Scrollclaw first frame

First Frame

The viewer's brain makes a stay-or-scroll decision in 3-5 seconds at the visual processing level — before conscious attention. By the time they've read the hook, the decision is already made.

Frame 1 is a visual design problem, not a copywriting problem. Read references/first-frame-psychology.md for emoji pattern interrupts, composition rules, and testing approach.

Prerequisites

• Script approved (from /persona)
• Creator profile exists in workspace/campaigns/<slug>/creators/ or workspace/creators/
• Brand context loaded

Prompt Structure (Three Layers)

Read references/first-frame-prompting.md for the full system. This is what makes images look like phone photos instead of AI.

1. Photorealism pre-prompt — "Raw iPhone 14 photo. Candid moment, unfiltered..." + realism rules + negative prompt (CGI, 3D render, perfect skin, cinematic color grading)
2. Color reference JSON — extracted from a real reference photo, not described in text. See _system/references/color-reference-system.md.
3. Scene description — creator identity + environment + expression + composition

Include "no text, no words, no letters, no writing" in negative prompt to prevent Sora from hallucinating text into the video.

Generation

python3 scripts/generate-first-frame.py \
  --prompt-file workspace/campaigns/<slug>/frame1-prompt.txt \
  --output-file workspace/campaigns/<slug>/frames/frame1.png \
  --creator workspace/campaigns/<slug>/creators/creator-<name>.md \
  --log-file workspace/campaigns/<slug>/output-log.md \
  --aspect-ratio "9:16"

Multi-Frame Formats (Visual Reference Chaining)

For formats with multiple settings (podcast Dan, gym Dan, car Dan):

1. Generate frame 1 FIRST (the canonical face)
2. Review and approve frame 1
3. ALL subsequent frames MUST use frame 1 as visual reference
4. Do NOT generate frames in parallel from text descriptions alone — causes face drift
5. For Visual Transformation (5-6 frames): generate sequentially, each referencing frame 1

Generate context-specific first frames — one per setting. But always chain from the canonical face. Don't feed a podcast frame into gym B-roll.

Iteration

Iterate on frame 1 more than anything else. Expect 2-4 attempts before it passes.

Quality checks:

• Does it look like a phone photo or an AI render?
• Is the creator believable (not model-pretty)?
• Is there real-world clutter in the environment?
• Does the color match the reference JSON?
• Is skin texture visible (pores, not porcelain)?

Once frame 1 is locked, the rest chain from it.

Brand Memory Integration

Reads

File	Purpose
workspace/campaigns/<slug>/creators/creator-<name>.md	Creator identity — face, hair, build, wardrobe, energy
workspace/creators/creator-<name>.md	Global creator profile (fallback if no campaign override exists)
workspace/campaigns/<slug>/scripts/<format>-script.md	Which scene/setting the first frame needs to establish
workspace/campaigns/<slug>/brief.md	Brand context, color cues, environment direction

Writes

File	Notes
workspace/campaigns/<slug>/frames/frame1.png	Canonical face — referenced by all subsequent frames and animation
workspace/campaigns/<slug>/frames/environment-frame.png	Context-specific frames for multi-setting formats
workspace/campaigns/<slug>/output-log.md	Prompt params, model version, generation time (append-only)

Context loading

🖼️ First Frame context loaded:
  ✓ Creator: Maya (workspace/campaigns/ridge-q1/creators/creator-maya.md)
  ✓ Script: talking-head (workspace/campaigns/ridge-q1/scripts/talking-head-script.md)
  ✓ Campaign: ridge-q1

Contract

Input

• Required: approved script plus a creator profile
• Optional: campaign brief, color reference inputs, global creator fallback profile
• Format: workspace markdown files plus prompt text
• Source: /persona, workspace/campaigns/<slug>/brief.md, and _system/references/color-reference-system.md

Output

• Produces: one canonical face image and any context-specific environment frames needed for the format
• Format: PNG files in workspace/campaigns/<slug>/frames/ plus prompt log entries in output-log.md
• Default behavior: generate frame1.png first, stop for review, then chain any additional frames from it
• Downstream use: /animate and /b-roll

Validation

• Pre-conditions: script is approved, creator profile exists, and the target setting is clear enough to visualize
• Post-conditions: frame looks like a phone photo, preserves creator identity, and is strong enough to chain into later generations
• Failure checks: reroll or revise if the image looks like an AI render, the identity drifts, or the scene lacks believable real-world clutter

Output

• Canonical first frame image (PNG) in workspace/campaigns/<slug>/frames/
• Context-specific frames for multi-setting formats in workspace/campaigns/<slug>/frames/
• Prompt params logged to workspace/campaigns/<slug>/output-log.md

Next Step

Frame 1 approved → run /animate for A-roll or /b-roll for environment shots.