anim-skill · End-to-End Animation Generation

One topic in, one video out. Frames are fully drawn by code (Remotion + React), with no dependency on stock libraries, generative video models, or external images.

Style is a plugin, voice is a plugin, the pipeline is fixed. Adding a new style does not change the pipeline; swapping the voice engine does not change the pipeline.

When to Use

"Create an explainer animation for X"
"Turn this document/lesson plan into a video"
"Add narration to this video"
"Remake it in X style"
"Check this video for flickering or empty scenes"

Do not use for: live-action footage, talking-head videos, 3D rendering (use Blender for that).

Input: Two Modes, Determine First

Mode	Trigger	Extra Step
A · Topic-Driven	User gives only a topic keyword or phrase	First do research (facts, figures, common misconceptions), then proceed to design
B · Document-Driven	User already provides a topic document (lesson plan, white paper, PPT script)	First do refinement and correction (fill data, unify terminology, mark paragraphs that cannot be visualized), then proceed to design

After determining, restate the plan before acting: I am in A/B mode, making an N-second video, in style X, for audience Y. If two of these are unclear, ask before guessing.

Six Stages (Order Is Fixed, No Skipping)

0 PREFLIGHT → 1 Content → 2 Storyboard → 3 Build → 4 Voiceover → 5 Accept & Deliver
                                    ↑                                 ↓
                                    └── Rework ←───────────────────┘

0 · PREFLIGHT (One-Time, 2 Minutes)

node <skill>/scripts/preflight.mjs

Check node / npm / ffmpeg / browser; list available voice engines on this machine. If anything is missing, state it clearly. Don't wait until rendering to find out.

1 · Content (A: Research / B: Refine)

Output research.md: one-sentence thesis, 3–6 supporting facts, figures with sources, mistakes to avoid, audience knowns and unknowns.

Mode A: Fact-checking first. Every figure must have a source; if not found, mark as "pending confirmation" — never invent.
Mode B: Keep the original text, only supplement. Mark sentences that cannot be visualized (animation can only depict actions, comparisons, changes).

2 · Storyboard

Output plan.md (format in references/storyboard.md): one line per shot — frame range, visual, motion primitives, subtitle.

Hard criteria (rewrite if not met):

Each shot has one protagonist, occupying ≥1/3 of the content area height (graphic ≥170px, large text ≥96px and full line ≥300px)
The protagonist enters on beat 1 or 2
Only three size classes: protagonist ≥170 / supporting 60–110 / annotation 22–30
At least one "highlight moment" per video (thesis sentence, protagonist entrance, closing payoff — pick one). Do not pad with decorations.

3 · Build

node <skill>/scripts/scaffold.mjs <project-dir> --title "Title" --style <style-id>

Following the style brief (styles/index.json), write src/scenes/*.tsx and connect them with <Series>.

Iron Rules (violation will break rendering or crash the video):

Everything must derive from useCurrentFrame(). Forbidden: Math.random(), Date.now(), setState counters. For randomness, use deterministic hashing: hash(seed) = fract(sin(seed*127.1+311.7)*43758.5453).
Every interpolate() must include {extrapolateLeft:'clamp', extrapolateRight:'clamp'}. Miss one, and frames outside the shot will fly out.
No CSS transition / animation / @keyframes. Use only interpolate and spring for motion.
No external assets or heavy libraries (three.js / GSAP / images). Draw everything with SVG and CSS gradients.
Maximum 12 SVG blur primitives per frame, maximum 15 filter: blur() elements, or rendering will be ten times slower.

4 · Voiceover (Optional)

node <skill>/scripts/srt.mjs <project-dir>            # Generate SRT from storyboard
python <skill>/scripts/narrate.py <video> --provider edge|volc --voice <id>

Can deliver without voiceover (e.g., for classroom use when teacher explains).
With voiceover: generate SRT first, then synthesize. Timeline is derived from rendered frame numbers, not estimates.
Engine is a plugin; see references/voice.md.

5 · Acceptance and Delivery

python <skill>/scripts/qc.py <video>          # Empty scenes / object size
python <skill>/scripts/qc_flicker.py <video>  # Flicker / hard cut detection

Two hard metrics (see references/qc.md):

Empty scene: largest object in content area <110px for >45 frames → defect; <80px → blocker
Flicker: luminance offset from local median >6 (0/1 jump) → defect

Delivery package: <slug>.mp4, <slug>.srt, plan.md, research.md, src/ (editable project), README.md (how to modify, how to re-render).

Style Plugins

styles/index.json contains 15 built-in styles. Each is an executable brief (background color, ink color, seven-color palette, stroke width, jitter, glow, texture, font, character, projection readability, suitability).

Usage: --style neon / --style paper / --style crayon …

Adding a new style: append one entry to index.json and place a brief in styles/. No need to change the pipeline, scripts, or templates. See the extension contract in references/styles.md.

Style selection tip: first filter by "projection readability" (can the back rows see it?), then by "character". Paper/crayon suit young children; neon/minimal suit classrooms with poor projectors.

Voice Plugins

Two built-in providers:

Provider	Cost	Requirements	Notes
`edge`	Free	None	Microsoft Azure neural voice, more broadcast-like
`volc`	0.0003 CNY/character	APPID + API Key	Doubao voice, close to human; free credits for new accounts

Adding a new engine: implement synthesize(text, voice) -> mp3 bytes and register it in scripts/tts/<provider>.py. narrate.py does not need changes. See references/voice.md.

Pre-Delivery Self-Check (Check Each Item)

remotion render exit code 0, ffprobe frame count = design frame count
qc.py has no blockers; qc_flicker.py has no peaks except at chapter hard cuts
All interpolate calls include clamp (grep -c "interpolate(" = grep -c "extrapolateLeft")
No Math.random / Date.now / CSS animation
Subtitles match visuals (check 3 beat frames)
Chinese fonts exist on this machine (otherwise rendering will show squares)

Quick Reference

node scripts/preflight.mjs
node scripts/scaffold.mjs ./videos/my-film --title "Title" --style paper --seconds 20
cd videos/my-film && npm install
node <skill>/scripts/render.mjs . main out/my-film.mp4
node <skill>/scripts/srt.mjs .
python <skill>/scripts/narrate.py out/my-film.mp4 --provider volc --voice zh_female_cancan_uranus_bigtts
python <skill>/scripts/qc.py out/my-film-volc.mp4

Details in references/: pipeline.md (each stage expanded), storyboard.md (storyboard template), styles.md (style catalog and extension), voice.md (voice and extension), qc.md (acceptance criteria), render-env.md (environment pitfalls: Windows sandbox, TEMP, browser path, ffmpeg parameters).

代码逐帧动画短片