// the find
alchaincyf/huashu-design
Huashu Design · HTML-native design skill for Claude Code · Claude Code 里 HTML 原生的设计 skill · 高保真原型 / 幻灯片 / 动画 + 20 设计哲学 + 5 维评审 + MP4 导出 · Agent-agnostic
A Claude Code skill (installable into other agents that read SKILL.md) that turns a plain-language request into HTML-native design work: single-file app prototypes, slide decks, motion pieces exported to MP4 and GIF, infographics, and a five-axis critique. It is for developers and solo builders who would rather stay in the terminal than open Figma, and who are comfortable reading the output as code.
The brand-asset protocol makes the agent pull hex values out of real brand assets with grep and frequency-sort them, instead of guessing from memory. That addresses a common failure in AI-generated design, where the brand colours are plausible but wrong. html2pptx.js reads computedStyle from the DOM and emits real PowerPoint text boxes, so the exported deck stays editable, unlike tools that flatten slides to images. The cloud features (Doubao TTS, AI video review) are isolated under scripts/cloud and gated behind an explicit --yes, and SECURITY.md lists the outbound domains, which makes the data-flow claim easy to check. Rendering and export are scripted (render-video.js, convert-formats.sh for 60fps and palette-optimised GIF), so the output can be verified by running it rather than by trusting the model's description of it.
Most of the quality sits in prompt-level rubrics: the 60-style library, the twenty design philosophies, and the five-axis review are judgment calls written as text, and the repo has no evaluation harness showing how stable the results are across runs. The 5x variance claim comes from six agents per arm, which is too small to rely on. The README's own self-assessment of 80/100 is more useful than the marketing around it, but the gap it describes (no layer-level PPTX-to-Figma editing, no Framer Motion-grade animation, 60 to 65 quality from a blank brand) matters to anyone deciding whether to adopt it. Installation is fragile: the README admits that skills CLI 1.5.15 and earlier copied only SKILL.md, and the 99 references to references/, assets/ and scripts/ break without those subdirectories, so the fallback is a manual git clone. The MIT grant covers the code, but assets/ ships six BGM tracks and a large sound-effects library, and nothing in the file listing says where those came from, so anyone shipping generated videos commercially should check those licences first.