Router and orchestrator for AI UGC video production

ScrollClaw

AI-generated UGC videos that look like a real person pulled out their phone and started talking. Brands pay $500–$5,000 per UGC video from human creators. This pipeline produces them for $5–$50 in API costs.

The Pipeline

Brand name + URL
      ↓
/brand-setup → One-time brand research + file generation (Step 0)
      ↓
Brand + Audience
      ↓
/persona    → Persona research, creator profiles, approved script
/first-frame → Canonical face image (Nano Banana 2)
/animate    → A-roll talking head clips (Sora 2 i2v)
/b-roll     → Environment + product shots (Kling 3)
/assemble   → Stitch + unified voice + post-production + captions
/score      → Virality gate (70+ to publish)
      ↓
Scroll-stopping UGC video

Core doctrine: Read _system/SKILL.md for format selection, anti-patterns, taste calibration, and pipeline routing rules.

Brand & campaign context: Read _system/references/brand-campaign-context.md for workspace structure, context matrix, and persistence protocol.

---

Routing

When the user invokes ScrollClaw or asks about UGC video, route to the right sub-skill based on where they are in the pipeline:

User says / wants	Route to
"Set up a brand" / "Initialize brand" / "New brand" / "Brand research"	/brand-setup — one-time, before first campaign
"Make me a UGC video" / "Start a campaign" / "I need ugc"	/persona — start at the beginning
"Research this brand" / "Create a creator profile" / "Write a script"	/persona
"Generate the first frame" / "Make a face image" / "Nano Banana"	/first-frame
"Animate this" / "Make a talking head clip" / "Sora" / "A-roll"	/animate
"Generate B-roll" / "Product shot" / "Kling"	/b-roll
"Stitch the clips" / "Add captions" / "Post-production" / "Assemble"	/assemble
"Score this video" / "Is this ready to publish?"	/score
"Run the full pipeline"	Start at /persona, proceed sequentially

Broad outcome requests stay at the root. Start the pipeline instead of asking the user to pick a sub-skill.

Stage-specific requests bypass the root and go straight to that sub-skill.

If the user isn't sure where they are, ask: "Where are you in the pipeline?" and show them the 6 steps above.

Contract

Input

• Required: brand/product context plus a campaign goal, or an existing campaign workspace if resuming
• Optional: format preference, creator profile, screen recordings/screenshots, existing clips, campaign slug
• Format: raw text, URL, file paths, or workspace/campaigns/<slug>/ files
• Source: user prompt, campaign brief, and upstream workspace/brand/ files

Output

• Produces: either a route decision to the correct stage or a sequential full-pipeline run starting at /persona
• Format: inline orchestration plus saved workspace artifacts produced by sub-skills
• Default behavior: broad requests like "make me a UGC video" or "start a campaign" start at /persona and continue stage by stage until blocked by missing inputs, an approval gate, or a dependency failure
• Downstream use: /persona, /first-frame, /animate, /b-roll, /assemble, and /score

Validation

• Pre-conditions: workspace exists or can be initialized before generation steps begin
• Post-conditions: the user knows the current stage, the next stage, and which workspace artifacts were produced
• Failure checks: do not leave the user at a vague route; if blocked, name the exact missing file, asset, or approval needed

---

Quick Start

First campaign:

1. Set up brand context (one-time per brand):
   • Recommended: Run /brand-setup with your brand name + URL — it researches and generates all three brand files automatically
   • Alternative: Copy templates from assets/ and fill in manually (see assets/ for templates)
2. Initialize workspace:

   CAMPAIGN="my-campaign"
   mkdir -p workspace/campaigns/$CAMPAIGN/{creators,scripts,frames,clips,scores}
   cp assets/campaign-brief-template.md workspace/campaigns/$CAMPAIGN/brief.md
   touch workspace/campaigns/$CAMPAIGN/output-log.md
   touch workspace/campaigns/$CAMPAIGN/learnings.md
   

3. Fill in workspace/campaigns/$CAMPAIGN/brief.md
4. Run /persona → /first-frame → /animate → /b-roll → /assemble → /score

Check setup first:

bash scripts/check-deps.sh

---

What ScrollClaw Needs

Key	Required	Used by
FAL_KEY	Yes	Sora 2 (A-roll) + Kling 3 (B-roll)
REPLICATE_API_TOKEN	Yes	Nano Banana (first frames)
OPENROUTER_API_KEY	Recommended	Gemini (virality scoring)
ELEVENLABS_API_KEY	Optional	Multi-clip voice consistency (S2S)

---

Six Formats

Format	Duration	Best for
Talking Head	15-25s	Product review, honest take
Hook Face + Demo	15s max	App/tool demos
Podcast Clip	8-20s	Authority, credibility
Wall of Text	4-8s	Hot take, faceless
Visual Transformation	10-25s	Before/after concept
Hybrid Transformation	20-30s	Complex mechanism explanation

---

Key Findings (from testing)

• Sora's native voice > ElevenLabs TTS for talking head. TTS sounds fake.
• B-roll must be environment-matched. Extract a frame from A-roll → feed to Kling.
• Captions go LAST — after post-production. Grain degrades caption pills.
• AI cannot generate realistic app screens. Use real screenshots.
• ~1 in 3 Sora generations have hand artifacts. Reroll, don't fix the prompt.
• Multi-frame formats: chain from frame 1. Parallel generation causes face drift.

---

For full doctrine, format blueprints, anti-patterns, and taste calibration: _system/SKILL.md