905 ·GitHub
What is this?
huashu-motion is an Agent Skill that lets coding agents like Claude Code or Codex turn one sentence ('make my product a 30-second launch video') into a full commercial-grade video: the agent researches the product, writes the script and an original score, renders the film, then opens a browser editor so you can click on any frame and ask for a change.
huashu-motion is an Agent Skill that lets coding agents like Claude Code or Codex turn one sentence ('make my product a 30-second launch video') into a full commercial-grade video: the agent researches the product, writes the script and an original score, renders the film, then opens a browser editor so you can click on any frame and ask for a change.Is this a standalone video editor or something that plugs into my coding agent?It's a skill package following the Agent Skills protocol, not a standalone app — it installs into agents such as Claude Code, Codex, Grok Bot, Muse and Dots, and its bundled bin/om CLI does the actual rendering and compositing.
It's a skill package following the Agent Skills protocol, not a standalone app — it installs into agents such as Claude Code, Codex, Grok Bot, Muse and Dots, and its bundled bin/om CLI does the actual rendering and compositing.What do I need installed before trying it?Node 22 or newer, and a Chrome/Chromium/Edge browser already on your machine for frame capture and export (no separate renderer is downloaded); FFmpeg is required for exporting video and mixing audio. Real instrument samples (~55MB) and Kokoro/whisper-cli for voiceover are optional extras.
Node 22 or newer, and a Chrome/Chromium/Edge browser already on your machine for frame capture and export (no separate renderer is downloaded); FFmpeg is required for exporting video and mixing audio. Real instrument samples (~55MB) and Kokoro/whisper-cli for voiceover are optional extras.Any catches before I rely on this?It's only been tested on macOS so far, and Codex's sandbox blocks local preview and browser access, requiring manual approval; it can't generate real actors or live footage since every frame is code-drawn, and the agent can only judge audio by waveforms and spectrograms — you still have to listen yourself.
It's only been tested on macOS so far, and Codex's sandbox blocks local preview and browser access, requiring manual approval; it can't generate real actors or live footage since every frame is code-drawn, and the agent can only judge audio by waveforms and spectrograms — you still have to listen yourself.Does it rely on a paid video-generation AI model?No — every frame is drawn live in the browser with code (GSAP, Canvas, Three.js, SVG) based on time, and the music is performed locally from real instrument samples, so no paid generation API is called.
No — every frame is drawn live in the browser with code (GSAP, Canvas, Three.js, SVG) based on time, and the music is performed locally from real instrument samples, so no paid generation API is called.How does the point-and-click editing actually work?Press C and click anywhere on the frame; the comment is sent with the timestamp, scene, element and its source-code location, the agent zooms into that exact spot before editing, and only that one thing changes while every version stays undoable with Cmd+Z.
Press C and click anywhere on the frame; the comment is sent with the timestamp, scene, element and its source-code location, the agent zooms into that exact spot before editing, and only that one thing changes while every version stays undoable with Cmd+Z.Can I use a cheaper model than Claude Opus with this skill?Yes — a full launch video was made by swapping in DeepSeek V4.1 Flash instead of Opus, cost about 1.95 yuan with a 99.4% cache-hit rate, and the skill's style rules still held without extra prompting.
Yes — a full launch video was made by swapping in DeepSeek V4.1 Flash instead of Opus, cost about 1.95 yuan with a 99.4% cache-hit rate, and the skill's style rules still held without extra prompting.What's inside the skill besides the renderer?20 documented motion-language styles distilled from 142 real launch videos, 10 starter templates, 23 sound effects, 12 fallback music directions, 3 annotated example scores, and 27 bundled fonts.
20 documented motion-language styles distilled from 142 real launch videos, 10 starter templates, 23 sound effects, 12 fallback music directions, 3 annotated example scores, and 27 bundled fonts.What are similar projects?openfilm, guizang-product-video-skill, onetake, saas-motion-kit, eian-riso-broll, herk-edit-kit, mg-styles-15, huashu-art-motion
openfilm, guizang-product-video-skill, onetake, saas-motion-kit, eian-riso-broll, herk-edit-kit, mg-styles-15, huashu-art-motionWhich topics is it in?gsap, music-generation, motion-graphics, creative-tools, audio-generation, threejs, canvas, content-creation, svg, ffmpeg, video-generation, video
gsap, music-generation, motion-graphics, creative-tools, audio-generation, threejs, canvas, content-creation, svg, ffmpeg, video-generation, video