raisaksham377-bot/youtube-shorts-ai-editor

make-tsx — build the visual beats over the master

Step 2 of the pipeline: take an approved master cut and its timing spine and produce the animated beats — UI/browser mockups, full-screen statements, diagrams, image reveals — then composite them over the master into a preview. Work collaboratively and beat-by-beat; the user gives notes per beat.

This is the project orchestration. For the low-level rules of writing a single crash-free Remotion file (frame-based only, monotonic interpolate, Easing.bezier, no useState/useEffect, import gotchas), use the vidtsx-2d-generator skill — shots here follow those rules.

Inputs (read these first, every session)

• The plan — videos/video-N/work/edit-plan.md (or a change list like v2-update-plan.md). The source of truth for what to build. If there's a change list from a review, that overrides; apply it, don't re-litigate settled beats.
• videos/video-N/work/timeline.json — maps each shot → its span on the master + type. You edit this. (Schema below.)
• videos/video-N/work/edited-transcript.json — word-level times in master timeline (ms). Everything syncs to these.
• brand.md — colors/fonts/motion/delivery tokens. Every shot reads the brand so all videos feel like one channel.
• remotion/src/lib/kit.tsx (+ lib/vscode.tsx, lib/browser.tsx) — the reusable component library; src/brand.ts, src/fonts.ts; existing shots in remotion/src/shots/video-N/.

Workflow

1. Read the plan + timeline + transcript + brand + kit. Pull the exact word times for each cue you must hit (grep edited-transcript.json for the phrase → start/end ms → local frame = (cue_s − shot.master_in_s) × fps).
2. Ask only what's genuinely open at the start (a small batch), then build. Don't re-ask settled decisions.
3. Library-first. Reuse a kit component before writing UI from scratch. If you build something reusable (a window chrome, a browser frame, a card slot), put it in lib/ so later shots and videos get it. Promote good patterns back into the library.
4. Author / edit the shot as a vidtsx-format .tsx in remotion/src/shots/video-N/ (auto-registered by scripts/gen-registry.mjs). One shot per file. Follow the brand tokens and the principles below.
5. Sync to the words (see principles). Time each sub-element's entrance to its transcript cue; never show a thing before it's said or after the point has passed.
6. Update timeline.json — add/replace/retime the beat's span and type. Leave a clear note when a slot is a placeholder (e.g. a pending screen recording).
7. Render + VERIFY BY SCREENSHOT (never skip): cd remotion && node scripts/render-all.mjs --scale=1 <ShotIds> (opaque → out/<id>.mp4, transparent overlays → .mov). Validate with a still first (--still); for word-synced reveals/scrolls, render the clip and pull a frame at each cue (ffmpeg -ss <t> -i out/<id>.mp4 -frames:v 1 f.jpg) — one 60% still is not enough.
8. Bake the preview: python tools/bake.py (reads timeline.json → videos/video-N/output/video-N-preview.mp4, 1080p30). --end SECONDS to bake a prefix while iterating.
9. Spot-check composited frames from the baked mp4 at the changed beats (the shot over the real master), not just the isolated shot.

timeline.json (the compositing map)

{
  "master": "videos/video-1/reference/master.mp4", "master_fps": 60,
  "remotion_out": "remotion/out",
  "shots": [
    { "id": "CreditsStat", "type": "cutaway", "master_in_s": 170.0, "master_out_s": 179.0, "cue": "..." }
  ],
  "preview": { "end_s": 329.6, "out": "videos/video-1/output/video-1-preview.mp4", "width": 1920, "height": 1080, "fps": 30 }
}

• type: "cutaway" — the shot's .mp4 REPLACES the master video for [master_in_s, master_out_s]. Master audio always continues underneath.
• type: "overlay" — a transparent shot (ProRes 4444 .mov, transparent: true in its compositionConfig) is composited OVER the master for that span. For small persistent CTAs/badges in the top/bottom band.
• bake.py splits [0, end] at every shot boundary, renders each segment (cutaway from the shot, overlay = master+alpha, else master pass-through), concats, and muxes master audio 0..end. Frame-accurate; a shot's file must exist in remotion/out/ first.
• Design shots at 1920×1080; render-all --scale=2 → 4K for final. Preview bakes 1080p30; final delivery is 4K60.

Principles (the house style — apply them)

• Sync to the words (P1). Reveal sub-elements on their narration cue, timed from edited-transcript.json. "Show everything then hold" is a smell.
• Don't pre-empt (P2). Never put on screen a thing mentioned later (don't show "skill" before the skill is introduced; don't show example tiles before "more examples").
• Real assets over generic icons — the DEFAULT (P3, brand.md §6). Any beat naming a real-world thing (a company, a person, a city/country, a real app/product, a real job posting) shows a REAL photo or REAL screenshot of it, not a lucide-icon graphic or a hand-drawn clone — icons are the fallback only for abstract concepts (a stat, a process step) with nothing real to show. Two cases:
  • Real website / app / page → capture it LIVE with tools/capture_web.py (Playwright — see requirements.txt) pointed at the real public URL, saved to media/projects/<video>/real/; frame it in WebBrowserFrame (single static page) or replay it with /fake-screencast (a walkthrough with simulated cursor/scroll/click). This is a real state capture, not a redrawn mockup.
  • Real photo / logo / person / place → WebSearch/the Browser tool to find the actual asset (the entity's own site, Wikipedia, Wikimedia Commons, a YC/LinkedIn-style profile), then curl/download it into media/projects/<video>/real/. Never redraw a logo or skyline from memory — fetch the real file. Animate it with Ken Burns (slow scale 1.0→1.12 + a small translate drift over the beat's hold) rather than showing it static.
  • Prefer official/openly-licensed sources over random image-search hits, especially for anything leaving the machine (a pitch sent to the real company, a public upload). Pair the reveal with the camera-snap SFX cue (brand.md §10).
• Config/setup → real UI (P4). Keys, dashboards, .env, skill folders → the actual VS Code / dashboard UI, not an abstract graphic.
• Real results in the real editor (P5). A generated result is shown opened inside the real editor (image in a pane next to the chat), not floating on a brand card.
• Full-screen vs overlay. Concept beats = full-screen cutaways (cut away from the talking head). Overlays only for small persistent CTAs/badges in the top/bottom band; never cover the center-framed talking head.
• Real UI realism. UI clones match the real product (Claude Code = dark + coral, real logos), not the indigo brand. Brand indigo is for overlays/full-screen beats.
• Recording vs TSX. TSX when you want controlled highlight/scroll/zoom; a real screen recording only when the point is genuine PROOF (real output, real speed). Leave a clearly-noted placeholder slot for recordings the user will provide.
• On-screen text follows the brand voice — no em-dashes (the site copy avoids them; keep it out of overlays too).

Those are the full principle set. When a video's plan calls for a revision pass, record what changed and why in videos/<project>/work/v2-update-plan.md so the next pass inherits the reasoning.

Tooling quick reference

• Register shots: node remotion/scripts/gen-registry.mjs (auto-runs discovery; needed after adding a file).
• Render: node remotion/scripts/render-all.mjs [--still] [--scale=1|2] <ShotIds...>.
• Bake: python tools/bake.py [timeline.json] [--end S] [--keep].
• Scratch renders/frames go in the scratchpad, not the project.

Done = the changed beats render, you have looked at stills/frames at each cue, the preview is re-baked, and composited frames are spot-checked. Update the plan/timeline and any relevant memory when the pass is complete.