siweimidu/code-video-infinite

code-video-infinite 是做什么的？

Frames aren't "generated" — they are "written". There is no image/video generation model here: a video is either an HTML page or a Remotion project. Give it a time t, it draws that frame; headless browser or Remotion renders frame by frame, ffmpeg composites audio. Sound can also be code-synthesized. Therefore every video can be arbitrarily seeked, re-rendered at will, and a new version never drifts from the previous one.

This skill is the merger of five engines: one entry point, one pipeline, three gates — each engine retains its full depth documentation and assets.

0 · Run these two first (30 seconds, don't skip)

node scripts/doctor.mjs          # Environment: Node / ffmpeg / ffprobe / Python / Chrome / all five engines present
node scripts/find.mjs --stats    # Directory: 880 assets counted by source and collection

If doctor shows FAIL, fix the environment before creating. Install Chrome or npx playwright install chromium if Chrome is missing; on Windows, if npm is blocked by execution policy, use npm.cmd.

1 · Pick the engine first, then discuss style

The user's requirements decide the engine, not the richness of the catalog. A video uses only one main engine by default; the second engine is only used to fetch assets.

What the requirement looks like	Engine	Its strength	What you pay out of the box
15–30s product/launch/skill motion reel, needs rhythm, camera movement, "not PPT-like"	onetake	Single HTML composition, real shutter motion blur, synthesized sound effects, rhythm and transition oracle	Python + Chrome; 4K60 rendering is slow (estimate time first)
Product promo, cinematic feel, real web page footage, multi-shot narrative, needs to edit in CapCut later, for social platforms	shotcraft	124 shot cards + Ink Press template + real page capture + Freesound audio selection + workbench	Node + Remotion (npm install); Remotion has its own license
Specific shapes, materials, 3D objects, characters & animals, hand-drawn papercraft, runnable scenes	claude-motion	47 source code cases + 20 source components + 30 object models + 14 scene starters, fidelity source code not style tags	Node ≥ 22 + Playwright Chromium + ffmpeg; large asset volume
MG style (flat/line-art/isometric/3D-rendered/frame-by-frame/collage/liquid/morph/Bauhaus/Synthwave/glass/flower-letter/sticker/pixel/cyber-HUD)	mg-motion	15 verifiable signature techniques + 103 sound fx toolkit + 10 machine gates + 2 rounds of review	Requires Chrome (MG_CHROME) + ffmpeg + self-bootstrapped fonts
Only need a single motion, transition, or text animation's real source code for use in another project	wise-motion	462 motion entries, previews, real source code entry points, and export scripts	Only hands over assets; does not design the full video for you

Routing details, disproven combinations, and clear boundaries for each engine are in references/engine-selection.md.

Search any asset with the same command (across all five engines):

node scripts/find.mjs "HUD radar"
node scripts/find.mjs "coffee hand-drawn" --limit 5
node scripts/find.mjs fade --source wise-motion
node scripts/find.mjs "title entrance" --collection shot-card --json

2 · Six-stage pipeline (order fixed; stages can be removed but gates cannot be skipped)

Stage	What to do	Deep reference
0 Requirements	Ask for delivery specs (duration / aspect ratio / fps / voiceover / language), asset sources, audience & placement; ask everything once, don't follow up one by one	references/pipeline.md §0
1 Reference	If a reference video exists, analyze it: energy graph, cut points, static ratio, t80 & motion blur (onetake analyze_ref.py); if no reference, write the next set of verifiable numeric targets	references/motion-grammar.md
2 Concept	Three distinct central ideas, user picks one; simultaneously select a look/style by name, never silently default	onetake §2 / mg-motion §2
3 Rhythm & Transition	Storyboard designed with both rhythm and transition legs: no equal-length shots, no beat replacing the previous beat	references/motion-grammar.md
4 Assets & Shots	Real page capture / screenshot UI restoration / object models / shot cards / motion entries; first grab the smallest runnable snippet	Each engine's ENGINE.md
5 Implementation	Every frame is a pure function of time; actions picked from library, curves set by t80; freeze first, then full video	Each engine's ENGINE.md
6 Render & Delivery	1080p30 draft → approval → 4K60/final; pass three gates; report per delivery contract	references/delivery.md

3 · Iron rules (deduplicated across five engines; each is a lesson learned)

1. Every frame depends only on t. Forbidden: Date.now(), rAF cumulative state, CSS animations triggered by state; use fixed-seed pseudo-random.
2. Don't mistake the template's story for the user's story. Pick one of three concepts; a new video using a template's beat will read as a template.
3. Don't carry the look of the previous video into the next one. Look/style must be chosen by name; silent defaults are not allowed.
4. Don't cut equally by metronome. Equal-length shots — no matter how good individually — read as a slideshow.
5. Don't replace a beat; transition into it. At every segment boundary, something must survive and visibly move/scale into the next segment; mutually exclusive scenes (display:none page switch) are PPT even with perfect rhythm. Hard cuts (explosive beats & end cards) excepted.
6. White space is where the punchline lands. Reference videos are roughly 1/3 static; a non-stop video has no climax.
7. Small elements, big ground. These videos often have 70% white / background; full-frame content is the exception, not the default.
8. Cause must be visible. Hand drawing a path, ball landing, point splitting — if the page animates on its own it reads as a screensaver.
9. Fast motion must have shutter. Above ~80 px/frame without motion blur causes strobing; use the renderer's shutter, not CSS blur.
10. Camera must follow the subject. Static camera gets reviewed as "lacking lively tracking"; declare inFrame, let the camera lag, tilt, settle.
11. Draft at 1080p30. Unapproved cuts are never rendered at 4K60.
12. No particles/confetti/glitter as decoration. It reads as template; if you want flair, let the letters land themselves.
13. Sound: one room, one material set. One synth per event = audio slideshow.
14. No faking: unreferenced examples are not reproduction; silent videos are not talking videos; technical pass is not aesthetic pass.

4 · Three gates, all mandatory

Gate	Who judges	What is judged	If failed
Machine gate	node scripts/qc.mjs gate <video>	Duration / aspect ratio / fps / audio track presence / LUFS / true peak / black frames / static ratio	Re-render the video; don't lower the standard
Keyframe gate	You yourself read_image	Can 6–10 keyframes serve as posters: composition / depth / white space / typography	Rearrange the frames
Review gate	Independent subagent with clean context	7-dimension scoring + frame-numbered fix list; re-review after fixes	Fix by impact order

The machine gate only measures what machines can measure; aesthetics are always left to independent review, as the creator's self-check suffers from confirmation bias. Each engine's own gates (mg-qc.mjs, verify_promo.py, final-review.md, jury.md) usage and criteria are documented in references/quality-gates.md.

5 · Sound

• Goal: one room, one material set; event count less than beat count; music defaults to commercially licensed tracks only, with royalty tracking.
• Common targets: −14 LUFS (social/web), −16 LUFS (long-form/podcast-style), true peak ≤ −1 dBTP (AAC encoding raises it); for more headroom target −3.5 dBTP.
• Ready tools from three engines: onetake sfx_palette.py / score.py (synthesis + recording mixdown), mg-motion audio.py + 103 samples, shotcraft sound-design.md + sfx-source.py (Freesound, CC0 by default).
• Details and licensing red lines in references/sound.md.

6 · Delivery contract

Every video must include these seven items (missing one means incomplete): two-sentence concept / timecoded beat sheet / signature techniques mapped one by one / which styles were used and how they transition / technical approach / measured numbers (duration aspect ratio fps LUFS true peak black frames static ratio) / honest top-3 weaknesses.

Engine-specific deliverables: onetake multi-language same-source (?lang= no fork), claude-motion ADAPTATION.md (retained/rewritten/licensed/unresolved issues), shotcraft motion workbench & CapCut project, mg-motion version directories v1/ v2/. See references/delivery.md.

7 · Directory map

Code-Video-Infinite.skill/
├── SKILL.md                  # This file: routing + iron rules + gates + delivery
├── README.md                 # For humans: what the five engines are, how to install, license table
├── LICENSES.md               # Per-engine licenses and third-party declarations
├── references/               # Merged shared knowledge (newly written at this level, only cross-engine decisions)
│   ├── engine-selection.md   # Decision matrix and boundaries
│   ├── pipeline.md           # Depth of each of the six pipeline stages
│   ├── motion-grammar.md     # Unified vocabulary for rhythm/curves/transitions/camera
│   ├── sound.md              # Sound and licensing red lines
│   ├── quality-gates.md      # Criteria for the three gates
│   ├── delivery.md           # Delivery contract
│   ├── assets-and-licenses.md# Where assets are, can they be redistributed
│   └── provenance.md         # Five sources, versions, what merging did
├── catalog/
│   ├── INDEX.md              # 880 asset list (generated)
│   ├── index.json            # Machine-readable version of same data
│   └── assets-index.json     # 537 heavy asset list
├── scripts/
│   ├── doctor.mjs            # Environment self-check
│   ├── find.mjs              # Cross-engine search
│   ├── qc.mjs                # Unified machine gate (measure / gate / delegate)
│   ├── build-catalog.mjs     # Rebuild index
│   └── check-links.mjs       # Documentation cross-link health check
└── engines/                  # Five engines preserved as-is (each with its own docs/assets/license)
    ├── onetake/       ENGINE.md + lib/motion.js + scripts/ + looks/ + cases/
    ├── shotcraft/     ENGINE.md + references/shots(124) + template/ + demos/ + assets/
    ├── claude-motion/ ENGINE.md + references/*.json + assets/{cases,extensions,models,scenes}
    ├── mg-motion/     ENGINE.md + mg-styles-15/{harness,assets} + references/ + scripts/
    └── wise-motion/   ENGINE.md + catalog/registry.json + references/effects(462) + scripts/

8 · What to read when

When	Read
Don't know which engine to use	references/engine-selection.md
Ready to start, need specific actions & commands per stage	references/pipeline.md
Rhythm is off, looks like PPT, motion too soft/too hard	references/motion-grammar.md
Voiceover / sound effects / background music / loudness	references/sound.md
Video finished, worried about failing gates	references/quality-gates.md
Need to deliver, write a report, or re-edit	references/delivery.md
Unsure if an engine's assets are commercially usable	references/assets-and-licenses.md
Want to upgrade an engine or query its origin	references/provenance.md
Engine depth (single source of truth)	engines/<engine>/ENGINE.md and its references/

9 · License & asset boundaries (one-liner)

The five engines have different licenses; merging the repository does not change each one's license: onetake is PolyForm Noncommercial (free for non-commercial use), wise-motion is Apache-2.0, claude-motion and the upstream mg-styles-15 are MIT (third-party assets retain their own licenses), shotcraft has its own license and Remotion has a separate license (commercial company use must be verified independently). Fonts, models, sound effects, BGM authorization is recorded per entry in references/assets-and-licenses.md and each engine's LICENSE / ATTRIBUTION / NOTICE files. Brand names are always fictional; don't use real trademarks.