name: motion2codedescription: >- Rebuilds an AI-generated (or any) flat motion design video as clean, editable code: inline SVG shapes animated by one GSAP timeline, seekable frame by frame, looping seamlessly, checked side by side against the source; and optionally turns an interface recording into a real interactive website (GSAP ScrollTrigger, real clicks). Ships scripts to study the video (frames, cuts, contact sheet, exact palette), render the code to mp4/GIF through anim.seek, compare it with the source frame by frame and check the loop seam and page contract. Use when the user wants a video, GIF, Lottie-like clip, motion design shot, animated logo, UI screen recording or AI video "recreated in code", "converted to SVG/GSAP", "made into an HTML animation" or "made into a real site". Not for photographic footage, 3D rendering or video editing.license: MITmetadata: hermes: tags: - svg - gsap - motion-design - animation - video-to-code - scrolltrigger category: development
Motion2Code
The video is the spec, and you cannot watch it. Treat the job as reconstruction: turn the clip into images you can
look at, write down what happens and when, build the shapes, drive them with one clock, then prove every moment
against the source with images side by side. Never animate from a memory of the video.
Workflow
- Study the source.
python3 scripts/study_video.py source.mp4 -o study/ saves a frame every 0.25 s, the
frames around every cut, a contact sheet, the exact flat palette and whether the source already loops. Open the
sheet, then the frames. Read the prompt file next to the video if there is one: it holds the exact words.
- Write the shot list. One line per beat: time range, what each shape does, how it eases, what it becomes
next. Note the stage size (1920x1080 or 1080x1920) and the real duration (ffprobe, e.g. 15.10 s means a 15 s
loop). See
references/study-the-source.md.
- Build the stage from
templates/index.html: one inline SVG at the source resolution, every shape drawn
around its own origin, flat fills from the sampled palette, the closest Google Font subset to the exact letters.
- Animate on one GSAP timeline. Tweens for the shots, plus a pure
draw(t) for procedural parts (particles,
morphs, physics) called from the same clock. Match the timing of the frames, not your idea of it.
See references/animation-architecture.md and references/motion-vocabulary.md.
- Close the loop in code. The last moment returns to the exact opening frame, even when the source does not.
See
references/seamless-loop.md.
- Verify, fix, repeat.
python3 scripts/frame_diff.py source.mp4 index.html -o diff/ writes source|code pairs and a score per time;
open the pairs, fix the worst first, then zoom on transitions with --every 1/24 --start a --end b.
python3 scripts/check_anim.py index.html --runtime checks the contract, the loop seam, determinism, console
errors and network. Ship only when both are clean. See references/verification.md.
- Deliver:
index.html, compare.html (from templates/compare.html), and when asked the proof videos:
python3 scripts/render_anim.py index.html -o anim.mp4 then
python3 scripts/side_by_side.py source.mp4 anim.mp4 -o side_by_side.mp4 --html index.html.
- Optional real site. For an interface recording, after the faithful replay, build
site.html from
templates/site.html: scroll becomes ScrollTrigger, fake clicks become real buttons.
See references/interfaces-and-sites.md.
Scripts need Python 3.9+, ffmpeg, and Playwright for anything that opens a browser
(uv run --with playwright python scripts/..., or pip install playwright && python -m playwright install chromium).
Run them from this skill's directory or with full paths.
The page contract
window.anim = { tl, duration, seek(t), play(from), pause() } // seek(t) is a pure function of t
index.html?t=6.5 freezes at 6.5 s; ?external=1 does not autoplay (compare.html and the scripts drive it);
with no parameter it plays by itself and loops forever.
<html data-ready="1"> once fonts are loaded and frame 0 is drawn. Renderers wait for it.
frame(duration) equals frame(0); frame(duration - 1/24) is one step away from it.
- Self-contained: inline SVG, GSAP 3 from cdnjs, flubber from jsdelivr only when a morph needs it, Google Fonts.
Nothing else is loaded.
Hard rules
- Use only the given source video. Never regenerate it, never fetch "a better version", never trace a frame
into the page. No
<img>, <video>, SVG <image>, canvas drawImage or raster data URI in the animation.
- Words exactly as in the source prompt (or the video if there is no prompt), even when the video misspells
them. Invented gibberish words in UI recordings become flat bars or are removed.
- Flat shapes as SVG. Solid fills, the sampled hex colors, no gradients, shadows or filters the source does not
have. Strokes with
stroke-linecap="round" for sticker outlines.
- GSAP timeline is the single clock. No
requestAnimationFrame loops, setTimeout choreography, CSS
animations or Date.now() driving visuals. Randomness is seeded and precomputed.
- The loop is done in code. If the source does not return to its opening frame, add a short clean transition
back to it. The seam is checked numerically.
- Every moment is checked against the source side by side, at least every 0.5 s and frame by frame around
every transition and the loop seam. "Looks right" without a pair image is not a check.
- Same timing as the source. Measure from the frames; do not compress, pad or reorder beats.
Reference routing
Read only what the task needs:
Templates: templates/index.html (the contract, a tiny looping example), templates/compare.html (video and code
side by side, synced to the video clock, frame stepping, onion skin), templates/site.html (ScrollTrigger pin and
scrub, menu, tabs, flip card, draggable slider, cart counter, keyboard, reduced motion).
Output expectations
- Deliver
index.html (+ compare.html, + site.html for interfaces) and say what was checked and how:
frame_diff times, worst scores and what they were, check_anim.py --runtime result, page size.
- List honestly what differs from the source (a simplified texture, a merged beat), never "pixel perfect".
- Keep names readable: one group per character or object, ids that say what they are, the palette as constants,
shots commented with their time range, so a designer can retime or recolor without reading everything.