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.

The Six Dimensions

Setting. Cadence. Influencer realism. Story. Product. Audience. Every decision routes through these. Most people skip consumer psychology — that's why their AI looks obvious.

Format Selection

Ask: what are we making? Then route to the right format.

Read references/format-library.md for shot-by-shot blueprints. Read references/hook-emotions.md for the emotion taxonomy.

Pipeline Overview

Anti-Patterns

Visual: Cinematic drift (override with handheld energy), model-pretty creators (specify normal-looking), clean room syndrome (demand real clutter), porcelain skin (specify visible pores), smooth steadicam (phone has micro-shake).

Color: Grey-scale Nano Banana default (fix with JSON color prompts). Too-clean color (lift shadows, slight fade). See references/color-reference-system.md.

Audio: Clean studio audio (add room ambience). Library ElevenLabs voices (voice design or instant clone only). Stock music (real UGC rarely has music).

Script: Testimonial cadence ("I've been using this for three weeks..."). Wardrobe drift across same-day clips.

Pipeline: Output loss (download immediately). Hallucinated text (add "no text" to negative prompt). Extra limbs (reroll, don't fix prompt). Resolution mismatch (normalize before stitching). Multi-clip voice drift (use ElevenLabs S2S).

Taste Calibration

Read references/taste-calibration.md for before/after examples that show what "anti-polish" actually sounds and looks like in practice.

Contract

Input: product/brand context (URL, description, or brand voice file). Optional: creator profiles, format preference, color references.

Output: video clips (MP4), first-frame images (PNG), creator profiles, scripts, prompt logs. Default 9:16 vertical.

Env: FAL_KEY (primary), REPLICATE_API_TOKEN (Nano Banana + fallback), ELEVENLABS_API_KEY (multi-clip voice), OPENROUTER_API_KEY (Gemini virality scoring).

Brand & Campaign Context

ScrollClaw persists work across sessions using a structured workspace. Campaign 10 takes a fraction of campaign 1 because creator profiles, brand context, and learnings accumulate.

Full protocol: Read references/brand-campaign-context.md.

Workspace structure

Copy code
workspace/
├── brand/                    ← Read-only for ScrollClaw (written by /brand-setup, GrowthClaw, or manually)
│   ├── voice-profile.md      ← Brand voice → informs script tone
│   ├── positioning.md        ← Differentiation → informs persona research
│   └── audience.md           ← ICP → informs creator archetype selection
├── creators/                 ← Global creator profiles (reusable across campaigns)
└── campaigns/<slug>/
    ├── brief.md              ← Campaign brief (from assets/campaign-brief-template.md)
    ├── persona-research.md   ← Written by /persona
    ├── creators/             ← Campaign-specific creator overrides
    ├── scripts/              ← Approved scripts
    ├── frames/               ← First frames + context frames
    ├── clips/                ← A-roll, B-roll, assembled finals
    ├── scores/               ← Virality score cards
    ├── output-log.md         ← Prompt log, generation params (append-only)
    └── learnings.md          ← What worked, what didn't (append-only)

Context matrix

Rules for reading brand memory

• Check if each brand file exists before reading. Never error on missing files.
• Show what was loaded: ✓ Loaded brand voice: conversational-direct / ✗ No audience file — proceeding standalone
• ScrollClaw reads workspace/brand/ but never writes there

Rules for writing campaign files

• output-log.md and learnings.md are append-only — never overwrite
• Creator profiles: global ones go in workspace/creators/, campaign overrides in workspace/campaigns/<slug>/creators/
• Skills own their outputs. /persona owns persona-research.md. /score owns scores/ and learnings.md.

Common Mistakes (from live production sessions)

These mistakes have been made in real sessions. Every one of them wasted time and tokens. Do not repeat them.

Never generate random people — always use creator reference

The creator system exists to lock visual identity. Every first frame must be generated from the creator's canonical reference image using i2v. Generating 7 random people instead of using Jess's locked reference was the biggest waste in session 2.

Wrong: text-to-image → random person → animate Right: reference image → Nano Banana 2 → first frame → Sora 2 i2v

Never use text-to-video for established creators — always i2v

t2v generates a completely different person every time. Once a creator has a reference image, always use image-to-video. The only acceptable use of t2v is initial exploration before a reference is locked.

Never caption before post-production

Grain degrades caption text. Color grading shifts caption colors. The order is non-negotiable:

Copy code
raw clips → stitch → post-produce → THEN captions (LAST)

Always QA with vision model after stitch

Use Gemini Flash to scrub every clip frame-by-frame after stitching. It catches content drift (wrong person appearing mid-clip), hand artifacts, and scene inconsistencies that are easy to miss on a quick watch. Don't wait for the user to catch problems.

Always use compatible encoding defaults

bashCopy code
-c:v libx264 -profile:v main -pix_fmt yuv420p -crf 23 -preset medium -movflags +faststart -c:a aac -b:a 128k -ar 44100

yuv444p and High 4:4:4 Predictive profile break Telegram Web, many mobile browsers, and some social platforms. Always yuv420p + main profile + faststart. Do this from the first encode, not as a fix after delivery fails.

Always run pre-flight before generating

bashCopy code
bash scripts/pre-flight.sh <campaign-slug>

Validates workspace structure, creator profiles, reference images, scripts, and caption plans. Catches missing prerequisites before you waste API calls.

Follow the production pipeline order

Read references/production-pipeline.md for the strict step-by-step sequence. Every step depends on the previous one being done correctly.

Setup

Run scripts/check-deps.sh to verify all API keys and dependencies.