vector-scene-layout

Turns a validated storyboard.json into a Remotion project with one static scene per beat. It fixes the failures that show up when an agent goes straight to animating: pixel sizes typed by hand that are unreadable on a phone, text placed under the feed's controls, a composition whose duration disagrees with the storyboard, fonts that swap mid-video, and scenes that only look right at one aspect ratio. Motion is not added here; the deliverable is a set of stills the agent has looked at.

When to use / when NOT to use

Use when the storyboard exists and the project, tokens, scenes or frames need building or fixing, when text is too small or clipped, or when a colour pair needs a contrast check for video.

Do NOT use for: writing the beats (animation-storyboard), choosing or licensing assets (vector-asset-sourcing), timing and easing (smooth-motion-choreography), rendering video files (animation-export-presets), reviewing a finished render (animation-frame-qa), web page or slide-deck layout, or Remotion API questions the official remotion-dev/skills answer.

Workflow

1. Scaffold once (steps and file layout in references/project-structure.md): npx create-video@4.0.522 --blank <name>, install @remotion/transitions, @remotion/google-fonts, @remotion/paths at the same version, add "resolveJsonModule": true to tsconfig.json, install remotion-dev/skills, copy storyboard.json and assets/manifest.json in, and write animation.config.json with remotion_license_basis (free tier only for individuals, companies of ≤3 people and non-profits; otherwise a Company License).
2. Generate tokens, never type pixel values:

   python3 "${CLAUDE_SKILL_DIR}/scripts/frame_tokens.py" --target linkedin-4x5 --target pptx-16x9 --ts > src/tokens.ts

   Tokens give width, height, safe insets, type scale (caption 32 / body 48 / headline 84 / display 120 px at 1080 wide; feed targets scale with width, slides with height), minimum weights, stroke widths, icon sizes, grid columns and the phone preview scale. Why those numbers: references/legibility.md. fps and duration come from the storyboard through timeline.ts, never from tokens.
3. Copy the templates from assets/templates/ into src/: timeline.ts (frame schedule from the storyboard), fonts.ts, Video.tsx (TransitionSeries over the beats), SceneOnly.tsx and Root.tsx (one composition per target plus one per beat, <target>--<beatId>, for isolated stills), scenes/. Check brand colours:

   python3 "${CLAUDE_SKILL_DIR}/scripts/frame_tokens.py" contrast "#F5F7FF" "#0B1020"

4. Write one scene per beat in src/scenes/<beatId>.tsx and register it in scenes/index.ts. Each scene is a pure function of beat and tokens: text at or above the token size for its role, everything essential inside tokens.safe, one focal element, icons from public/ inline as SVG sized from tokens.icon, strokes ≥ tokens.stroke.min. No useCurrentFrame() yet.
5. Render and look at every beat alone (no transitions blended in), full size and at phone width:

   python3 "${CLAUDE_SKILL_DIR}/scripts/render_stills.py" storyboard.json --target linkedin-4x5 --out out/stills
   python3 "${CLAUDE_SKILL_DIR}/scripts/render_stills.py" storyboard.json --target linkedin-4x5 --phone --out out/stills-phone

   Open the PNGs, the phone-width ones especially. Fix clipping, undersized text, weak contrast, crowding; repeat for every target in animation.config.json. npx tsc --noEmit and npx eslint src must be clean.
6. Hand off: list the scenes, the tokens used per target, and anything the storyboard asked for that the layout could not honour. Next: smooth-motion-choreography.

Output spec

• A Remotion project with animation.config.json, src/tokens.ts (generated), src/timeline.ts, src/Video.tsx, src/Root.tsx, src/scenes/<beatId>.tsx for every beat, and out/stills/*.png per target.
• Every still checked by eye at phone scale; text inside safe areas at token sizes.
• Composition duration equals the storyboard's total_frames (from timeline.ts).

Gotchas

• A 1080-wide video is ~390 logical px wide on a phone: 48 px body text is 17 pt there, 32 px is the floor (Apple's 11 pt). A 1920-wide landscape post in the same feed column needs 1.78× larger text; the tokens scale feed targets by width. Hairlines under 3 px vanish after encoding.
• Vertical (9:16) feeds cover the bottom third with UI and some feeds crop to 4:5: tokens reserve 35% at the bottom and 15% at the top; do not fight them.
• Contrast belongs in luminance. Thin saturated red or blue text fringes after 4:2:0 chroma subsampling; light-on-dark with weight ≥ 600 is the safe default.
• Load fonts with waitUntilDone() behind delayRender() (the template does) or the first frames render in a fallback font.
• Never animate or vary font-size, width or left between frames later; lay out with transforms in mind so the motion step only touches transform and opacity.
• The composition id is the target id; durationInFrames comes from timeline.ts. If the storyboard changes, regenerate nothing by hand: re-validate, re-render stills.
• Frame 0 is the LinkedIn thumbnail: the hook scene must read as a poster on its own.
• Scenes must be pure: no Math.random, Date.now, fetches or CSS transitions.

Pointers

• scripts/frame_tokens.py — tokens (--ts, --target, custom --width --height), contrast; --self-test.
• scripts/render_stills.py — one isolated still per beat at its midpoint; --phone renders at 390 px wide.
• assets/templates/ — timeline.ts, fonts.ts, Video.tsx, SceneOnly.tsx, Root.tsx, scenes/ (typecheck-clean).
• references/project-structure.md — scaffold commands, file layout, how the pieces connect, loops.
• references/legibility.md — type scale derivation, contrast, safe areas with sources.