huashu-motion: let a coding agent write your launch film, then point at the frame to fix it

The usual way to get video out of an agent is to call a video model: pay per generation, give up control of the frame, and regenerate everything to fix one word. huashu-motion takes another route. The agent writes the whole film as a single index.html using GSAP, Canvas, Three.js or SVG, so every frame is a function of code and time, and the score is synthesized on the local machine. The repository was created on October 18, 2026, is MIT-licensed, and had 226 stars as of October 19, 2026.

Core features

• One sentence to a finished cut: tell Claude Code, Codex or another agent “make a 30-second launch film for my product” and it reads the product, offers three key-frame directions, writes a rhythm sheet, writes the film, composes music, then opens an editor in your browser.
• Point at the frame to change it: in the editor, press C, click an element and type what is wrong. The comment carries the timestamp, scene, clicked element and its location in the source; the agent zooms into the patch you pointed at, changes only that, and refreshes the preview in place. Every version is kept and undo is one keystroke away.
• Score first, then shoot: in composer mode the agent writes an ABC-notation score whose scene markers line up with the cuts, played back on piano, strings and drums; sound effects land on the frames where animation resolves. All of the audio is generated locally, with no paid API.
• Directing rules from measurement: the author used FFmpeg to measure cut points, beat, loudness and energy curves across 21 real launch films, and deconstructed 142 high-engagement films into 20 motion languages. The rules are concrete: a hook within three seconds, one burst in the middle, and a held beat before the loudest moment.
• A check pass before delivery: om check validates scripts, missing files and disallowed animation patterns before rendering, then samples frames for clipped, occluded or low-contrast text; om qa looks for single-frame flashes, dead stretches and hard cuts in the finished video.

Typical use cases

• Small teams shipping a launch film: no budget for a motion team, but you need 30 seconds for the homepage or a Product Hunt page.
• Data explainers: hand it a CSV and it follows “axis first, then data, label one thing” so the change in a metric is legible; numbers become editable inputs in the studio.
• Packaging an existing voice-over: word-level alignment puts the picture on the exact word being spoken, adds captions automatically, and ducks the music under speech.
• Deliverables that stay editable: the output is a web page, not a rendered file, so it can live in a repo, be diffed, and be handed to the next agent.

Know the boundaries up front. It has only been tested on macOS, and needs Node 22+, Chrome and FFmpeg. The real-instrument sample pack is optional; without it you get synthesized timbres. Codex’s sandbox blocks local preview and the browser, so expect to approve those manually. It does not produce photoreal footage — the footage is drawn by code, and a video model is still the tool when you need live action or photographic realism.

Quick start

npx skills add alchaincyf/huashu-motion

Then tell your agent: “Use huashu-motion to make a 30-second launch film for my product, the site is xxx.com.” At runtime, om doctor checks the environment, om studio opens the editor, and om render exports MP4, MOV with alpha, WebM or GIF.

Summary

huashu-motion fits teams that want the launch film inside the code workflow: the film is code, one change touches one line, and music and effects are generated locally, which makes cost and reproducibility more controllable. The trade-off is that it cannot draw live-action texture and it depends on the local toolchain (Chrome, FFmpeg, Node 22+), with a browser sample download on first use. It belongs to the same “draw every frame in code” school as the animate skill we covered, but its distinguishing feature is the editor-and-agent revision loop. MIT-licensed, primarily JavaScript, stats checked October 19, 2026.