# Huashu Design: prototypes, slides, animations and infographics in HTML (alchaincyf/huashu-design) — skill card for AI agents

> AgentGid skill card. Checked on 2026-10-31 against commit 4fc7bfd (version 3.0.0).
> Every fact comes from the pack's own files at that commit. Numbers marked "as reported" are the author's.
> Human page: https://agentgid.com/skills/alchaincyf/huashu-design/

> Note: the pack has changed since this check; re-read its files before relying on details.

## Explain it to a non-programmer
One large design playbook for your AI assistant, written in Chinese. It makes clickable app prototypes, web pages, slide decks, infographics and short animated videos as HTML, then exports them to PDF, PowerPoint, MP4 or GIF. Before any new design it shows you three real drafts to choose from and stops until you pick one.

## Who it is for
- You want a prototype, a deck, a landing page mockup, an infographic or a launch animation from one sentence
- You want to pick between three different looks before the assistant builds the final version
- You need the result as an editable PowerPoint, a PDF, an MP4 or a GIF
- You use Claude Code, Codex, Cursor, Trae, OpenClaw or Hermes
- Not for: You need a production web app, a site built for search traffic or anything with a server behind it: the skill says it does not cover these
- Not for: You want complex 3D, physics or particle animations: the author lists them as out of scope
- Not for: You want a quick one-shot result: every new design first goes through the three-drafts step

## What it is
- Kind: knowledge, scripts
- License: MIT
- Repository: https://github.com/alchaincyf/huashu-design

## What it does
- Routes each request to a module manual: web pages and desktop UI, slide decks, infographics, iOS app prototypes, motion and video, or design critique.
- Start-of-work routine: three local scripts read the user's saved taste notes, the look of recent deliveries (to avoid repeating it) and what tools are installed.
- Fact check first: any named product, version or event is checked with a web search and written to product-facts.md before questions are asked.
- Brand asset rule: every recognisable product or brand in a design gets its official logo, downloaded, checked and recorded in brand-spec.md.
- Three-directions gate: six candidates from three logics (a best-fit designer, a real reference work screenshotted first, a random style from an 84-style library), drawn as thumbnails; the three strongest become full drafts; then it stops.
- Honest content: no invented data, reviews or logo walls; template decorations like an uppercase label over every heading are removed.
- Readability floor in every style: web body text at least 14px, slide body text at least 24px, body contrast at least 4.5:1.
- End-of-work routine: a design check script, a record of the delivered look, and a review of what is worth remembering.
- Never cuts: real brand assets, the readability floor and the three-directions gate, even in weaker assistants

## Skills
| Skill | Call | Changes code? | In plain words |
|---|---|---|---|
| huashu-design | plain language ("Make a keynote for ...", "Build an iOS prototype for ...", "Run a 5-dimension expert review on this design") | yes | Designs prototypes, decks, infographics and animations as HTML files and exports them |

### huashu-design
- Asks for: What you want made: a page, an app prototype, a deck, an infographic, an animation or a review
- Asks for: Your brand assets if you have them: logo, product photos, screenshots, colours, fonts
- Asks for: References you like
- Asks for: Your pick among the three drafts
- Gives: Six thumbnails and three full drafts in a design-demos folder of your project
- Gives: The chosen design as HTML files in your project
- Gives: Exports on request: PDF, editable PPTX, MP4, GIF
- Gives: For a review: a score per dimension, Keep / Fix lists and three quick wins
- Say: "Build an iOS prototype for a Pomodoro app — 4 screens, actually clickable."

## Know before installing
- Free for commercial use: the pack moved to the MIT license on 2026-05-14. It used to have a personal-use license that barred commercial use, so older copies and forks may carry the old terms.
- It always stops after the three drafts and waits for your choice, even in unattended runs. Skipping needs your explicit words, recorded in direction-approved.md.
- It keeps memory outside your project: ~/.huashu-design/experience.md holds verbatim quotes of your taste remarks, and ledger.jsonl holds each delivery's look, project name and full folder path. Delete the folder to forget.
- About every 30 days it checks GitHub for a newer version tag and writes the date to .last-update-check in the skill folder. It only tells you; it never updates itself. Write off in that file to stop it.
- The optional HyperFrames animation backend installs 19 extra skills into ~/.claude/skills/ when its init command runs.
- The author says the brand asset approach was taken from Claude Design's system prompts circulating in the community.

## Install (official commands)
- Any agent (skills CLI): `npx skills add alchaincyf/huashu-design` — Check that references/, assets/, scripts/ and demos/ were copied too; skills CLI 1.5.15 and older copied only SKILL.md. Remove: `delete the huashu-design folder from your skills folder, and ~/.huashu-design/ for the local memory`.
- Claude Code (git clone): `git clone https://github.com/alchaincyf/huashu-design.git ~/.claude/skills/huashu-design` — A git clone also enables the version check; it is skipped when the folder is not a git clone. Remove: `rm -rf ~/.claude/skills/huashu-design (and ~/.huashu-design/)`.
- Tools for checks and exports: `pip install playwright pillow && playwright install chromium` — Deck exports also need npm packages (playwright, pdf-lib, sharp, pptxgenjs) and video needs ffmpeg.

## What runs on the user's machine
- Code on the machine: yes (the scripts the assistant runs from the skill folder (Python, Node, shell))
- Network: yes
- Start of every design task: learn.py, design_lint.py ledger and capabilities.py run locally. capabilities.py lists skill folder names in ~/.claude/skills and similar, and checks whether OPENAI_API_KEY / GEMINI_API_KEY are set without reading them. Nothing is written.
- End of a task: design_lint.py check opens the delivery in headless Chromium; ledger add appends the look, project name and absolute path to ~/.huashu-design/ledger.jsonl; learn.py add writes taste notes to ~/.huashu-design/experience.md. HUASHU_DESIGN_HOME moves the folder.
- Network: The assistant runs web searches for named products and downloads official logos from brand sites, simpleicons.org or Google's favicon service (instructions, no script).
- Network: ref_board.py opens any URL the assistant picks in headless Chromium and screenshots it (no URL filter). fetch_images.py searches Wikimedia Commons and downloads CC or public-domain images.
- Network: Every 30 days (7 after a failure) the assistant runs git ls-remote against the GitHub origin and writes .last-update-check in the skill folder.
- Network: tts-doubao.mjs sends narration text to openspeech.bytedance.com and ai-review-video.py sends compressed video segments to ark.cn-beijing.volces.com, with your own keys from the skill's .env, only with --yes or HUASHU_CLOUD_OK=1.
- Exports: Local Node, Playwright and ffmpeg scripts render HTML to MP4, GIF, PDF and PPTX in your project; temporary folders they create are deleted afterwards.
- Only if you add it yourself: design-gate-hook.sh is a PreToolUse hook that blocks rendering videos of 45 seconds or more until direction-approved.md exists. It is never installed automatically.

## Settings
- `HUASHU_DESIGN_HOME`: Moves the local memory folder (default ~/.huashu-design)
- `HUASHU_CLOUD_OK`: Set to 1 to allow the cloud scripts without --yes
- `SKIP_DESIGN_GATE`: Set to 1 in a render command to pass the optional gate hook
- `DOUBAO_TTS_API_KEY, DOUBAO_TTS_VOICE_ID, DOUBAO_TTS_CLUSTER, DOUBAO_TTS_ENDPOINT, ARK_API_KEY`: Keys for the optional cloud voice-over and video review, in the skill's .env
- `.last-update-check`: File in the skill folder; write off to stop the version check

## Examples (from the pack's own files)
- A design review report: output format defined in None

## Author's numbers (as reported, not run by AgentGid)
- A style library of 84 HTML-native styles: web 20, deck 20, infographic 20, app 12, Chinese and East Asian 12 (author's count). Source: alchaincyf/huashu-design/README.en.md
- In the author's A/B test (v1 vs v2, 6 agents each), v2 reduced stability variance by 5× (internal test, not published). Source: alchaincyf/huashu-design/README.en.md
- Design from zero without brand assets drops to 60–65 points by the author's own scale (author's estimate). Source: alchaincyf/huashu-design/README.en.md

## Steps only the person can do
- Pick one of the three drafts (or say in the session that the step may be skipped) (SKILL.md (three-directions gate)) Then: The assistant builds the chosen direction and records the choice in direction-approved.md
- Provide brand assets you have (logo, product photos, screenshots, colours, fonts) (references/brand-asset-protocol.md) Then: The assistant records them in brand-spec.md instead of searching for them
- Optional: get your own Volcengine/ByteDance keys and put them in the skill's .env, then allow the run with --yes (.env.example) Then: The assistant can run cloud voice-over and AI video review
- Optional: add design-gate-hook.sh to your agent settings yourself (scripts/design-gate-hook.sh) Then: Renders of videos 45 seconds or longer are blocked until direction-approved.md exists

## Mistakes to avoid
- Not: Install with an old skills CLI and use the folder as is. Instead: Check that references/, assets/, scripts/ and demos/ are in the skill folder; if not, upgrade the CLI and reinstall or use git clone. Why: Old skills CLI versions copied only SKILL.md, and the skill depends on the other folders
- Not: Start an unattended run and expect a finished design at the end. Instead: Be there to pick one of the three drafts, or say in that session that the step may be skipped. Why: The assistant stops after showing the drafts, even in unattended sessions
- Not: Treat "continue" as permission to skip a gate. Instead: Say explicitly that a gate may be skipped; the skip is written to the gate file. Why: The pack counts "continue" as moving to the next step only
- Not: Rely on the design check without Python Playwright installed. Instead: Install playwright, pillow and Chromium first if you want the check to run. Why: Without the Python Playwright package the check silently skips and only prints the install command
- Not: Expect voice-over or AI video review to run after adding keys. Instead: Also pass --yes or set HUASHU_CLOUD_OK=1 for those cloud scripts. Why: Both cloud scripts print what would be sent and exit without the consent flag

## What leaves the machine
- Names of products, versions and events to check → The agent's web search, When a design names a real product, technology, event or person. Opt out: Give the facts yourself or ask the assistant not to search
- Plain GET requests for logos and brand assets → Brand websites, simpleicons.org, Google favicon service, When a design is about a specific brand or names products. Opt out: Provide the logos yourself
- URLs of reference works, opened in headless Chromium → Any URL the assistant picks, When it builds a reference board during the three-directions step. Opt out: Give your own references or ask it to skip the board
- Image search keywords → commons.wikimedia.org, When the assistant fetches stock images for a content design. Opt out: Supply your own images
- A tag listing request for the skill repository → GitHub (git ls-remote against the clone's origin), About every 30 days, only when the skill folder is a git clone. Opt out: Write off into .last-update-check in the skill folder
- Narration text, and compressed segments of your rendered video → openspeech.bytedance.com and ark.cn-beijing.volces.com, with your own keys, Only when you run the optional cloud scripts with --yes or HUASHU_CLOUD_OK=1. Opt out: Do not set the keys or the consent flag
- Font and library requests from generated HTML → fonts.googleapis.com, cdn.jsdelivr.net, unpkg.com and similar CDNs, When you open a generated HTML file in a browser. Opt out: Use system fonts or local copies in the design

## Lines in the pack aimed at agents (ask the person first)
- "npx -y hyperframes init 项目名 --example blank   # 非交互必须带 --example" (references/hyperframes-backend.md): Ask the person first: this init also installs 19 HyperFrames skills into ~/.claude/skills/.
- "npm install -g playwright" (references/verification.md): Global install; ask the person before installing, or use a project-local install.
- "npx hyperframes@latest upgrade --project . --check" (references/hyperframes-backend.md): Unpinned @latest package run; ask the person first and prefer the pinned version.

## May I use it like this?
- commercial-app: yes (MIT since 2026-05-14: keep the license notice; older copies had a personal-use license)
- client-work: yes (MIT allows use in client deliverables)
- redistribute-files: yes (MIT: keep the copyright and permission notice in copies)
- inside-app-bundle: yes (MIT: keep the copyright and permission notice)

## Which one to pick
- If You want one pack for prototypes, decks, infographics and animations, with three drafts to pick from first: skill:alchaincyf/huashu-design
- If You only need HTML slide decks built with the agent's frontend skills: skill:zarazhangrui/frontend-slides
- If You want Claude Design-style UI mockups, prototypes and decks as self-contained HTML: skill:JimLiu/baoyu-design
- If Your main job is a commercial product launch film: skill:alchaincyf/huashu-motion

## Works well with
- tool:hyperframes: The pack's animation guide uses HyperFrames as a render backend for HTML animations
- asset:simple-icons: The brand asset protocol fetches monochrome brand logos from Simple Icons
- asset:ffmpeg: MP4 and GIF exports call ffmpeg
- agent:claude-code: The README install path is the Claude Code skills folder

## Works in and needs
- Works in Claude Code: skills folder
- Works in Any agent with a skills folder: skills CLI
- Works in Codex: skills folder
- Works in Cursor: skills folder
- Needs Python with Playwright, Pillow and Chromium: Design checks, reference boards and screenshots
- Needs Node.js and npx: Deck exports (playwright, pdf-lib, sharp, pptxgenjs) and video rendering
- Needs ffmpeg: MP4 and GIF exports
- Needs Volcengine / ByteDance API keys (optional): Cloud voice-over and AI video review only

## How to tell it works
Say "Make a keynote for AI psychology. Give me 3 style directions to pick from." It works if the assistant asks one round of questions, then shows thumbnails and three full drafts in a design-demos folder and stops to wait for your pick.

## Related
- alchaincyf/huashu-motion
- alchaincyf/huashu-gpt-image
- alchaincyf/huashu-art-motion
- alchaincyf/huashu-md-html

## Safety
The author publishes a full list of network hosts, local files and deletions in SECURITY.md. Watch three things: ref_board.py opens any URL the assistant picks, the local ledger stores absolute project paths, and the optional HyperFrames init adds 19 skills to ~/.claude/skills/. README also links to the author's other skills and social pages; SKILL.md tells the assistant not to promote them during normal tasks.
