Extract Design System

Turn a public URL into a local git repo full of that site's real design assets and tokens. Only capture what is actually in the page and its linked resources — never invent colors, fonts, or brand facts from visual inspection.

Inputs to collect

1. URL — the page to extract from. Require a full https://....
2. Parent directory — where to create the repo. Default: ~/design-systems/. Ask; do not assume.
3. Repo name — derive from the host (e.g. example.com → example-com). Offer the default, let the user override.

Confirm all three before running anything.

Workflow

Step 1 — Try WebFetch for a first look

Use WebFetch on the URL and ask it to return: page title, meta description, and any mentions of brand or site name. This is cheap and gives you enough to confirm the target is what the user expects. If WebFetch is blocked, skip to Step 2 without retrying.

Step 2 — Run the extractor

mkdir -p "<parent>/<repo-name>"
uv run "${CLAUDE_PLUGIN_ROOT}/extract-design-system/scripts/extract.py" \
  "<url>" --output "<parent>/<repo-name>"

${CLAUDE_PLUGIN_ROOT} is set by Claude Code to the plugin's install directory. If it is unset (skill loaded outside a plugin), fall back to the path shown in the skill's own directory — e.g. resolve the script relative to this SKILL.md.

The script uses requests + beautifulsoup4 + tinycss2 to:

• Fetch the HTML and every linked stylesheet (following @import one level).
• Parse @font-face rules and download each font file into fonts/.
• Extract :root / html CSS custom properties into tokens/variables.css.
• Download favicons, Apple touch icons, and web-manifest icons into images/favicons/.
• Save inline <svg> blocks (likely logos) into logos/.
• Download the Open Graph image into images/.
• Snapshot the raw HTML into source/page.html.
• Write a machine-readable manifest.json and a fact-only README.md.

On success it prints a JSON summary (fonts, stylesheets, favicons, inline SVGs, custom-property counts).

Step 3 — Fallback chain if the page is JS-rendered

If the summary shows near-zero CSS/fonts/SVGs, the site is likely a JS SPA. Render it with a real browser and re-run with --html:

3a. Playwright CLI (preferred fallback):

playwright install chromium  # only if not yet installed
cat > /tmp/render.js <<'EOF'
const { chromium } = require('playwright');
(async () => {
  const b = await chromium.launch();
  const p = await b.newPage({ userAgent:
    'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 Chrome/126.0 Safari/537.36' });
  await p.goto(process.argv[2], { waitUntil: 'networkidle', timeout: 60000 });
  process.stdout.write(await p.content());
  await b.close();
})();
EOF
node /tmp/render.js "<url>" > /tmp/rendered.html

3b. MCP Playwright tools (if Node/Playwright CLI fails): use mcp__plugin_playwright_playwright__browser_navigate, then mcp__plugin_playwright_playwright__browser_evaluate with () => document.documentElement.outerHTML to capture HTML. Write it to /tmp/rendered.html.

3c. claude-in-chrome (last resort): load the tool with ToolSearch select:mcp__claude-in-chrome__navigate,mcp__claude-in-chrome__get_page_text, navigate, then use javascript_tool to read document.documentElement.outerHTML.

Then re-run the extractor with the rendered HTML:

uv run "${CLAUDE_PLUGIN_ROOT}/extract-design-system/scripts/extract.py" \
  "<url>" --output "<parent>/<repo-name>" --html /tmp/rendered.html

Step 4 — Initialize git and commit

cd "<parent>/<repo-name>"
git init -q
git add -A
git commit -q -m "Initial extraction from <url>"

Report the repo path, a one-line summary of what was captured, and any obvious gaps (e.g. zero fonts — probably means a JS-rendered site; rerun with Step 3).

Ground rules

• Only record facts present in the page or its linked resources. Never guess colors from visual inspection.
• If a field is missing (no og:image, no theme color, no @font-face), leave it out of the README — do not invent.
• Respect robots.txt and Terms of Service. If the user asks to extract from a site that disallows scraping, surface this and wait for confirmation.
• The script is idempotent: re-running into the same directory overwrites manifest.json and README.md and re-downloads assets.

Output layout

<repo-name>/
├── README.md           # fact-only summary
├── manifest.json       # structured extraction record
├── fonts/              # downloaded font files
├── logos/              # inline SVGs (candidate logos)
├── images/             # og:image and other page images
│   └── favicons/       # favicons + apple-touch + web-manifest icons
├── css/                # downloaded stylesheets
├── tokens/
│   ├── variables.css   # extracted :root custom properties
│   └── fonts.css       # rewritten @font-face rules pointing to local files
└── source/
    ├── page.html       # raw/rendered HTML snapshot
    └── manifest.webmanifest  # if the site exposes one