huashu-motion: The MIT Agent Skill That Makes Launch Films and Lets You Click a Frame to Edit

A new open-source agent skill is climbing GitHub with a simple promise: tell your coding agent "make a 30-second launch film for my product" and get a finished video with a composed score, then click on any frame to change just that one thing. huashu-motion, by the developer Huashu, had about 1,300 stars and 132 forks when we read its GitHub README, and it works with Claude Code, Codex and other agents that follow the Agent Skills format.

This guide covers what it does, how to install it, what it needs, the one cost figure the author gives, and where it falls short. It sits in the same space as Claude's Motion beta for animated explainers, the OpenFilm HTML film studio and the Lemo skill that makes short films without a video model.

TL;DR

What it does

The README's description: you tell the agent to make a 30-second launch film. It first reads your product, offers three directions to choose from, writes the full film, composes the music, then opens an editor in the browser. Where you are unhappy, you click on the frame and write what to change. The comment goes to the agent with the timestamp, scene, the clicked element and its position in the source, and the agent looks at that frame, changes only that spot and replies in plain language. Every version is kept, Cmd+Z undoes, and titles, numbers, colors and durations appear as input fields you can edit yourself without waiting.

The film itself is an index.html: GSAP, Canvas, Three.js and SVG, anything a browser can draw. There is no private format, and the README stresses you can take the page away and it still works. The loop is simple: the agent writes the HTML, the studio shows it, your clicks land in an inbox file that the agent reads.

Install and requirements

1. Install the skill. Either ask your agent to install github.com/alchaincyf/huashu-motion and run the demo, or run npx skills add alchaincyf/huashu-motion. Add -a claude-code or -a codex to target an agent. To clone instead, symlink skills/huashu-motion into ~/.claude/skills/.
2. Check the environment. Run om doctor. It needs Node 22 or later, a local Chrome, Chromium or Edge for frame capture, checks and export, and FFmpeg for video export and audio mixing. Animation runs on GSAP, which keeps its own license.
3. Optional audio. For real instrument samples (grand piano, strings, drums) run om samples build once; it downloads about 55 MB of CC0 and CC-BY samples. Without it you get synthesized timbres. Voice-over can use macOS system voices, or Kokoro for Chinese text (needs uv, Python 3.12 and about 325 MB of models) with whisper-cli for word alignment.
4. Platform note. The README says it is only tested on macOS. Codex's sandbox can block the local preview and browser, so approve manually when prompted.

If you are new to the format, our complete guide to agent skills explains how a skill is loaded and invoked.

What you can ask for

After export, om share produces a share pack: a GIF under 10 MB plus a share.md with post copy and README header code.

How it tries to be good, not just working

The author's "director rules" are the most interesting part. Per the README, they come from measuring 21 real launch films (naming Apple, Nothing, GoPro, Google, Vercel, Linear and Anthropic) with FFmpeg for cuts, beats, loudness and energy curves, and from breaking down 142 high-engagement launch and motion films into 20 motion languages. Each film starts with a beat sheet: a hook within 3 seconds, one burst in the middle, and a pause before the loudest moment. Colors are taken from the product rather than chosen freely, with an om swatch command to show palettes and contrast.

Sound is treated as half the film. Composer mode has the agent write a score in ABC notation with markers aligned to scenes. There are also 12 fallback score directions by product type, three annotated example scores, and 23 sound effects placed on the same line of code where an animation lands.

Quality checks run before delivery: om check for script errors, missing files and rule-breaking animation, plus sampled frames for clipped or low-contrast text; om qa for single-frame flashes, dead gaps and the motion curve. The skill asks the agent to look at real frames and watch at normal speed, and to use an independent reviewing agent when it can.

The cost claim, and how to read it

The author describes one 22-second film made by swapping the model behind Claude Code to DeepSeek V4.1 Flash. It took three rounds of feedback and under half an hour, and cost 1.95 yuan at the time, with a 99.4% cache hit rate. The README concludes cheap models can produce decent films because the method is in the skill.

That is a self-reported figure for one project, with a heavily cached workflow, and the author also says taste differs across models and that most work was done on Opus 5.5 in Claude Code. Treat the number as an indication that token cost is low relative to editing time, not as a guarantee.

Honest limits, from the README

• It cannot hear. The agent checks scores through notation, energy curves and spectrograms, so a human ear is the last check.
• No real people or live footage. Visuals are drawn by code. Bring your own footage and it will composite and package.
• Model-dependent. Results vary with the model behind the agent.
• Needs a local environment. Preview and export run on your machine.
• Branding and licensing quirks. Films over 4 seconds show a small "Made with huashu-motion" line in the last 1.2 seconds unless you tell the agent to turn it off. A cartoon character in a demo template is excluded from the MIT license.

Compared with other options

Video generation models produce photographic motion; code-drawn films produce crisp, editable graphics. For product launches, UI demos and data stories, the second often fits better, and editing a number is trivial because it is text.

Safety and trust notes

You are installing a skill that runs shell commands, drives a browser and calls FFmpeg. Read SKILL.md and the CLI before running it, keep it in a project directory rather than your home folder, and avoid giving it credentials. The README says installation opens no browser and performs no inference, and preferences are stored only on your machine. As with any trending repo, pin a commit rather than tracking main. For running agents with less trust, see our look at how Auto Mode monitors fail, and consider AgentBeam, the agent security platform from the explainx.ai team.

The backstory

The README says the project began after Anthropic released Claude Motion on October 8, 2026, with the author telling Claude Code the skill should fully reproduce its abilities and be much stronger. The first day produced an editor, version management and over a hundred tests but a boring 19-second sample, so the author discarded it and restarted, then spent the time measuring real launch films and turning each "not cool" reaction into a rule. This is the author's account, not independently verified.

Related reading

• Complete guide to agent skills
• Claude Motion beta: animated explainers
• OpenFilm open-source HTML film studio
• Lemo: short films with no video model
• MG Styles: 15 motion films with Claude Opus 5.5
• Claude Duet launch
• Official: huashu-motion on GitHub · Agent Skills registry on explainx.ai

Details are from the repository README at the time of writing; the project is young and may change quickly.