| name | ai-ugc-video-pipeline |
| description | End-to-end AI UGC ad production workflow for realistic AI creator videos. Use when a user wants Codex to turn a short ad brief into UGC video assets, coordinate UGC script writing, prepare creator/product references, create chunked scene prompts, set fal video generation settings, review clip failures, plan regenerations, and assemble TikTok/Reels/Shorts-ready ads. |
AI UGC Video Pipeline
Use this skill to run the transcript-derived workflow for realistic AI UGC ads: prepare references, write a natural script, split it into controllable clips, generate/review/regenerate clips, then produce an edit-ready assembly plan.
Required inputs
Collect only what is missing:
- Product/app name and landing page or product summary.
- Offer or CTA.
- Target viewer and awareness level.
- Format:
mid-funnel punchy or full-stack.
- Creator persona or reference package. For AutoVisuals, prefer
@imagen/ugc-influencers and marcHalePersona.
- Product/app image assets or screenshots.
- Desired aspect ratio, usually
9:16.
- Video generator target: default to fal Kling O3 Standard for cost-effective tests when available, with Seedance 2.0 as the higher-cost quality baseline.
If the user gives a brief and the repo already has persona/product assets, proceed using those assets.
Skill handoffs
- Use
$ugc-script-writing for the script, hook framework, voiceover, chunking, captions, and natural UGC language.
- Use
$fal-seedance-video for fal endpoint choices, request payloads, queue execution, clip generation, and output validation.
- Use project persona data when available before inventing a new creator.
Workflow
- Define the ad strategy.
- Pick
mid-funnel punchy for cheaper 18-22s warm-audience ads.
- Pick
full-stack for 28-32s education-led cold-audience ads.
- Choose one viewer pain, one mechanism/reframe, one soft CTA.
- Prepare source assets.
- Creator: use one high-quality consistent avatar/reference image.
- Product/app: use a clean isolated product image, app screenshot, or app demo frame.
- Remove references from scene prompts when a chunk says the asset should not appear.
- Build the script with
$ugc-script-writing.
- Require conversational voiceover that sounds spoken, not written.
- Require chunked output with estimated runtime, scene direction, product inclusion, and continuity notes.
- Create universal prompt blocks.
- Character lock: identical creator identity and wardrobe continuity.
- Product/app lock: exact asset name and when it is allowed to appear.
- UGC realism: natural camera, off-camera phone unless selfie is required, hand gestures, imperfect-but-believable creator delivery.
- B-roll sequencing: no product/app reveal before voiceover names it.
- Generate clip prompts.
- Each chunk gets one clip prompt.
- Paste universal blocks into every clip.
- Replace only the chunk voiceover, framing, runtime, and asset-inclusion line.
- Keep each clip independently regenerable.
- Generate clips with
$fal-seedance-video.
- Default provider for TikTok/Reels tests: Kling O3 Standard image-to-video with audio on.
- Seedance fallback/baseline:
aspect_ratio: "9:16", resolution: "720p".
- Prefer short 5-6s chunks unless the line needs more space.
- Use 10-12s only for longer mechanism or payoff chunks.
- Review each clip.
- Keep: natural voice, stable face, matching persona, no unwanted product/app asset, acceptable hand motion.
- Regenerate: voice changes sharply, product appears when omitted, lip sync is distracting, face drifts, pacing is unusably fast/slow, or hands/product deform badly.
- First try runtime adjustment before rewriting a good scene.
- Assembly plan.
- Order clips by chunk number.
- Note any speed changes such as 90% for fast lines.
- Add simple captions: readable sans/classic style, white text with dark outline, raised above platform UI.
- Add light music only if it helps mask small voice differences.
Output shape
When asked to produce an ad package, create:
Ad name:
Format:
Target viewer:
Hook framework:
Runtime target:
Creator/persona:
Source assets:
Full voiceover:
Universal direction block:
Chunks:
1. Runtime:
Include creator asset:
Include product/app asset:
Voiceover:
Scene prompt:
Review risks:
Generation settings:
Regeneration rules:
Edit assembly notes:
Transcript rules to preserve
- More control beats one-shot generation. Generate separate clips and stitch them.
- A realistic creator is not enough; the script must sound authentic.
- Do not include product/app assets in prompts for chunks where they should not appear.
- Use runtime as a pacing control. If a voice sounds off, try changing duration and regenerating that chunk.
- Expect at least one bad clip; plan review and regeneration into the workflow.
- Prefer 720p for cost-efficient social ads unless the user explicitly requests 1080p.
- Final review is against the assembled video, not isolated clips only.