huashu-motion: an open skill that writes the launch film, scores it, and lets you point at the frame to fix it

花叔 (Huashu, @AlchainHust), the developer behind huashu-design, published huashu-motion on Oct 18 under the MIT licence. His launch post passed 840 likes, and the repo had 1,120 stars and 116 forks when we checked this morning. That's two days old. Daily repo pickers in Japanese, Korean and English picked it up, and people are already posting films they made with it. He says he built it because Claude Motion (Oct 8) launched for Team and Enterprise only. His pitch is about storytelling, not effects: most motion skills polish the animation and forget what the product is supposed to say.

The flow, per the README: ask for 'a 30-second launch film for my product' and give it your site. The agent reads the product, shows three directions as key frames, writes a rhythm sheet (a hook within 3 seconds, a burst in the middle, a beat of silence before the biggest hit), builds the film and scores it. Then it opens a browser studio. Press C, click the spot that's wrong and type the change. The note carries the time, the scene, the object under your cursor and its source line, so the agent changes only that and the preview reloads where you were. Every version is kept, and titles, numbers and colours show up as editable fields.

Sound is what sets it apart. In composer mode the agent writes the score in ABC notation, with marks that land on scene changes. After a one-time om samples build (about 55 MB of CC0/CC-BY samples), a sampled piano, strings and drums play it; without the pack, synths do. Sound effects land on the frame where each move lands. Narration can use macOS voices, Kokoro, or your own recording aligned word by word, with the music ducked under the voice. Each film is a single index.html (GSAP, Canvas, Three.js, SVG). It exports MP4, transparent MOV, WebM, GIF or a share page, and has a mode that packages talking-head footage with name plates and captions as a transparent overlay for CapCut or Premiere.

The README has some honest notes. It can't hear its own music: it checks the score, the energy curve and a spectrogram, and the final listen is yours (pair it with earshot if you want an agent check). It doesn't generate people or live footage. Results depend on the model. The author mostly uses Opus 5.5, but he says a 22-second film made with Claude Code running on DeepSeek V4.1 Flash took under half an hour with three rounds of notes and cost ¥1.95. Requirements: Node 22+, Chrome, Chromium or Edge, and FFmpeg. It's tested on macOS only so far, and Codex's sandbox will ask before it opens the preview. Licensing: the code is MIT, GSAP keeps its own licence, and the cartoon 'bean Huashu' character is the author's likeness and is excluded from MIT. Install with npx skills add alchaincyf/huashu-motion. Spotted, not tested. Credit 花叔.