alchaincyf/huashu-design
Media & DesignAn agent skill that turns a one-line design brief into a deliverable HTML webpage, clickable app prototype, or editable PPTX, mainly for people who work without a graphics tool.
If you already work inside a coding agent and would rather not leave it to produce a web layout, pitch deck or app mockup, this skill is worth a trial on one real deliverable. It is less compelling if you need real layer editing in Figma or Keynote, or 3D and physics-grade motion.
The problem
You want a finished page, a clickable prototype or a slide deck for a launch or client, but your work lives in a coding agent or a terminal rather than a design tool. Assembling it means writing layout code by hand or bouncing between apps. With this skill the step that changes is the start: you give one sentence of intent, the agent picks candidate directions, renders HTML, and exports to pages, PDF or PPTX. The result is an asset you can open and show, not a text description of a layout.
What's different
Its core idea is to avoid the agent's first safe answer. The README describes a 3.0 flow that generates six thumbnails, narrows to three real directions, and draws on an 84-entry style library with sample sheets — and, when you supply a logo, palette or UI screenshots, reads your brand assets. There are also a 'design health check' and an appearance ledger in ~/.huashu-design/ that records your recent choices so later deliverables use a different typeface or layout. That addresses a real failure of one-shot generation: every output looking the same.
The README's own cases show the mechanism: for a Hokusai exhibition site, the health check flagged that two directions shared the same hero image and swapped one; for a birding app, it built a reference board before drawing. The author describes it as an 80-point skill, not a 100-point product.
The trade-off is format: the output is HTML-native, so you can screenshot, screen-record, or export images, video, PDF and PPTX, but you cannot drag it into Keynote or Figma and edit text positions as layers. Complex Framer Motion-level animation, 3D, physics and particle systems are outside its scope. The README also says a brand with nothing to start from drops to roughly 60–65 points — handing it real brand assets matters more than prompt wording.
Who should try it
Good fit: a founder, marketer or engineer who writes in Claude Code, Codex, Cursor or a similar agent and needs a decent-looking site, prototype or deck in hours, especially if you can supply a logo, palette or screenshots. The 3–30 minute range in the README and the Hokusai, game birding and Qianli Jiangshan examples are the kind of single-sentence briefs that fit.
Poor fit: a designer who will need to refine the result in a vector editor, or anyone needing cinematic motion or gameplay-quality interaction. If your brand is entirely blank, expect to be the 'last resort' case the README warns about, and consider building a starting visual system first.
Requirements and limitations
It is a skill, so you need a compatible agent host; the author lists Claude Code, Codex, Cursor and 'any agent', with skills.sh compatibility. The core design-to-render and MP4/PDF/PPTX export path is documented as local, with no network and no key (project README). That matters because the main deliverable loop does not depend on a cloud account or an API budget.
Optional cloud pieces — Doubao TTS voice-over and AI 'watch the cut' review — live under scripts/cloud/, use your key, and require explicit --yes confirmation on first use. Local taste and appearance records stay in ~/.huashu-design/; reference boards visit a site only when you ask the agent to capture a reference. The security document enumerates outbound domains and deletion boundaries, which is worth checking if client material is involved.
The practical limits are those above: no layer-editable PPTX round trip to Figma or Keynote, no heavy animation, and weaker results for brands with no assets. The output is a starting deliverable that likely needs human polish, not a pixel-locked final file.
HotRepo's take
On 2026-10-26 the repository appeared once on the GitHub Trending daily board at 26,397 stars, gaining 177 that day, according to HotRepo's single recorded sample. That confirms a dated snapshot of interest, not a trend; the sample covers one day only, so persistence cannot be judged. The repository is young (created 2026-04-19) and has moved to MIT, which removes the earlier commercial-use restriction. Before investing, check the security document's outbound-domain list against your client's data policy and confirm your agent host can load skills from skills.sh.
Written by AI from the project's README, updated October 26, 2026. The project's own docs are the source of truth.