Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Step 2 of the AI Video Editor pipeline — build the visual beats (Remotion TSX shots) over a project's master cut and bake a composited preview. Use when the user wants to add/edit overlays or full-screen animation segments, "make a shot / beat", implement an edit plan or a v2 update, retime a beat to the narration, re-render shots, or re-bake the video-N preview in this repo. Covers reading the plan + edited-transcript + brand + reusable kit, authoring shots library-first, syncing reveals to wor
.claude/skills/hassancs91-make-tsx/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-08 | ✗→✓ | ▲ Improved | 72% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 76% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 60% | 0% |
| case-18 | ✗→✓ | ▲ Improved | 10% | 0% |
| case-11 | ✓→✗ | ▼ Worse | 66% | 0% |
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.
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/.edited-transcript.json for the phrase → start/end ms → local frame = (cue_s − shot.master_in_s) × fps).lib/ so later shots and videos get it. Promote good patterns back into the library..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.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).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.python tools/bake.py (reads timeline.json → videos/video-N/output/video-N-preview.mp4, 1080p30). --end SECONDS to bake a prefix while iterating.jsonc{ "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.--scale=2 → 4K for final. Preview bakes 1080p30; final delivery is 4K60.edited-transcript.json. "Show everything then hold" is a smell..env, skill folders → the actual VS Code / dashboard UI, not an abstract graphic.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.
node remotion/scripts/gen-registry.mjs (auto-runs discovery; needed after adding a file).node remotion/scripts/render-all.mjs [--still] [--scale=1|2] <ShotIds...>.python tools/bake.py [timeline.json] [--end S] [--keep].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.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 6,865 | 41,026 | +498% | 1 | 1 | 0% | 231 | 2,338 | +912% | 0 | 0 | — |
case-02 | fail→fail | 8,096 | 16,439 | +103% | 1 | 1 | 0% | 232 | 2,362 | +918% | 0 | 0 | — |
case-03 | fail→fail | 6,452 | 7,160 | +11% | 1 | 1 | 0% | 285 | 2,271 | +697% | 0 | 0 | — |
case-04 | fail→fail | 16,538 | 5,256 | -68% | 1 | 1 | 0% | 1,077 | 2,380 | +121% | 0 | 0 | — |
case-05 | fail→fail | 28,638 | 14,822 | -48% | 1 | 1 | 0% | 4,581 | 4,535 | -1% | 0 | 0 | — |
case-06 | fail→fail | 12,214 | 10,857 | -11% | 1 | 1 | 0% | 1,932 | 2,188 | +13% | 0 | 0 | — |
case-07 | pass→pass | 13,214 | 5,745 | -57% | 1 | 1 | 0% | 2,301 | 2,924 | +27% | 0 | 0 | — |
case-08 | fail→pass | 11,251 | 4,998 | -56% | 1 | 1 | 0% | 1,592 | 2,745 | +72% | 0 | 0 | — |
case-09 | pass→pass | 11,627 | 5,196 | -55% | 1 | 1 | 0% | 1,554 | 2,526 | +63% | 0 | 0 | — |
case-10 | fail→fail | 10,189 | 5,973 | -41% | 1 | 1 | 0% | 1,428 | 2,846 | +99% | 0 | 0 | — |
case-11 | pass→fail | 11,668 | 4,203 | -64% | 1 | 1 | 0% | 1,480 | 2,451 | +66% | 0 | 0 | — |
case-12 | fail→fail | 13,866 | 12,601 | -9% | 1 | 1 | 0% | 1,872 | 2,441 | +30% | 0 | 0 | — |
case-13 | pass→pass | 11,906 | 4,078 | -66% | 1 | 1 | 0% | 1,671 | 2,498 | +49% | 0 | 0 | — |
case-14 | fail→pass | 10,986 | 7,490 | -32% | 1 | 1 | 0% | 1,593 | 2,799 | +76% | 0 | 0 | — |
case-15 | fail→pass | 14,865 | 2,564 | -83% | 1 | 1 | 0% | 1,490 | 2,388 | +60% | 0 | 0 | — |
case-16 | pass→pass | 11,954 | 8,546 | -29% | 1 | 1 | 0% | 1,979 | 3,301 | +67% | 0 | 0 | — |
case-17 | fail→fail | 13,775 | 3,527 | -74% | 1 | 1 | 0% | 1,946 | 2,438 | +25% | 0 | 0 | — |
case-18 | fail→pass | 25,846 | 3,067 | -88% | 1 | 1 | 0% | 2,151 | 2,374 | +10% | 0 | 0 | — |
case-19 | pass→pass | 12,562 | 5,442 | -57% | 1 | 1 | 0% | 1,795 | 2,781 | +55% | 0 | 0 | — |
case-20 | pass→pass | 10,423 | 3,827 | -63% | 1 | 1 | 0% | 1,462 | 2,428 | +66% | 0 | 0 | — |
case-21 | pass→pass | 11,369 | 3,547 | -69% | 1 | 1 | 0% | 1,438 | 2,433 | +69% | 0 | 0 | — |
case-22 | pass→pass | 16,286 | 8,782 | -46% | 1 | 1 | 0% | 2,356 | 3,264 | +39% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 18 counted toward the lift figure. The other 4 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +14 percentage points is the difference between those two pass rates over the 18 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.