Design Distiller

Design Distiller

Give any website URL, extract the complete design system into a Design MD + composable design tokens.

Triggers

• /design-distiller
• "scrape this website's design" / "extract design system"
• "design scrape" / "make a Design MD"
• "analyze this site's design"

Core Concepts

Design MD

A pure Markdown design system document that AI agents can read directly to generate matching UI. Format follows awesome-design-md.

Reference Pool

references/{slug}/ — One folder per website, containing the full Design MD + screenshot assets. This is the raw material library.

Digest Pool

digests/ — Cross-site design tokens decomposed by dimension (typography, colors, spacing, components, etc.). Each entry is tagged with its source website. Used for mix-and-match composition when building new sites.

Output Structure

design-distiller/
├── references/                # Reference Pool
│   └── {slug}/
│       ├── DESIGN.md          # Full Design MD (9 modules)
│       ├── meta.json          # Metadata
│       └── screenshots/       # Key page screenshots
│           ├── homepage.png
│           ├── homepage-mobile.png
│           └── components.png
└── digests/                   # Digest Pool
    ├── typography.md          # All font systems
    ├── colors.md              # All color systems
    ├── spacing.md             # All spacing systems
    ├── components.md          # All component patterns
    ├── depth.md               # All shadow & elevation systems
    ├── motion.md              # All motion specs
    ├── layouts.md             # All layout systems
    └── philosophy.md          # All design philosophies

Pre-loaded References

The Reference Pool ships with 55 pre-loaded brands imported from awesome-design-md and the nothing-design skill. These text-only imports differ from freshly scraped references:

• No screenshots/ — visual data was not captured during batch import. Screenshots are generated only when running the full 4-phase pipeline on a new URL.
• No meta.json — metadata is generated only by the pipeline's Generate phase. Pre-loaded references have their metadata embedded in the Digest Pool cross-references instead.
• Nothing uses an extended format — references/nothing/ contains multi-file documentation (DESIGN.md + tokens.md + components.md + platform-mapping.md) from a standalone design system skill. It does not follow the standard 9-module format because its craft rules and composition philosophy cannot be reduced to that template.

All other references follow the standard 9-module Design MD format.

Workflow — 4 Phase Pipeline

Phase 1: Scrape

-> prompts/scraper.md

Goal: Extract design information from the target website using all available tools.

Tool Priority

Capability Detection: Before starting a scrape, check which browser tools are available in the current environment. Try the highest-priority tool first; if unavailable or erroring, fall back to the next tier. Do not assume any specific tool is present.

Extraction Checklist

For each website, collect:

A. Visual Screenshots (via browser tools)

• Homepage full-page screenshot (desktop 1440px + mobile 390px)
• Dark/light mode toggle (if available)
• 2-3 key subpages (about, product, pricing)
• Interactive states (button hover, input focus)

B. CSS Source Extraction (via browser JS execution or source analysis)

• CSS custom properties (--* variables)
• @font-face declarations
• Computed styles for key selectors (h1-h6, body, button, input, card)
• Media query breakpoints
• Animation/transition definitions

C. Asset Files

• External CSS file URL list
• Font file URLs
• Design token files (e.g., design-tokens.json, theme.ts)

D. Page Structure

• DOM hierarchy (main landmark elements)
• Component pattern identification (nav, card, list, form, etc.)

Browser JS Extraction Scripts

Execute via browser tools to extract design data:

// 1. Extract all CSS custom properties
(() => {
  const styles = getComputedStyle(document.documentElement);
  const vars = {};
  for (const sheet of document.styleSheets) {
    try {
      for (const rule of sheet.cssRules) {
        if (rule.style) {
          for (let i = 0; i < rule.style.length; i++) {
            const prop = rule.style[i];
            if (prop.startsWith('--')) {
              vars[prop] = rule.style.getPropertyValue(prop).trim();
            }
          }
        }
      }
    } catch(e) {} // cross-origin sheets
  }
  return JSON.stringify(vars, null, 2);
})()

// 2. Extract computed styles for key elements
(() => {
  const selectors = ['h1','h2','h3','h4','h5','h6','p','a','button',
    'input','nav','footer','[class*=card]','[class*=hero]'];
  const result = {};
  for (const sel of selectors) {
    const el = document.querySelector(sel);
    if (!el) continue;
    const s = getComputedStyle(el);
    result[sel] = {
      fontFamily: s.fontFamily,
      fontSize: s.fontSize,
      fontWeight: s.fontWeight,
      lineHeight: s.lineHeight,
      letterSpacing: s.letterSpacing,
      color: s.color,
      backgroundColor: s.backgroundColor,
      padding: s.padding,
      margin: s.margin,
      borderRadius: s.borderRadius,
      boxShadow: s.boxShadow,
      transition: s.transition
    };
  }
  return JSON.stringify(result, null, 2);
})()

// 3. Extract fonts
(() => {
  const fonts = new Set();
  for (const sheet of document.styleSheets) {
    try {
      for (const rule of sheet.cssRules) {
        if (rule instanceof CSSFontFaceRule) {
          fonts.add({
            family: rule.style.getPropertyValue('font-family'),
            src: rule.style.getPropertyValue('src'),
            weight: rule.style.getPropertyValue('font-weight'),
            style: rule.style.getPropertyValue('font-style')
          });
        }
      }
    } catch(e) {}
  }
  return JSON.stringify([...fonts], null, 2);
})()

Phase 2: Analyze

-> prompts/analyzer.md

Goal: Structure the raw scraped data into the 9 Design MD modules.

Design MD — 9 Modules

Analysis Principles

1. Observation over guessing — Only document what screenshots and code show. Never fabricate.
2. Precise values — Colors must be hex, font sizes must be px, spacing must be concrete numbers.
3. Semantic naming — Don't write #171717, write Near Black (#171717) — primary text color.
4. Source attribution — Tag each value as extracted from CSS variables vs. visually estimated.
5. Uncertainty marking — If visually estimated from screenshots, mark as (visual estimate).

Phase 3: Generate

-> prompts/generator.md

Goal: Generate the final Design MD file + meta.json.

meta.json Schema

{
  "name": "Site Name",
  "slug": "site-name",
  "url": "https://example.com",
  "scraped_at": "YYYY-MM-DD",
  "version": "v1",
  "theme_modes": ["light", "dark"],
  "primary_font": "Font Name",
  "primary_color": "#hexval",
  "style_keywords": ["minimalist", "warm", "industrial"],
  "confidence": {
    "colors": "high",
    "typography": "high",
    "spacing": "medium",
    "components": "medium",
    "motion": "low"
  },
  "pages_analyzed": [
    "https://example.com/",
    "https://example.com/about"
  ]
}

Design MD Template

See the 9-module template in prompts/generator.md. Each module follows the structure:

• Section heading with number
• Key data in Markdown tables
• Contextual notes and usage guidelines

Phase 4: Digest

-> prompts/digester.md

Goal: Decompose the newly generated Design MD and merge into the Digest Pool.

Digest File Format

Each digest file is organized by source website for easy mix-and-match:

# Typography Digest

## Cross References
- Vercel and Linear both use Geist Sans
- Claude and Notion both use warm-toned backgrounds

## Vercel
- **Display Font**: Geist Sans, -2.4px letter-spacing
- **Body Font**: Geist Sans, 400 weight, 16px/1.6
- **Mono**: Geist Mono
- **Style**: Extreme compression, modern geometric

## Stripe
- **Display Font**: sohne-var, variable weight
- **Body Font**: sohne-var, 16px/1.5
- **Style**: Sophisticated, OpenType features

Digest Rules

1. Auto-update after each new reference — No manual trigger needed.
2. Tag confidence — (CSS var) vs (visual estimate).
3. Tag uniqueness — Mark distinctive choices with [unique].
4. Cross-reference — If two sites share the same font/color, note the connection.

Command Interface

Basic Usage

/design-distiller https://vercel.com

-> Full pipeline: Scrape -> Analyze -> Generate -> Digest

Extended Commands

Compose Example

/design-distiller compose "Developer tool landing page with Vercel typography + Stripe colors + Linear component style"

-> Extracts matching dimensions from Digest Pool, composes a new Design MD.

Scraping Strategies

Browser Use Cloud (Priority)

1. Create session -> mcp__browser-use__run_session
2. Navigate to target URL
3. Screenshot homepage (desktop + mobile)
4. Execute JS to extract CSS variables and computed styles
5. Navigate to 2-3 subpages for screenshots
6. Try toggling dark mode for screenshots
7. Close session

Playwright MCP (Fallback)

1. browser_navigate -> target URL
2. browser_take_screenshot -> homepage screenshot
3. browser_evaluate -> execute JS extraction scripts
4. browser_resize -> mobile viewport, screenshot again
5. browser_navigate -> subpages

Source Code Analysis (Last Resort)

1. WebFetch -> get HTML
2. Extract <link rel="stylesheet"> URLs from HTML
3. WebFetch -> fetch each CSS file
4. Regex extract CSS variables, @font-face, key selectors
5. Analyze <meta> tags for theme colors

Parallel Scraping Strategy

When multiple scraping tools are available, launch parallel Agent tasks:

Quality Standards

Design MD Completeness Check

Confidence Levels

Agent Behavior Rules

Before Starting

1. Capability detection — ToolSearch to check browser-use / playwright / chrome availability
2. Check Reference Pool — Whether the target site already has a reference
3. Inform user — Which scraping strategy will be used

During Scraping

• Save screenshots to references/{slug}/screenshots/
• Raw extraction data does not need persistence (use and discard)
• If anti-bot/CAPTCHA encountered -> inform user, degrade to source analysis
• If the site has public design system docs -> prioritize scraping those

After Generation

• User must review Design MD before it's committed to the pool
• Digest Pool updates should show diff to user
• meta.json confidence field must be honestly tagged

Prohibited Actions

• Never fabricate design values that cannot be observed
• Never screenshot login-gated or paid content
• Never store authentication credentials
• Never overwrite existing references without versioning (see Version Management below)

Version Management

When re-scraping an existing reference:

1. Check existing version — Read meta.json version field (e.g., "v1")
2. Increment — Bump to next version ("v2", "v3", etc.)
3. Archive previous — Rename existing folder to {slug}-v{N} before writing new version
4. Update Digest Pool — Replace the slug's entries in all digest files with the new data
5. Log in meta.json — Record "previous_version": "v1" and "updated_at": "YYYY-MM-DD"

For pre-loaded references (no meta.json), the first re-scrape creates "version": "v2" with "previous_version": "imported".