cut raw footage (silences, dead air, mistakes) · editAsk Claude to trim silences/pauses/mistakes from raw footage using FFmpeg and fast Whisper, applying a 'last take' rule for repeated phrases, and produce a side-by-side HTML comparison of the original vs. proposed clean script with word-level timestamps.
Prompt
edit my raw recording into a clean final cut ready for hyperframes. My only instructions were to cut all the silences and the dead air out, and also transcribe the audio so we word-level timestamps. I asked to use the last take rule, so if I say a phrase twice, then keep the second one and cut the first. And also to build me this HTML page that shows the current script versus the proposed clean script side by side.
- 4Storyboard title cards · motion-copyAsk Claude to build a one-page HTML storyboard mocking up planned title-card motion: part one full-bleed video with subtitles and floating cards, part two collapsing into a corner with a motion-graphics center stage, styled to the brand design system.
Prompt
build me a one-page HTML storyboard that visually shows my plan for the video. My plan is two parts: part one, when I talk about floating cards and subtitles, keep the video full bleed, show subtitles, and sample cards on the right at around 80% opacity, then transition into part two, where the video collapses into a vertical rounded rectangle at the bottom right. Fill the center stage with placeholder boxes which I will add in later, using our Robo Labs design system for colors.
- 5Refine storyboard · motion-copyGive feedback to tighten the storyboard: one card animating per beat (enter, hold, exit), a more 'liquid glass' style, relevant logos added, and card text repeating the spoken line in 2-6 word phrases.
Prompt
edit that storyboard, what I want is only one card animating per beat, where the card enters, holds while I say the line, and then exits, then the next one comes in. Also, I want the cards to be more liquid glassy in style, and add the relevant logos. Also the card text should just repeat what I'm saying word for word with two to six-ish word phrases.
- 6Render in Hyperframes Studio · edit · HyperframesOpen the built-in Hyperframes Studio (a local timeline UI on localhost) to apply the storyboarded card motion graphics and subtitles to the cut video and preview the result before final render.
Prompt
open the hyperframes studio and apply those slide card motion graphics to our cut video.
- 7Generate rotating 3D logo asset · Picture · Nano Banana 2Generate a green-screen image of the brand logo with Nano Banana 2, then animate it into a rotating 3D logo with Kling 3.0; Hyperframes applies color-keying to remove the green screen when compositing.
- 7Animate the 3D logo · Video · Kling 3.0Animate the green-screen logo image into a rotating 3D logo clip via Kling 3.0, for later color-keyed compositing into the main video.
- 8Composite motion graphics (wireframe globe, icons, 3D logo) · edit · HyperframesCopy an HTML/CSS component (e.g. a wireframe globe) from 21st.dev, give it to Claude with instructions on where in the timeline to place it, and composite it alongside the 3D logo and brand icon assets into the final Hyperframes video.
Prompt
get the wireframe globe from 21st.dev, put it in the center, have it move to the left, and then add in a text in there.
- 9Calibrate / bank learnings into the skill · CheckAt session end, ask Claude to capture the session's learnings and feedback and fold them into the Hyperframes Helper skill (the 'Integrate' and 'Tune' steps of the BIT — Build, Integrate, Tune — framework), so future runs already know what 'good' looks like.
Prompt
I want to capture all of the learnings in this session and update the hyperframes helper skill based on my feedback to you and all the best practices that you captured based on our conversation.
Missing: Needs Hyperframes + the Hyperframes Helper skill installed, FFmpeg and fast-whisper available locally, and access to Nano Banana 2 and Kling 3.0 (provider for these two not named in this lesson) for the 3D logo step; 21st.dev component prompts are copied manually per use.
RoboNuggets · n51 - Hyperframes + Claude- 1
Sora watermark removerManuala Sora-generated video with its watermark removed, via an n8n automation.
watermark-free video- 1Run watermark removal workflow · editAn n8n workflow (template only, no lesson text or transcript) that takes a Sora-generated video and strips its watermark.
Missing: this lesson has no lesson text, video or transcript, only an n8n workflow JSON resource file that was not opened; no model, provider or method could be confirmed beyond the title.
RoboNuggets · b1 - Sora Watermark Remover- 1
Infinite short clips from longform video (Klap)Runsmultiple short vertical clips (with captions, reframing and emoji overlays) auto-extracted from one longform video and published to multiple social channels
Klap (video-to-shorts analysis + export)short-form clip extraction from longform source- 1Get longform video input · edit · Google SheetsPull the next longform video URL to process from a Google Sheets tracker.
- 2Analyze longform video (Klap) · Video · Klap video-to-shorts analyzer · Klap.app (/v2/tasks/video-to-shorts)POST the longform video to Klap's video-to-shorts task to analyze it for short-clip candidates, with editing_options set for captions, reframe and emojis and to remove silences.
- 3Poll for analysis completion · Check · Klap.app (get status / get shorts details)Wait/poll the task status until the analysis is ready, then get the list of candidate shorts.
- 4Export shorts · Video · Klap export · Klap.app (export endpoint, preset_id)For each candidate short, POST an export request using a preset_id to render the final captioned/reframed clip.
- 5Poll for export completion and get final shorts · Check · Klap.app (get shorts)Wait/poll each export task until ready, then GET the final rendered short clip URLs.
- 6Log clips to sheet · edit · Google SheetsLog the finished clip URLs back to the Google Sheets tracker.
- 7Publish clips to multiple channels · publish · BlotatoEach clip is published per social channel through Blotato, batched roughly one request per 60 seconds to avoid rate limits.
RoboNuggets · R32 | The Infinite Clips AI Machine- 1
AI vertical reframe system (Lumalabs)Runsa 9:16 vertical version of a landscape source video, reframed by AI (not just cropped) in 29-second chunks and stitched back together with a black-bar transition animation
Lumalabs Dream Machine reframe (ray-flash-2)ffmpeg (via Rendi.dev)horizontal-to-vertical AI reframe- 1Optional: scrape a source YouTube video · research · Apify (YouTube scraper + YouTube video downloader actors)Optionally use an Apify YouTube scraper actor to get video details, then an Apify YouTube downloader actor to fetch the actual video file.
- 2Chunk into 29-second timestamps · edit · n8n (Code node)A JS code node calculates timestamp ranges to split the source video into consecutive 29-second chunks (Lumalabs' reframe API caps clip length at 30s).
- 3Split video into chunks (ffmpeg via Rendi) · edit · ffmpeg · Rendi.dev (/v1/run-ffmpeg-command)For each timestamp pair, run an ffmpeg split command via Rendi.dev's FFmpeg-as-a-service to cut out that 29-second segment.
- 4Poll for split completion · Check · Rendi.devWait/poll Rendi.dev until each split job is ready, then GET the resulting segment URLs.
- 5AI-reframe each chunk to 9:16 · Video · Lumalabs Dream Machine reframe, model ray-flash-2 · Lumalabs (/dream-machine/v1/generations/video/reframe)Each 29-second segment is sent to Lumalabs Dream Machine's reframe API to be intelligently reframed from landscape into 9:16 vertical (tracking subjects/action, not a plain crop).
- 6Poll for reframe completion · Check · LumalabsWait/poll Lumalabs until each reframe job is ready, then GET the resulting reframed segment URLs.
- 7Build stitch command · edit · n8n (Code node)A JS code node generates an ffmpeg concat command that stitches the reframed segments back together in order, adding a black-bar transition animation between chunks.
- 8Stitch final vertical video · edit · ffmpeg · Rendi.devRun the generated ffmpeg concat command via Rendi.dev to produce the final single stitched 9:16 video.
- 9Log final video · edit · Google SheetsLog the final video URL to Google Sheets.
Missing: exact JS code (chunking + stitch-command builder) and exact ffmpeg/Lumalabs/Apify request bodies were viewed in the source lesson during an earlier read but were not retained verbatim into this rewrite pass; re-open knowledge/courses/robonuggets-school/all-lessons-and-templates/r33-the-ai-vertical-reframe-system--873d7b96.md directly to recover the exact code/bodies before building.
RoboNuggets · R33 | The AI Vertical Reframe System- 1
Other42
Winning-ad concept script adaptationRunsan adapted ad script plus 10 ad hooks (text only), built by feeding a proven competitor ad's transcript and a brand research document into Poppy
Poppyscript/hooks- 1Source a winning ad transcript · research · Foreplay / PoppyFind a winning ad from the course swipe file or via Foreplay (transcript section), or paste a TikTok/Instagram link directly into Poppy and ask its chat to confirm the transcript.
Prompt
What is the transcript of this video?
- 2Build the research document · researchPaste each sourced transcript under its own 'Concept N' heading with concept-type metadata (e.g. 'founder type ad'), alongside 3 key personas, into one document; export as PDF.
- 3Adapt the script · Script · Poppy · PoppyDrag the PDF into a Poppy project, open chat, paste the script-adaptation prompt naming the concept number and chosen avatar. Review for product accuracy, persona's exact phrases, no hallucination.
Prompt
You are an expert direct response copywriter who has generated over a hundred million in sales with Facebook ads and scripted thousands of UGC ad creatives. Analyze the structure, flow, and length of concept number [X]. In the concept document, adapt it to our product/brand and use categorized insights from the research document. Make sure the insights are relevant to [avatar — e.g. avatar three from the key avatar section, the skeptical first-time owner]. Explain your reasoning and go.
- 4Generate hooks · Script · Poppy · PoppyIn the same chat, ask for 10 ad hooks with different emotional/humorous/problem-solving approaches, using common pain points from the research document.
Prompt
Now knowing what you know, come up with ten ad hooks for this brand. Think of different emotional, humorous, or problem-solving approaches. Use common pain points listed in the research document to create the hooks. Ensure each angle is different. Ensure each angle captures the attention of a buyer persona who does not know or care about the brand.
AI Ads (Ad Creators Lab) · Creating winning ad concepts with AI- 1
Triple G scriptwriting framework (Goal, Gap, Gains)Runsa script structure: not itself a generated asset, but the reusable script method that feeds any UGC/video workflow's script step
script framework- 1Goal · ScriptSet up the viewer's goal so they keep watching. Verbatim framing: 'you wanna set up the goal and be like, "Hey, you know, if you have this goal, basically keep watching."'
- 2Gap · ScriptName the gap between what someone wants and what they're currently doing that isn't working (e.g. working out isn't enough without diet/protein).
- 3Gains · ScriptState the benefits — how they get closer to the goal and close the gap.
AI Ads (Ad Creators Lab) · Triple G Framework - Scriptwriting- 1
Transformation trifecta scriptwriting frameworkRunsa script-angle-staggering method built on the Value Equation (increase dream outcome, increase perceived likelihood, reduce time delay, reduce effort), staggering ads by perceived viewer effort
script framework- 1Level 1 effort · ScriptPresent the product as the fastest immediate solution, like a Google search of the best tool. Verbatim: 'a Google search of the best tool to quit vaping... present your product as an immediate solution... the fastest to get the result.'
- 2Level 2 effort · ScriptCall out the problem directly and focus on the dream outcome. Verbatim: 'Are you experiencing this? Do you have this problem? The solution, it's that, or let's change that this year.'
- 3Level 3 effort · ScriptHelp build a habit rather than solve an immediate problem, for people ready to start a change. Example: 'make hydration a habit.'
AI Ads (Ad Creators Lab) · Transformation Trifecta - Scriptwriting- 1
Curiosity loop scriptwriting frameworkRunsa stacked-curiosity-loop script structure for high-converting UGC video hooks, with a seasonal/gifting applied variant
script framework- 1Clarify the emotion · ScriptOpen by naming the emotional state the viewer is in (baseline low-converting version is just 'hook, problem, solution, payoff/CTA' — the improved version stacks loops instead).
Prompt
So if your skin is breaking out, don't get upset. I'm about to let you in some detox secrets so you don't get overwhelmed looking for face creams. All you need to do is take this powder...
- 2Open loop 1 · ScriptTease the solution without naming it.
- 3Open loop 2 · ScriptShow the solution while still withholding key information.
- 4Open loop 3 · ScriptReveal what the solution does, still holding back the final detail.
Prompt
You have until the age of 40 to change this one habit if you want to live longer... You don't need a whole new diet or lifestyle change. You just need the one micronutrient that 60% of the U.S. population is deficient in, and that's fiber.
- 5Payoff · ScriptReveal the product/brand.
AI Ads (Ad Creators Lab) · Mastering curiosity loops for high converting UGC videos / Valentine's Curiosity Loop - Scriptwriting- 1
CURE listicle scriptwriting frameworkRunsa listicle-style ad script structure (Curiosity, Uncover the failure, Reveal the solution, Evidence+CTA)
script framework- 1Curiosity (hook) · ScriptOpen with curiosity, not a generic list count. Bad: 'Here are three reasons why.' Better: 'Three reasons your boyfriend hates snuggle time.'
- 2Uncover the failure · ScriptCall out current solution flaws. Examples: 'melatonin supplements that disrupt natural hormone balance cycles', 'white noise machines that drown out vital environmental alerts.'
- 3Reveal the solution · ScriptIntroduce your solution as the answer. Example: 'NeuroRestore uses chronobiotic amino acids.'
- 4Evidence + CTA · ScriptGive proof and a call to action. Example: 'eighty percent of users have had this.' CTA example: 'try risk-free for thirty nights.'
AI Ads (Ad Creators Lab) · Mastering listicle formats for engagement- 1
UGC brief template with 8-beat script arcRunsa UGC creator/editor brief document — a reusable template (not a worked example) with a named 8-beat emotional script arc and a B-roll shot list
brief template- 1Concept and hooks table · ScriptState the concept, then list 5 hooks in a table.
- 28-beat talking-head script arc · ScriptStructure the body script through 8 named beats: The Realization, The Time Drain, The Avoidance Pattern, The Discovery, The Time Reclaimed, The Frequency Shift, The Real Benefit, The Invitation.
- 3B-roll shot list · ScriptList 10 B-roll shot slots to brief a creator/editor, plus filming DO/DON'T guidelines and delivery instructions.
- 4Video editor brief · ScriptSecond half of the document: editing principles plus a 3-hook table (Audio Hook / Caption Hook / B-roll Guide) to brief the editor.
AI Ads (Ad Creators Lab) · Swipe File / scriptwriting assets- 1
Seedance 2.0 facial-recognition-filter bypass (black-circle method)PartlyA real person's likeness used as a Seedance 2.0 video reference despite the model's built-in face-detection filter, by splitting the reference into a masked body image and a separate unmasked face cutout.
Seedance 2.0any Seedance format- 1Prepare the starting frame · PictureSet a photo of the person/character to a 9x16 canvas.
- 2Mask the face · editDraw a black circle over the face on the 9x16 image so the facial-recognition filter does not trigger.
- 3Cut out the face · editOn a second copy of the same photo, erase everything except the face, leaving a transparent-background face cutout.
- 4Attach both references and prompt by role · Video · Seedance 2.0 · Higgsfield / Dreamina / Kie.aiAttach the masked body PNG and the face cutout PNG to Seedance 2.0 as two separate reference images, then reference each by role in the prompt before giving the action/dialogue instruction.
Prompt
this is the body I want you to use for the person [reference masked/body image]; this is the face I want you to use of the person [reference face-cutout image]; create me a video of her talking about how her work day went.
Missing: The Canva masking and PNG export steps are manual desktop-editor actions with no API shown; only the final Seedance submission is an API/tool call.
The AI Content System · Seedance 2.0 Bypass Method- 1
One-agent image/video/voice/music generation (MiniMax)Runsan image, a 6 or 10-second video (text or image-to-video), cloned/40-language speech, or a music track, all from one MCP server
image-01T2V / I2V / S2V / Hailuo-02speech-01music-1.5imageT2V/I2V/S2V product demo videocloned/multilingual voiceovermusic track- 1Install and connect · By hand · MiniMax MCPInstall UV, add the MiniMax MCP with API key + matching API host (key region must match host or auth fails).
Prompt
claude mcp add minimax-mcp -e MINIMAX_API_KEY=your-api-key -e MINIMAX_API_HOST=https://api.minimax.chat -- uvx minimax-mcp
- 2Generate an image · Picture · image-01 · MiniMaxSynchronous image generation from a text prompt.
Prompt
Generate an image of a sunset over mountains using MiniMax.
- 3Generate video · Video · T2V / I2V / S2V / Hailuo-02 · MiniMaxAsynchronous video generation; submit then poll the task endpoint with exponential backoff.
Prompt
Use MiniMax to create a 6-second product demo video showing a drag-and-drop interface in action. Use 1080p resolution.
- 4Generate speech / clone a voice · Voice · speech-01 · MiniMaxSynchronous TTS, 40 languages, clone from a 10s+ sample; cap requests under 10,000 characters, chunk and stitch longer scripts.
Prompt
Use MiniMax voice cloning with this reference audio file: /path/to/my-voice-sample.mp3. Then generate speech for: "This is a product update from the team."
- 5Generate music · music · music-1.5 · MiniMaxGenerate a music track from a text brief.
- 6Optional skill reference · By hand · manualClone the MiniMax skills repo and drop a skill markdown file (e.g. fullstack-development.md) into the project or CLAUDE.md for the full API-usage discipline (async polling, chunking, build-time vs runtime generation, caching, fallback content).
AIOps · MiniMax Media Skills: Image, Video, Audio, Music From One Agent- 1
Headless local ComfyUI video renderRunsa self-hosted, zero-marginal-cost video (MP4), run entirely on local hardware via the ComfyUI API, no visual canvas required once set up
Wan2.2LTX-2.3general video (Wan2.2)audio-synced video (LTX-2.3)- 1Install and launch · By hand · ComfyUIpip install comfy-cli then comfy install then comfy launch (server at 127.0.0.1:8188).
Prompt
pip install comfy-cli comfy install comfy launch
- 2Export a workflow template · By hand · ComfyUI web UILoad a video template in the browser (e.g. Wan2.2 5B), enable Dev Mode, export the workflow in API format as JSON.
- 3Install model dependencies · By hand · comfy-cliAuto-download the (multi-GB) models and custom nodes the workflow needs.
Prompt
comfy node install-deps --workflow=workflow.json
- 4Render via API · Video · Wan2.2 (general video) or LTX-2.3 (audio-synced) · ComfyUI local APIPOST the API-format JSON to /prompt with a description, resolution and duration.
Prompt
curl -X POST http://127.0.0.1:8188/prompt -H "Content-Type: application/json" -d @workflow_api.json
- 5Agent-driven end-to-end · By hand · as step 4 · Claude CodeHand the full sequence to Claude Code as a fill-in-the-blank prompt with model/description/resolution/duration placeholders so the agent drives install-to-render end to end.
Prompt
Set up a ComfyUI video generation pipeline on this machine. 1. Install comfy-cli if not already installed: pip install comfy-cli 2. Install ComfyUI: comfy install 3. Launch ComfyUI: comfy launch 4. Load the [MODEL_PREFERENCE] video generation template 5. Modify the workflow to use these settings: - Prompt: "[VIDEO_DESCRIPTION]" - Resolution: [RESOLUTION] - Duration: [DURATION] seconds 6. Install any missing model dependencies: comfy node install-deps --workflow=workflow.json 7. Save the workflow in API format for future reuse 8. Run the workflow via the API 9. Tell me where the output video file was saved If any model downloads are needed, let them finish before running the workflow. Show me the file path when the video is ready.
Missing: hardware-gated: 1080p needs a 24GB GPU, which the current Mac fleet cannot host
AIOps · ComfyUI Video Pipeline- 1
Monthly content calendar to approved social postsPartlya month of per-platform social posts (images, short videos, captions) from a client's content calendar, gated by price approval and per-post human review
per-platform social image postshort social video post- 1Read the calendar · By hand · SocialForge (add-on for Claude Code, Cowork, Hermes, or OpenClaw)Read a client's monthly content calendar.
- 2Read the brand profile fresh · By hand · SocialForgeRead the brand profile fresh on every post so a month-long run doesn't drift.
- 3Source real photos · By hand · SocialForgeIndex and drop in real brand photos untouched, never repainted.
- 4Generate images · Picture · not stated · Google Vertex AIGenerate images per post brief and brand context.
- 5Generate short videos · Video · not stated · WaveSpeed (billed per second)Generate short videos per post brief.
- 6Price and approve · Check · SocialForgePrice the full month's run and wait for a yes before generating; start with one tier/week to catch a wrong brand profile cheaply.
- 7Review gallery · Check · SocialForge UIOne card per post: approve, revise, swap asset, or regenerate.
- 8Finalize · publish · SocialForgeRefuses to deliver any unapproved post.
Missing: the pipeline and its three guardrails are specified; API-key setup detail and exact generation prompts are not captured
AIOps · SocialForge: Content Calendar to Approved Social Posts- 1
Shorts QC and iteration desk: thumbnail + hookManuala stronger thumbnail image and matching hook-text options for one already-drafted short video, approved before posting
short-form video thumbnail + hook pairing- 1Review the draft · Check · not stated · not statedReview a draft short video and flag its weakest seconds.
- 2Generate thumbnail options · Picture · not stated · not statedGenerate stronger thumbnail options.
- 3Generate hook text options · Words · not stated · not statedGenerate hook text options.
- 4Human approval · Check · manualHuman approves before posting.
Missing: no models, tools, or prompts given; this lesson is a menu/overview of six 'desks', naming the job, not the method
AIOps · Organic Social: Channel-Fed Content Desks- 1
Text-to-video clip via Kling 2.6 (KIE.AI)Runsa short video clip generated directly from a text prompt
Kling 2.6text-to-video clip- 1Set parameters · Words · n8nA Set node ('Set Text to Video Parameters') stores the prompt, duration, aspect ratio and sound flag.
Prompt
The prompt field's shipped default value is the literal placeholder text: "YOUR_VIDEO_PROMPT_DESCRIPTION" (confirmed in the real template JSON — this is not a real example prompt, it is a fill-in instruction).
- 2Submit generation job · Video · Kling 2.6 (kling-2.6/text-to-video) · KIE.AISends the parameters to KIE.AI's kling-2.6/text-to-video model endpoint using Bearer auth.
- 3Poll status · Check · KIE.AIWaits, then repeatedly polls the KIE.AI job-status API and branches on queuing/generating/complete/failed until the job resolves.
- 4Download video · edit · KIE.AIExtracts the generated video file URL from the completed job response and downloads it locally.
Missing: confirmed by the real template JSON: the prompt field ships only the literal placeholder "YOUR_VIDEO_PROMPT_DESCRIPTION", no worked example prompt exists in this template
Free online · Generate text-to-video and image-to-video clips with Kling 2.6 via KIE.AI- 1
Image-to-video clip via Kling 2.6 (KIE.AI)Runsa short video clip generated by animating a supplied still image with a motion prompt
Kling 2.6image-to-video clip- 1Set prompt & image URL · Words · n8nA Set node ('Set Prompt & Image Url') stores the motion prompt, a publicly accessible source image URL, duration and sound flag.
Prompt
The prompt and image_urls fields ship with literal placeholder values: prompt = "YOUR_VIDEO_PROMPT_DESCRIPTION", image_urls = "YOUR_IMAGE_URL" (confirmed in the real template JSON — not real examples).
- 2Submit generation job · Video · Kling 2.6 (kling-2.6/image-to-video) · KIE.AISends the parameters to KIE.AI's kling-2.6/image-to-video model endpoint using Bearer auth.
- 3Poll status · Check · KIE.AIIdentical wait/poll/status-branch loop as the text-to-video workflow, until the job completes.
- 4Download video · edit · KIE.AIRetrieves the generated video URL and downloads the final file.
Missing: confirmed by the real template JSON: prompt/image_urls ship only literal placeholders ("YOUR_VIDEO_PROMPT_DESCRIPTION", "YOUR_IMAGE_URL")
Free online · Generate text-to-video and image-to-video clips with Kling 2.6 via KIE.AI- 1
Reference clip breakdown + brand-adapted scriptRunsA structured markdown breakdown of a proven-winning reference video plus a brand-adapted, scene-split (≤10s each) script and 4-6 hook variations.
gemini-2.5-flashUGC testimonialstreet interviewday-in-the-lifepodcast clipauthority explainerbefore-after demo- 1Load brand context · WordsloadBrandContext pulls the brand's real product/ICP facts so the script never borrows the reference clip's product category.
- 2Analyze + rewrite · Video · gemini-2.5-flash · GoogleGemini watches the attached reference clip and reverse-engineers it into a breakdown, a brand-adapted script split into ≤10s scenes, and hook variations.
Prompt
You are the UGC Analyzer. Watch the attached ${clip.platform} video — a PROVEN-WINNING short-form ad reference — and reverse-engineer it into a replicable ad for the brand "${brand}", following the course method: model winners, don't invent; the hook is everything; iterate the winning formula, don't copy it verbatim. ========== BRAND CONTEXT — THIS is what "${brand}" actually is (source of truth) ========== ${brandContext} ========== END BRAND CONTEXT ========== CRITICAL: The reference video is ONLY a structural template (its hook shape, pacing, scene flow, performance style). The PRODUCT, the buyer, the claims, and the language of YOUR script come ONLY from the BRAND CONTEXT above — NEVER from the reference's product category. If the reference sells skincare but ${brand} is something else, your script sells ${brand}'s real product, not skincare. Pick ONE persona and install ONE belief from the brand context; speak in the customer's real voice from it. Reference context (template only): platform ${clip.platform}, ~${duration}s, engagement: ${eng}. Reference caption: "${clip.caption}" OPERATOR STEER (obey this for angle / ICP / vibe / hook direction, within the brand context): "${instruction}" Produce: 1) A full BREAKDOWN of why this reference works: transcribe the spoken script; describe the camerawork, framing, lighting, the on-camera performance, the pacing, and the single core message. Plain English, markdown. 2) A ready-to-use SCRIPT selling ${brand}'s REAL product (from the BRAND CONTEXT above — never the reference's product category), adapted to the operator steer, broken into SCENES of AT MOST 10 SECONDS each (a hard cap — Omni generates one ≤10s clip per scene). A typical UGC ad is ~3 scenes (HOOK → BODY → CLOSE) but use as many ≤10s scenes as the idea needs. Front-load the dialogue; the first 3 seconds are the whole ballgame. Write it spoken-style, conversational, in the buyer's own words — never an actor reading an ad. 3) 4–6 HOOK VARIATIONS to test — iterate the proven hook's formula, each a different angle/emotion. Do not just reword; vary the lead (problem, result, curiosity, social proof, urgency). Return STRICT JSON only, no prose, no code fences, shaped exactly: {"concept_type": "<the video style, e.g. Customer UGC Testimonial / Street Interview / Day-in-the-Life / Podcast Clip / Authority Explainer / Before-After Demo>", "angle": "<the buyer (ICP) + the belief/angle this installs, one line>", "analysis_md": "<the full breakdown, markdown>", "scenes": [{"n": 1, "label": "HOOK", "time_range": "0–10s", "on_screen": "<what's visible + the action>", "dialogue": "<the spoken line / VO>", "purpose": "<one line: why this scene>"}], "hooks": ["<hook variation 1>", "<hook variation 2>", "<hook variation 3>", "<hook variation 4>"]} Every scene's time_range must be within a single 10-second window. Output ONLY the JSON object.
Ours ·- 1
Reference clip sourcing (Source & Swipe)RunsA scored, kept/discarded set of downloaded reference clips pulled from TikTok/Instagram for a brand to model.
gemini-2.5-flash- 1Search · research · ApifyhandleSourceSearch → pullReferences → consumer/source.mjs handlePull calls the Apify actors apify~instagram-scraper and clockworks~tiktok-scraper.
- 2Score · Check · gemini-2.5-flash · GoogleGemini 2.5 Flash scores each candidate for reference quality.
- 3Download + store · By handdownloadAndStore saves the kept clips; operator presses keep/discard.
Missing: Facebook Ad Library sourcing is deliberately unbuilt (source.mjs:171)
Ours ·- 1
Google Flow hand-off assembly (manual step)ManualA set of script-card PNGs plus a fixed instruction string, assembled so an operator can manually run Google Flow (Frames-to-Video) using the pipeline's own start frames and scripts — no generation call itself.
- 1Render script cards · PicturerenderCardPng draws each scene's script as an SVG→PNG card.
- 2Fixed instruction · Wordsflow.mjs writes a fixed instruction string per scene for the operator to type into Google Flow.
Prompt
read [script card] use [start frame]
- 3Open Flow · By hand · Veo (via Google Flow) · Google Flowopen-flow/route.ts opens labs.google/fx/tools/flow in Opera (macOS only) for the operator to run Frames-to-Video by hand.
Missing: Per the file's own header comment (flow.mjs:2-27): bringing finished Flow clips back into the pipeline is 'not yet built — next slice'; no final stitch/edit stage exists.
Ours ·- 1
Ads dashboard creative pipeline (n8n hand-off)PartlyStatic image, video, or AI UGC batch, brief submitted from the ads dashboard and handed to an external n8n workflow that does the actual generation.
Kling ProKling O3Seedance v1.5 ProFLUX KontextSora 2 (hero only)- 1Submit brief · Words · picker: Kling Pro / Kling O3 / Seedance v1.5 ProOperator fills a brief and picks a model in the UI.
- 2Hand off to n8n · Video · n8n webhook https://n8n.icarusdigitalmarketing.com/webhook/creative-pipeline-v2POST to an n8n webhook; all real generation happens outside this repo.
Missing: Trigger/hand-off is wired; the actual generation logic lives in n8n, not inspectable from this repo.
Ours ·- 1
Sora 2 Character Cameo (reusable @tag)PartlyA reusable character asset registered in Sora 2 as a permanent @tag, so future prompts can reuse the exact same identity.
Sora 2- 1Describe character · WordsDescribe the character via 6 micro-detail categories: eyes, nose, contours, lips, skin, subtle features.
- 2Generate base video · Video · Sora 2 · Sora.comGenerate a base video on a plain background from the description.
- 3Create Character · By hand · Sora.com3-dot menu → 'Create Character' → trim to the clearest section → saved with a permanent @tag; include the @tag in future prompts.
Missing: Character-description prompt is conceptual, not literal.
Ours ·- 1
Seedance 2.0 — Prompt Formula LibraryManualA reference library of fill-in prompt templates covering product showcase, beauty/supplement, app/SaaS, retro, POV UGC, testimonial, TikTok sketch, MrBeast-style, B-roll and before/after content, all in the same universal Seedance formula.
Seedance 2.0product showcasebeauty/supplementapp/SaaSretroPOV UGCtestimonialTikTok sketchB-rollbefore/after- 1Fill the universal formula · Words · Seedance 2.0Universal formula and per-content-type fill-in templates, plus a time-blocked multi-beat template (00-05s/05-10s/10-15s) and character-consistency reference-image patterns.
Prompt
[Subject]+[Action]+[Scene]+[Lighting]+[Camera]+[Style]+[Quality]+[Constraints]
Missing: Pure prompt reference, no platform/account/UI steps given.
Ours ·- 1
Seedance 2.0 — Standalone prompting skillPartlyA prompt-engineering reference: a 6-layer core formula, a 7-element cinematic Image/Video Prompt split, a ban-word list, @-syntax reference binding, and a 3-angle character-consistency sheet protocol.
Seedance 2.0- 1Apply the prompting rules · Words · Seedance 2.06-layer core formula + 7-element cinematic Image/Video Prompt split; ban-word list with replacements; mandatory literal quality suffix; @-syntax reference binding (12 slots); 3-angle character-consistency sheet protocol with a literal preservation-keywords block; time-coded multi-shot template (≤15s).
Prompt
sharp clarity, natural colors, stable picture, no blur, no ghosting, no flickering
Missing: Complete for prompting; platform routing described only, not literal.
Ours ·- 1
AI-Influencer network — Pipeline B character setup (one-time)RunsA one-time, fully-specified AI influencer character: a 7-layer persona blueprint, a locked face seed, and a trained LoRA.
Nano Banana Pro- 1Persona blueprint · Words7-Layer Persona Blueprint: Bio/Lore/Topics/Adjectives/Style/Examples/Voice.
- 2Face Seed · Picture · Nano Banana ProGenerate the face seed with a literal JSON schema and a literal Face-Lock Phrase and character-sheet prompt.
- 3LoRA training · Picture · fal.aiTrain a LoRA on fal.ai from the face seed.
- 4Recurring environment anchor · PictureEstablish a recurring environment anchor for consistency.
Ours ·- 1
UGC Video Analyzer Gem (video → reusable Sora 2 prompt)ManualA ready-to-use Sora 2 prompt plus 10 hook variations, extracted from an analyzed reference video.
Gemini- 1Analyze · Video · GeminiBuild a Gemini Gem from a course Master Prompt + Knowledge File (both not included in this file), then trigger with the literal prompt.
Prompt
Analyze this video in depth and make it into a Sora 2 prompt using all the knowledge you have
Missing: Core Gem instructions (Master Prompt + Knowledge File) are missing from these files.
Ours ·- 1
Reverse-Engineering Ads with Gemini (script extraction/rewrite)RunsA rewritten ad script, extracted and adapted from a reference ad via two literal Gemini prompt methods.
Gemini- 1Method 1: strategic breakdown · Video · GeminiLiteral Gemini prompt for a strategic breakdown of a reference ad.
- 2Method 2: script extraction + rewrite · Video · GeminiLiteral Gemini prompt for script extraction and rewrite.
Ours ·- 1
Universal Prompt-Writing / JSON Prompt SystemRunsA structured image/video prompt (Subject→Context→Action→Style→Constraints), and the JSON-round-trip technique used throughout the Nano Banana/medical/3D workflows.
Claude- 1Image to JSON · Words · ClaudeConvert a reference image to a structured JSON prompt via Claude.
- 2JSON to flat text · Words · ClaudeFlatten the JSON back to a plain-text prompt for the target model.
Ours ·- 1
Ad Script Writing FrameworksRunsAn ad script written in one of three frameworks (Feel-Felt-Found, PAS-A, Story-Sell).
ClaudeGPT- 1Generate script · Script · Claude / GPT3 frameworks (Feel-Felt-Found, PAS-A, Story-Sell) plus a literal 'Start Prompt' quoted in full, reused verbatim across multiple files.
Ours ·- 1
MUAPI Gateway — 6 executable API recipesRunsSix sub-recipes via one gateway: product-in-scene edit, label-swap edit, infographic/diagram generation, image-to-video, lipsync, and music/SFX generation.
gpt-image-2kling-prolatent/LatentSyncsync/Sync LabsSuno V5- 1Edit-mode product-in-scene · Picture · gpt-image-2 · muAPIPRESERVE-clause pattern, gpt-image-2 default for VERO — verbatim WRONG/CORRECT prompt examples, bash/python/CLI call paths, async polling, mandatory single smoke-test before mass regeneration.
- 2Label-swap edit · Picture · gpt-image-2 · muAPIVerbatim bash + prompt for swapping a product label.
- 3Infographic/diagram (generate-mode) · Picture · gpt-image-2 · muAPIVerbatim bash + prompt for generate-mode infographic/diagram creation.
- 4Image-to-video · Video · kling-pro · muAPIVerbatim bash, model kling-pro.
- 5Lipsync · Talking video · latent/LatentSync or sync/Sync Labs · muAPIVerbatim bash, model latent/LatentSync or sync/Sync Labs.
- 6Music + SFX · music · Suno V5, MMAudio · muAPI3 verbatim bash examples: Suno V5 song, SFX clip, MMAudio video scoring.
Ours ·- 1
Scroll-stop scrolly-telling site buildRunsA scroll-driven HTML/JS/Canvas microsite built from an already-produced video (frame extraction, sticky-canvas scroll animation, snap-stop annotation cards, count-up specs).
- 1Extract frames · Videoffprobe/ffmpeg frame extraction, 60-150 frames, white first-frame required.
- 2Build the site · editFull scroll-driven site build: starscape, sticky-canvas scroll animation, snap-stop annotation cards, count-up specs, optional Three.js card scanner.
Ours ·- 1
Flora AI — 8 node-based canvas recipesPartlyEight distinct creative recipes on Flora's node canvas: logo, brand identity, social ad, video background/hero, product image, UGC-style video, layer compositing, and website assets.
Recraft V3Ideogram 3.0Nano Banana 2Nano Banana ProSeedream 4.5Luma Ray 2Kling 2.6 ProQwen Image EditGPT Image 1.5Wan 2.2 Animate- 1Logo creation · Picture · Recraft V3 / Ideogram 3.0 · Flora AIText Node brainstorm → Recraft V3/Ideogram 3.0 4-variant gen → refine → Magnific Precision upscale, transparent bg.
- 2Brand identity from seed asset · Picture · GPT-5/Claude Opus 4.5, Nano Banana 2, Luma Ray 2 · Flora AIName brainstorm → logo → brand guidelines (GPT-5/Claude Opus 4.5) → Instagram ad (Nano Banana 2) → animated hero (Luma Ray 2, Loop on).
- 3Social ad creation · Video · Seedream 4.5, Nano Banana Pro, Kling 2.6 Pro · Flora AIShot list (Claude/GPT-5) → Seedream 4.5/Nano Banana Pro stills → inpainting fixes → Kling 2.6 Pro video (native audio sync) → Topaz 4K.
- 4Video background/hero · Video · Luma Ray 2 · Flora AILuma Ray 2 or Nano Banana Pro still → Luma Ray 2 Loop-on video (only model with native loop toggle) → Topaz upscale. Sora 2 flagged shutdown-risk, redirect to Seedance 2.x/Kling 3.0.
- 5Product image creation · Picture · Nano Banana Pro, Qwen Image Edit · Flora AIReal photo → Nano Banana Pro 3x3 angle grid (Asset Vault reference) → lifestyle placement (mandatory 'match the lighting' phrase) → Qwen Image Edit Angles orbit → 4K upsample → Magnific/Topaz finish.
Prompt
match the lighting
- 6UGC-style video · Video · Nano Banana Pro, Veo 3.1 Ingredients mode · Flora AINano Banana Pro AI actor (literal selfie-style prompt) → product grid → composite (literal blend prompt) → Veo 3.1 Ingredients mode, literal template fusing actor+product+script+auto-lipsync audio in one generation → Wan 2.2 Animate fix if product details drop.
- 7Layer Editor compositing · Picture · Flora AIConnect layers → manual canvas placement → mandatory AI blend pass (literal blend prompt); explicit warning raw layer output 'will look pasted-in'.
- 8Website assets (partly) · Picture · GPT Image 1.5, Recraft V3, Nano Banana Pro, Flux Dev, Reve · Flora AIHero (reuse #4), icons (GPT Image 1.5 transparent bg), UI mockups (GPT Image 1.5/Recraft V3), social banners (Nano Banana Pro, ≤14 ref images), background textures (Flux Dev/Reve + outpainting).
Missing: Recipes 1-7 complete with literal prompts where cited; recipe 8's sub-steps partly described only.
Ours ·- 1
Higgsfield access/orchestration playbookManualNot a content recipe itself — an access/orchestration reference for reaching Higgsfield (browser suite / Supercomputer cloud agent / MCP+CLI) and a manual brand-kit workaround.
- 1Access + plan/generate/feedback · By hand · HiggsfieldThree access surfaces (browser suite / Supercomputer cloud agent / MCP+CLI); manual brand-kit engineering (no native brand kit); PLAN→GENERATE→FEEDBACK loop; cost discipline noted.
Missing: Self-flagged in the source with an 'Uncertainties, do not fabricate' section; OAuth-only auth (no CI/API key); quality ceiling ~4-6s/720p.
Ours ·- 1
Brand identity build (upstream of all creative work)RunsA full brand kit: positioning, name, voice doc, colour tokens, typography, and a finished logo/visual identity.
Midjourney v6Nano Banana ProRecraft- 1Positioning · WordsApril Dunford competitive-alternatives method.
- 2Naming · WordsAI-consultant brief, ranked names + trademark-risk flags.
- 3Voice · Words4-dimension mapping, voice doc, 3-level scaling ladder incl. PEFT.
- 4Colour + Typography · Words · Coolors.co, Khroma, Fontjoy, FontPair AIColour via Coolors.co/Khroma (max 2-3 colours, CSS tokens); typography via Fontjoy/FontPair AI (max 2-3 fonts).
- 5Visual Identity/Logo · Picture · Midjourney v6, Nano Banana Pro, RecraftLiteral Midjourney v6 prompt template with --no/--style raw/--sref/--sw/--cref flags; Nano Banana Pro for brand-kit imagery; Recraft for editable SVG; 360° 9-angle character sheet.
- 6Post-processing · Upscale · Topaz GigapixelMandatory post-processing: Midjourney U1-U4/Topaz Gigapixel → vectorise → replace all AI letterforms with licensed typography, never keep AI-generated logo type.
Ours ·- 1
Sora2/Sora2 Pro autopilot with Character Cameos (Sheet-driven)Runsmulti-scene Sora2 videos featuring a reusable branded/character cameo
Sora2Sora2 Pro10s or 15s clips- 1Collect inputs · input · Google SheetsGoogle Sheet row provides scene descriptions, optional image_reference, model choice (Sora2/Sora2 Pro), cameo username
- 2Generate scene prompts · agent · OpenAI (agent, model unspecified)AI agent generates N video prompts, dynamically choosing image-to-video or text-to-video based on presence of image_reference, tagging the cameo handle (@username) inline where relevant
- 3Generate video · generate-video · Sora2 or Sora2 Pro · Kie.aiHTTP POST to Kie.ai Sora2/Sora2 Pro endpoint with the 15-second request body template (prompt including @cameo tag, model, duration, aspect ratio)
- 4Poll and log · poll · Kie.aiPoll/switch node branches on success/error, logs result or error back to sheet
Missing: full verbatim prompt examples (UGC sun-hat/VINADA dialogue, historical POV bodycam, anime concept video) exist in n23 but are not reproduced field-by-field here
RoboNuggets · n24 - Autopilot Cameos with Sora2 Pro- 1
Veo3.1 multi-format video generator (Sheet-driven)Runsvideo across 4 generation modes: first/last-frame, ingredients-to-video (up to 3 reference images), reference-to-video, and text-to-video
Veo3.1 fastVeo3.1 qualityfirst/last-frameingredients-to-video (up to 3 images)reference-to-videotext-to-video- 1Select generation type · input · Google SheetsGoogle Sheet 'tool' column selects which of the 4 generation types to use, plus model choice (Veo3.1 fast vs quality)
- 2Generate video · generate-video · Veo3.1 fast or Veo3.1 quality · Kie.aiSingle Kie.ai HTTP request node handles all 4 generation types (first/last frame, up to 3 reference images for ingredients-to-video, single reference for reference-to-video, or plain text-to-video)
- 3Poll and log · poll · Kie.aiPoll job status and log result back to sheet
Missing: one engine/pipeline covering 21 distinct illustrative use cases (UGC try-on, real-estate promo, apparel, food videography, product showcase, 360 shoe shot, studio/people shots, low-res photo animation, editorial shots, commercial ads, viral animal videos, ASMR, CCTV-style, anime-vlog, character turnaround sheet, logo animation, lower-thirds, static-image animation, 3D motion graphics) — catalogued here as one workflow with formats as presets, not 21 separate objects; use-case-specific prompts not individually reproduced
RoboNuggets · n25 - 21 Creative Uses for Google Veo3.1- 1
Sora2 Character Cameo video (client-facing web form)Runsmulti-scene Sora2 videos featuring one or more reusable cameo characters, triggered per-client via a web form rather than a bulk sheet
Sora2Sora2 Pro10s or 15s clipsSora2 or Sora2 Pro quality- 1Create cameo character (manual, one-time) · manual-setup · Sora (app) · OpenAI Sora appCreate a Sora cameo character in the Sora app; for photorealistic real people, first generate an AI character via a detailed voice/face/clothes/UGC-scene prompt, then 'create a character' from that generated video
- 2Collect client inputs via web form · input · n8n form triggern8n form fields: clip count, scene descriptions, cameo username(s), optional reference image upload, aspect ratio, duration (10s/15s), quality (Sora2/Sora2 Pro)
- 3Upload reference image to URL · upload · Kie.aiConvert form-uploaded image to a web-accessible URL via Kie.ai (form uploads aren't public by default)
- 4Generate scene prompts · agent · OpenAI (agent, model unspecified)AI agent generates N scene prompts tagging the cameo handle(s)
- 5Generate video · generate-video · Sora2 or Sora2 Pro · Kie.aiKie.ai Sora2/Sora2 Pro HTTP POST + poll/switch (same pattern as n24)
- 6Optional upscale · Upscale · flashvsr · Wavespeed AIOptional upscale via Wavespeed's 'flashvsr' model, e.g. 15s video to 4K
- 7Deliver · deliver · Box.comUpload successful videos to Box.com
Missing: manual cameo-creation workaround for real people is a UI/app procedure, not automatable in n8n
RoboNuggets · n27 - Character Cameos with Sora2- 1
Brand-styled animated HTML slide deck (frontend-slides skill)RunsA full HTML/CSS/JS slide presentation, restyled to the user's own brand design system, with a library of custom animated components (charts, icons, motion pads) and photos pulled from Unsplash, viewable and presentable full-screen in the browser.
Claude Code (writes the HTML/CSS/JS itself; no image/video generation model)HTML slide deck / presentation- 1Install base slides skill · By handInstall Zara Zhang's open-source frontend-slides GitHub skill (or the community's Master Slides skill + component library) into Claude Code.
- 2Generate a vanilla test deck · WordsAsk Claude Code to create a sample deck from the skill to inspect its default (vanilla) look before customizing.
- 3Apply brand design system · editFeed Claude the brand book PDF (or a screenshot of an inspirational site) and ask it to update the installed skill to follow the brand's colors, typography and principles, with clarifying questions first.
Prompt
take a look at this brand book design system, and can you please update that skill we just installed so that it follows the principles behind this brand book, please? You can also ask me clarifying questions so that the intention is clear and aligned before you proceed.
- 4Regenerate deck on-brand · editRegenerate the earlier sample deck using the newly brand-aligned skill.
Prompt
update the previous slides you made with this new skill.
- 5Build a custom animated component library · motion-copyAsk Claude to design and build a set of original animated HTML components (motion pads, text effects, animated charts, UI mockups) suited to the user's business, laid out in a Bento grid on localhost for review.
Prompt
I need animated HTML components that we can include in this skill. Can you create six of them for me, put them in a Bento site, and open it for me in a new HTML page in my localhost in my device so that I can inspect them. Feel free to suggest different animations depending on our setup and based on everything you know about the work that I do.
- 6Rework deck with component library · editAsk Claude to rework the presentation to incorporate the just-built components (icons, SVGs, orbiting particles etc.).
Prompt
rework the slides that you were building to include components from the library that you just built.
- 7Add stock/own photos · editAsk Claude to add photos to specific slides, sourced from Unsplash (or the user's own images) by linking to them from within the HTML.
Prompt
add slide four and slide five... and to get them from Unsplash
RoboNuggets · n49 - Claude HTML Slides- 1
Claude Design slide deck with MCP-generated imagesPartlyAn editable, presentable slide deck built directly in Claude Design (with speaker notes and PPTX export), optionally including an AI-generated title image pulled in live through an MCP connector (e.g. Higgsfield's GPT Image 2), and a design system ported over from Claude Code via /design sync.
GPT Image 2 (invoked via Higgsfield MCP connector)slide deck (web-presentable, speaker-note view)PPTX export- 1Open Claude Design, choose Slides template · By handGo to claude.ai/design and select the Slides template to start a deck.
- 2Generate the deck · WordsPrompt Claude Design to build a short deck about a topic, with guidance notes on what to highlight, styled to the selected design system.
- 3Manual edit pass · editUse the rebuilt manual editor (Pro toggle gives a layered, Figma-like view) to click any element and directly retype text, change font, resize, bold/italicize.
- 4Markup feedback · CheckUse the markup tool to select elements or draw on the slide and leave a running list of comments for Claude to address.
Prompt
across all the slides, don't use em dashes.
- 5Generate a title image via MCP · Picture · GPT Image 2 · HiggsfieldWith the Higgsfield MCP connector configured, ask Claude Design to add a title slide and generate an appropriate image using Higgsfield's GPT Image 2 tool.
Prompt
add a title slide and use Higgsfield GPD image 2 to generate an appropriate image.
- 6Sync a design system from Claude Code · By handIn Claude Code, run the /design sync skill on an existing brand book/design system project to port it into Claude Design as a reusable design-system asset for future decks/dashboards/animations.
- 7Export to PowerPoint · publishClick share > export to download the deck as an editable .pptx file; select 'editable custom fonts' if the design system uses non-universal fonts, or they may be swapped on export.
Missing: Needs Claude Design access (claude.ai/design), the Higgsfield MCP connector added under Claude's connector settings for image generation, and either an existing design system built via /design sync in Claude Code or one created from scratch in Claude Design.
RoboNuggets · n61 - Claude Design just got Massive Upgrades- 1
Claude Design automated deck with live data and video clipsPartlyA full on-brand pitch/report slide deck with HTML/CSS animated transitions, live data pulled from a Google Drive or SharePoint spreadsheet, optional Studio-Ghibli-style AI video clips embedded per slide, and a Claude Code routine that refreshes the deck's data weekly.
Kling 3.0 (Studio Ghibli-style video clips, via Higgsfield; transcript renders the name as "Cloud 3.0")pitch deckdata/report deck with chartsdeck with embedded AI video clips (9x16 or other ratios)PPTX export- 1Prompt the full deck with a design system · WordsAt claude.ai/design, choose the Slides preset, toggle on a saved brand design system, and prompt for a multi-slide deck (optionally from a live website or uploaded files for context).
Prompt
make a five-slide pitch deck for Rubrik.
- 2Apply slide-principles skill · editAttach a 'slide principles' skill file (pyramid principle, takeaway-style titles etc.) as a chat attachment and ask Claude Design to apply it, which adds HTML/CSS animations the deck didn't have before.
Prompt
this skill file is the best practice principles for slides and I want Claude design to use it to improve this presentation.
- 3Manual + markup edits · editUse the edit sidebar for direct text/style changes (bold, italic, underline, retype) and the markup tool to leave element-level comments for Claude to address.
- 4Add a self-generated tweaks panel · editAsk Claude to add a tweaks panel with sliders (color, motion intensity, text verbosity/headline case); Claude self-prompts to build the panel, which you then use to batch-restyle many slides quickly.
Prompt
add tweakable controls to this pitch deck, to study this design and add a tweaks panel with two or three expressive controls.
- 5Pull in live spreadsheet data · WordsWith a Google Drive (or SharePoint) connector configured, give Claude Design a spreadsheet link and ask it to summarize the data into slides with tables/charts and takeaways.
Prompt
make three slides to summarize the data in that spreadsheet.
- 6Generate Studio-Ghibli-style video assets · Video · Kling 3.0 · HiggsfieldGenerate short stylized video clips via the connected Higgsfield MCP tool, specifying aspect ratio and model, for use as slide backgrounds/inserts.
Prompt
generate these Studio Ghibli style animations. I wanted it to be 9x16, although obviously you can do other aspect ratios as well. And I wanted to use Cloud 3.0 via Higgsfield.
- 7Assemble video-led launch slides · editGive Claude Design the URL/link to the generated video assets and ask it to build launch-video slides with the clips placed on the right and titles alongside.
Prompt
get the videos from here and make slides for a launch video.
- 8Send to Claude Code and schedule a weekly refresh · publishSend the finished deck's prompt/zip to Claude Code and ask it to set up a routine that pulls the latest spreadsheet data from the linked Google Drive folder on a weekly cadence, refreshing the deck's data slides automatically.
Prompt
set up a routine that grabs this information for us from this folder in Google Drive every week.
Missing: Needs both Claude Design and Claude Code, a Google Drive or SharePoint connector authorized for the data source, a Higgsfield MCP connector for Kling 3.0 video generation, and the 'slide principles' skill file (described as exclusive to the RoboNuggets community and not included in this harvest).
RoboNuggets · n62 - Beautiful Slides with Claude Design- 1
Fable 5 ad content factory (reverse-engineer top ads, regenerate)ManualA batch of new ad creative for a product, produced by having Claude Fable 5 reverse-engineer a product's most-viewed commercials (via the YouTube Data API) and then plan/generate a month's worth of new ad content through Higgsfield or fal.ai.
ad creative (image and/or video, count and exact format unspecified)monthly content calendar of ad pieces- 1Reverse-engineer top commercials · researchAsk Fable 5 to connect to the YouTube Data API, find the most-viewed commercials for the product, and reverse-engineer what makes them work.
Prompt
have it analyze, let's say, the most viewed commercials of this product by connecting to the YouTube data API and just asking it to reverse engineer them and generate 50 ads for your product.
- 2Generate ad variations · Picture · Higgsfield or fal.aiHave Fable 5 direct generation of new ad creative based on the reverse-engineered commercial patterns, via a connected Higgsfield subscription or the pay-as-you-go fal.ai aggregator.
- 3Plan a month of content pieces · Words · Higgsfield or fal.aiSend a templated prompt asking Fable 5 to act as content planner: it asks clarifying questions, then plans a month's worth of content pieces built around the Higgsfield/fal.ai connector.
Missing: No verbatim prompt or confirmed generation model in the harvested transcript (the source itself hedges: 'most likely with something like GPT Image 2'); the actual prompt lives in a linked PDF guide that was not harvested here; needs a Higgsfield or fal.ai account and YouTube Data API access.
RoboNuggets · n63 - Claude Fable 5 Use Cases you MUST try- 1
Marp markdown slide decksPartlyA multi-slide presentation deck written as a single markdown file, styled with HTML/CSS, previewable in the IDE and exportable to PDF or PowerPoint.
slide deckreport deckrecipe/travel/infographic card packspresentation template- 1Install Marp skill · WordsClone the robonuggets/marp-slides skill into Claude Code and follow its README to set it up.
Prompt
I'm giving you a skill called Marp Slides. Get the files with: git clone https://github.com/robonuggets/marp-slides Follow the README to set it up. Recommend how it would best apply to our setup in plain language, and ask me a few questions to clarify my intent before you install it.
- 2Feed source data · researchGive Claude Code a data file (e.g. a Meta Ads CSV) or topic to build the deck from.
Prompt
I have a CSV file here in this location with my April Facebook ad campaign results. Please create a Marp presentation that reviews the performance and a couple of things that I want to include in here and finally a recommendations slide with next steps.
- 3Generate Marp markdown deck · WordsClaude Code analyses the data and writes a markdown file styled with the Marp skill's design principles (fonts, bar/pie charts, HTML interactive elements).
- 4Preview in IDE · CheckInstall the 'Marp for VS Code' extension (with markdown.enableHtml set to all) to render and present the deck directly from the markdown file.
- 5Export · publishExport the finished deck to PowerPoint or PDF from within Claude Code.
- 6Save as reusable template · By handSave the tuned markdown file as a template so future data (e.g. weekly ad results) can be dropped in to regenerate the same deck, optionally scheduled.
Missing: Which image model (if any) sources non-Unsplash photos is not stated; the exact skill file contents were not harvested, only its install prompt.
RoboNuggets · b10 - Claude Code MARP Presentations- 1
Excalidraw diagrams via ClaudePartlyA hand-drawn-style diagram (flowchart, architecture map, decision chart, comparison chart) drawn live on an Excalidraw canvas and exportable as PNG/SVG or a .excalidraw file.
flowchartarchitecture diagramdecision chartvisual comparisoninfographic- 1Install Excalidraw skill/MCP · WordsClone the robonuggets/excalidraw-skill into Claude Code, which installs the open-source MCP Excalidraw project (a live-syncing canvas, distinct from the official Excalidraw MCP).
Prompt
I'm giving you a toolkit called Excalidraw Diagramming. Get the files with: git clone https://github.com/robonuggets/excalidraw-skill Follow the README to set it up. Recommend how it would best apply to our setup in plain language, and ask me a few questions to clarify my intent before you install it.
- 2Open canvas · By handClaude opens a local canvas at localhost:3000 in the browser where it will draw.
- 3Describe and draw the diagram · PictureClaude reads its design guide, clears the canvas, plans the layout (boxes/arrows/connections decided dynamically), draws the diagram, then zooms to fit.
Prompt
Hey Claude, can you do a diagram of how the Excalidraw MCP skill works? Make it simple, but visually compelling. And also, I want you to use the Excalidraw skill that we just installed for you to do this.
- 4Self-check via screenshot · CheckClaude takes a screenshot of what it drew to visually verify correctness before handing back to the user.
- 5Generate variations · PictureOn request, Claude clears the canvas and produces multiple distinct layout variations (e.g. simple left-to-right, conversation loop, vertical view) so the user can pick a direction.
Prompt
Give me three different variations, so that I can choose which one is directionally more correct and aligned to my intent.
- 6Refine and update the skill · WordsUser gives numbered feedback (language level, layout, visual elements, text handling, color scheme); Claude rewrites the skill file so future diagrams default to that taste.
- 7Export · publishExport the diagram as an image file or as a .excalidraw file for further manual editing.
- 8Schedule recurring diagrams · publishClaude Code can be scheduled (cron) to research a topic, draw a diagram with the skill, export it, and send it to Telegram on a set cadence.
Missing: The underlying MCP Excalidraw project is open source and would need to be verified/installed against our own setup; no generating image model is involved, Claude draws primitives directly.
RoboNuggets · b11 - BEAUTIFUL Diagrams with Claude + Excalidraw- 1
Claude Design brand asset generationPartlyBrand-consistent design assets generated from a saved design system and a single prompt: slide decks, email/newsletter templates, UI mockups, interactive web components, and short motion-graphic videos.
slide deckemail newsletter templateUI/app mockuplanding pageshort motion-graphic video (15s-1min)- 1Build a design system · WordsCreate a new design system in Claude Design by linking a code folder (brand book PDF + HTML design elements) or by using a website-to-brand-book skill; Claude codifies typography, accent colors, icons and brand elements from it.
- 2Grab reference web elements (optional) · researchUse the 'grab web element' bookmarklet to copy HTML elements from a reference site (e.g. iridescent cards) onto the clipboard, then paste them into Claude Design as additional visual inspiration alongside the design system.
- 3Generate the asset · PictureStart a 'high-fidelity' project, select the design system, and prompt for the asset (e.g. a monochromatic playing-card UI with hover glow, a monthly newsletter, a 12-slide Meta Ads report from an attached CSV).
Prompt
Have Claude design me a monthly email and newsletter for Rubric, which is our command center app.
- 4Tweak · editUse auto-generated custom sliders (iridescence, saturation, scale, tilt) or the edit panel (fonts, colors, copy) to adjust the result without re-prompting; annotate directly on the canvas for bigger change requests.
- 5Generate a motion-graphic video · VideoStart a project from the 'animation' template (which loads an animated-video/motion-design skill) and prompt for a short announcement video; Claude Design returns a short motion graphic (e.g. 15s brand end-tag, or a longer 1min sprite-based explainer).
Prompt
Create an animation to announce the Rubric app.
- 6Save as template / export / hand off · publishSave any asset as a reusable template (shareable org-wide on team/enterprise plans) for recurring reports; export as PDF, PowerPoint, or standalone HTML (e.g. into Klaviyo); or hand off the design + README as a prompt package to Claude Code to build the working application.
Missing: Requires a Claude Pro/Max+ subscription (claude.ai/design); usage limits were reported as severely restrictive at launch; no specific generating model is named for the images/video inside Claude Design itself.
RoboNuggets · b14 - Claude Design 10+ Use Cases- 1
Claude Code /design skill assetsPartlyEditable design artboards generated from one prompt inside Claude Code itself: landing pages (with optional video backgrounds), email templates, mobile app mockups, and 3D objects, all pulling from the user's existing memory/context rather than a blank slate.
landing pageemail templatemobile app mockup3D object render- 1Update and invoke /design · WordsUpdate Claude Code and type /design (built in, no install needed) with one prompt describing the page to build; Claude Code draws on the user's existing workspace context (memory, design systems, brand books, business details, skills) rather than starting blank.
- 2Review and tweak the artboard · editOpen the returned link to an editable Claude artifact; adjust accent colours, spacing, font scale/weight directly, or via a Figma-like sidebar, and edit copy inline.
- 3Add a generated video background · VideoInvoke the user's own /generate skill (a separate skill connecting to multiple AI video/image models on a pay-as-you-go basis) from within the same session to create a relevant video and place it as the page's background.
Prompt
Generate a relevant video for the site and place it as the background.
- 4Share or comment · CheckToggle link sharing (anyone-with-the-link) similar to a Google Doc, or leave comments directly on the canvas that feed back into the same Claude Code session for bigger changes.
- 5Deploy · publishAsk Claude Code to deploy the finished site to Vercel; it deploys and returns the live URL directly (unlike Claude Design, which only instructs the user to do it themselves).
Prompt
Deploy this site to Vercel.
Missing: The video-generation step depends on a separate /generate skill (models/providers not specified in this source) that was not itself harvested in this batch.
RoboNuggets · b24 - Claude Code's official /design skill- 1
Multi-model creative engine (image + video ad generator agent)Partlybatches of unique product/UGC ad images and videos (studio shots, UGC talking-style ads, product floating-object shots) generated across multiple named AI models in one agent session, logged to Airtable
GPT Image 1.5Nano Banana ProVeo 3.1Sora 2 ProKling 3.0Kling 2.6studio product shotUGC-style adflexible — general-purpose engine, not one fixed output shape- 1Brief the agent (batch image ads) · Words · agent orchestrates GPT Image 1.5 + Nano Banana ProUser tags a product reference photo and asks for N ads across a chosen mix of models.
Prompt
can you create 10 ads with this product? use a mixture of GPT image 1.5 and Nano Banana Pro across those 10 ads...
- 2Generate images · Picture · GPT Image 1.5, Nano Banana Pro · Kie.ai / fal.aiGenerates 2 images per concept to improve selection odds.
- 3Animate selected images to video · Video · Veo 3.1, Sora 2 Pro, Kling 3.0 · Kie.ai / fal.ai / WavespeedUser picks a model per shot; the agent crafts SEALCaM-style prompts itself.
- 4Maintain prompting-best-practices file · Words · agentAgent maintains a persistent prompt best-practices markdown (~400 lines) covering UGC realism cues, skin-detail keywords, the 'BOPA' consistency framework for influencer work, and SEALCaM for cinematic/hero content.
- 5Integrate a new model · Words · agentUser copies a model's API docs page (as markdown) from the provider site and pastes it to the agent, which writes the connector script.
- 6Spawn parallel multi-agent runs · By hand · Claude Code / Antigravity / CodexMultiple agent threads run different product briefs simultaneously.
- 7First-time setup (CREATE framework) · By hand · agentSets up .agent/ folder with CLAUDE.md instructions, .env API keys, references/ (brand images, style pegs, prompt best practices), tools/ (per-model Python connectors).
Prompt
I just imported the Creative Engine template into my workspace. Read README.md and CLAUDE.md and walk me through the first-time setup.
Missing: the actual prompt engineering (the 'prompt best practices.md' file — BOPA framework, UGC realism keyword list, ~400 lines) was not harvested as text, only described; would need the zip template's contents to get it verbatim.
RoboNuggets · R56 | The Antigravity Creative Engine- 1