_scrollclaw-system
ScrollClaw System
AI video defaults to cinematic. Sweeping drone shots. Perfect lighting. Orchestral energy. Nobody scrolls past that thinking "real person" — they think "ad" and keep moving. UGC works because it looks like someone pulled out their phone and talked.
Core Doctrine
Messaging is the actual skill. The tools improve weekly — visuals, voice, motion are getting solved. AI cannot solve having something worth saying. Most AI UGC fails not because it looks like AI but because the script sounds like a copywriter, not a customer. Persona research before production. Real language before prompts.
Anti-polish is the product. Every decision optimizes for "could a real person have shot this on their phone." Impressive is the enemy.
Creators are persistent. AI creators have locked identities — same face, same hair, same build across every clip. First-frame consistency: generate one canonical face with Nano Banana, feed it to every Sora/Kling i2v generation. Read references/creator-system.md.
First frame controls everything. Generate frame 1 with Nano Banana 2 (composition, character, environment, color locked), then animate with Sora 2 i2v. Text-to-video is the fallback.
Save first, watch later. Replicate URLs are ephemeral. Every output gets downloaded immediately.
System not clip. Every run leaves reusable creator profiles, prompt logs, color references. Campaign 10 takes a fraction of campaign 1.