Huashu Design is a design skill for AI coding agents that turns a one-sentence request into a deliverable file: a web page, a clickable app prototype, an editable PPTX deck, an animation or a print-ready infographic. It is aimed at developers and solo builders who already work inside Claude Code, Codex or Cursor and need output that looks designed without opening Figma or hiring anyone. The repository is worth a look partly because its own README is assembled from pages the skill produced, so the demo and the documentation are the same thing. One practical note before you start: the project documents itself mainly in Chinese, with an English README alongside the default one.
What it does
You install it once into your agent, then describe what you want in plain language. The skill handles the design decisions and hands back a finished artifact rather than a description of one. The five output types the project names are web pages, app prototypes you can click through, PowerPoint decks that stay editable, animations, and infographics. The authors put the turnaround at roughly 3 to 30 minutes depending on what you ask for; that is their claim, not an independently measured figure.
Two things are meant to keep the results from looking like generic AI output. If you hand the skill your brand assets (a logo, a colour palette, screenshots of your existing UI), it reads those for direction. If you hand it nothing, it falls back on a library of 84 styles, each shipped with a sample sheet so the agent is matching against a picture rather than a word like "minimal".
How it works
Instead of committing to the first idea, the skill sketches six thumbnails, picks three of them as real directions, and builds those into full drafts. When you point it at a reference design, it screenshots that reference before it starts work.
Version 3.0 adds two pieces of bookkeeping that are the more interesting part of the design:
- an "appearance ledger" that records the looks you have already shipped, so your next project does not come back wearing the same skin as the last one
- a design check that reports only measurable facts about the output rather than opinions about it
Everything is HTML-native, which is why the repository's main language is HTML and why it can run in any agent that reads skill files. The project lists Claude Code, Cursor, Codex, OpenClaw and Hermes as supported, and is skills.sh compatible. For the animation side, the authors edited their own 32-second 3.0 launch film with a sibling project, huashu-motion.
Getting started
The repository is github.com/alchaincyf/huashu-design, and installation is one command:
npx skills add alchaincyf/huashu-design
What it needs is an AI coding agent that is already installed and configured, plus that agent's own model access. That is the honest cost picture: the skill itself is free under the MIT license (the project switched to MIT on 2026-05-14, retiring its earlier "free for personal use, licence required for commercial use" terms, and now allows commercial use without prior permission), but the intelligence that does the work is billed by whoever supplies your agent's API. There is no local-model path described in the materials. No operating system, memory or GPU requirements are listed, which is consistent with a skill that is mostly instructions and HTML.
The main catch is the installer: older versions copy only one file instead of the full skill, so take the latest release from the repository rather than whatever your package cache already has. The release page is also where the launch video assets live, including a 1080p MP4 with music.
When to use it / when not
Use it instead of just asking Claude for a design when you want a file you can hand to someone: a deck that opens in PowerPoint, a prototype a stakeholder can click, a video for a product launch. A bare prompt to a chat model usually stops at a description or a single rough page.
Don't use it if you want editable Figma or After Effects source files to keep working in, if you have no coding agent set up and no intention of paying for one, or if a Chinese-first repository will slow you down. It is also still a young project, published in April 2026, with 3.0 arriving recently enough that the installer problem above was shipping until the latest release.
Alternatives
The closest comparison the project itself draws is to Claude Design, and the README keeps a section explaining how the two relate; read that before choosing between them. If you only need motion work, huashu-motion from the same author covers editing. Against conventional tools, this trades the control of Figma or After Effects for not having to open them.
Take this one seriously if you already live in a coding agent and keep shipping things that need to look presentable: landing pages, pitch decks, launch clips. The 26,000-plus stars collected in about six months say that combination is common. If your design work needs to stay editable in a designer's hands afterwards, this gives you a draft to start from, not a replacement for that pipeline.