intersight:design-analyze

Analyze a website and extract its design system as W3C DTCG tokens.

Arguments

Parse the user's arguments from $ARGUMENTS:

• URL (required, positional) — The target URL to analyze. Must start with http:// or https://.
• --depth (optional, default: standard) — Analysis depth: tokens, standard, or full
• --format (optional, default: json) — Output format: json, markdown, or tokens-only
• --fresh (optional) — Skip cache lookup, run full extraction
• --pages (optional) — Comma-separated relative paths for multi-page analysis (e.g., /,/dashboard,/settings)

If no URL is provided, ask the user: "What URL would you like to analyze?"

Validate the URL starts with http:// or https://. If not, prepend https:// and warn the user.

State Tracking

Track these variables throughout execution:

url = <parsed URL>
depth = tokens | standard | full
format = json | markdown | tokens-only
fresh = true | false
pages = [<parsed page paths>] or ["/"]
dembrandt_available = true (until proven otherwise)
intercache_available = true (until proven otherwise)
phases_completed = []
warnings = []
extraction_results = {}

Phase 0: Preflight

0a: Dependency checks

Check Dembrandt:

command -v npx >/dev/null 2>&1 && npx dembrandt --version 2>/dev/null

• If fails: set dembrandt_available = false, add warning: "Dembrandt not found. Continuing with DOM-only extraction. For better results: npm install -g dembrandt"
• Do NOT abort — DOM extraction works without Dembrandt.

Check Playwright MCP: Attempt: browser_navigate to about:blank

• If fails (tool not found or connection error): ABORT with error:

  "intersight requires Playwright MCP server to analyze websites. Install: claude mcp add playwright npx @playwright/mcp@latest --snapshot-mode none and add it to your Claude Code MCP settings (Settings > MCP Servers)."

• This is a hard requirement — cannot proceed without browser.

Token optimization check: After confirming Playwright MCP is available, check whether the server was started with --snapshot-mode none. If every tool response includes a large accessibility snapshot (thousands of tokens of ARIA tree), warn the user:

"Playwright MCP is running with default snapshot mode. For significantly lower token usage (~70-80% reduction per action), restart with --snapshot-mode none. The browser_snapshot tool still works independently — this only suppresses the auto-appended snapshots on every other tool response."

This optimization is strongly recommended because Intersight's pipeline makes 16-40 Playwright tool calls per analysis (depending on depth), and only Phase 4's browser_snapshot needs the accessibility tree. All other calls (browser_navigate, browser_resize, browser_evaluate, browser_take_screenshot, browser_hover) discard the snapshot.

0b: robots.txt compliance (mandatory)

Extract the target path from the URL:

// e.g., https://example.com/dashboard → /dashboard
const targetPath = new URL(url).pathname;

Navigate to <origin>/robots.txt:

browser_navigate: <origin>/robots.txt

Read the parseRobotsTxt.js script from ${CLAUDE_PLUGIN_ROOT}/scripts/extraction/parseRobotsTxt.js, replace __TARGET_PATH__ with the actual target path, then evaluate:

browser_evaluate: <modified parseRobotsTxt.js content>

Parse the result:

• If allowed: false: ABORT with message:

  "robots.txt disallows access to [path] for web crawlers (matched rule: [matchedRule]). intersight respects robots.txt. Suggestions: analyze a different page, or run against a local copy."

• If allowed: true: continue.

Also check meta robots tag after page navigation (in Phase 1).

0c: Cache lookup (if intercache available and not --fresh)

Only if fresh is false:

Attempt: cache_lookup(key="intersight:" + url + ":" + depth)

• If tool call fails (tool not found): set intercache_available = false, continue
• If cache hit: parse and return cached result immediately. Done.
• If cache miss: continue with extraction.

Add preflight to phases_completed.

Phase 1: Setup

Navigate to the target URL:

browser_navigate: <url>

Wait for page stability (SPA quiescence): Run via browser_evaluate:

(() => {
  return new Promise(resolve => {
    let lastCount = document.querySelectorAll('*').length;
    let stableChecks = 0;
    const interval = setInterval(() => {
      const current = document.querySelectorAll('*').length;
      if (current === lastCount) stableChecks++;
      else { stableChecks = 0; lastCount = current; }
      if (stableChecks >= 3) { clearInterval(interval); resolve(JSON.stringify({ stable: true, elementCount: current })); }
    }, 500);
    setTimeout(() => { clearInterval(interval); resolve(JSON.stringify({ stable: false, elementCount: lastCount })); }, 5000);
  });
})()

Challenge page detection: Run via browser_evaluate:

(() => {
  const title = document.title.toLowerCase();
  const bodyText = (document.body?.innerText || '').substring(0, 500).toLowerCase();
  const indicators = ['just a moment', 'checking your browser', 'please wait', 'access denied',
                       'verify you are human', 'captcha', 'challenge-platform'];
  const detected = indicators.filter(i => title.includes(i) || bodyText.includes(i));
  return JSON.stringify({ challengeDetected: detected.length > 0, indicators: detected });
})()

• If challengeDetected: true: ABORT with:

  "Target site served a bot-protection challenge (detected: [indicators]). intersight cannot bypass bot protection. Try analyzing a local dev server or a site without protection."

Meta robots check: Run via browser_evaluate:

(() => {
  const meta = document.querySelector('meta[name="robots"]');
  if (!meta) return JSON.stringify({ blocked: false });
  const content = (meta.getAttribute('content') || '').toLowerCase();
  const blocked = content.includes('noindex') || content.includes('nofollow');
  return JSON.stringify({ blocked, content: meta.getAttribute('content') });
})()

• If blocked: true: add warning: "Page has <meta name='robots' content='[content]'> — extraction may be restricted."
• Do NOT abort (meta robots is advisory for this use case, unlike robots.txt).

Set viewport:

browser_resize: { width: 1440, height: 900 }

Content hash: Read and evaluate ${CLAUDE_PLUGIN_ROOT}/scripts/extraction/contentHash.js via browser_evaluate. Store the result as content_hash.

Add setup to phases_completed.

Phase 2: Dembrandt Baseline

Skip if dembrandt_available is false.

Run via Bash:

npx dembrandt "<url>" --dtcg --json-only 2>/dev/null

• If succeeds: parse JSON output, store as extraction_results.dembrandt
• If fails (non-zero exit, invalid JSON, timeout after 30s): set dembrandt_available = false, add warning: "Dembrandt extraction failed. Continuing with DOM-only extraction."
• Do NOT abort on Dembrandt failure.

Add dembrandt to phases_completed (even if skipped — note in meta).

Phase 3: DOM/CSS Extraction

For each extraction script, read the file from ${CLAUDE_PLUGIN_ROOT}/scripts/extraction/<name>.js using the Read tool, then pass the full content to browser_evaluate.

Run these 7 scripts in sequence:

1. extractCSSCustomProperties.js → store as extraction_results.customProperties
2. extractColorTokens.js → store as extraction_results.colors
3. extractTypography.js → store as extraction_results.typography
4. extractSpacing.js → store as extraction_results.spacing
5. extractShadowsAndBorders.js → store as extraction_results.shadowsAndBorders
6. extractBreakpoints.js → store as extraction_results.breakpoints
7. extractComponentInventory.js → store as extraction_results.components

For each script:

• Parse the JSON result
• If error field present in result: add warning, continue with partial data
• If count: 0 and no error: note low count but continue

Low-count warning: If total unique colors < 3 AND custom properties count is 0: Add warning: "Very low token count detected. The site may use client-side rendering that wasn't fully hydrated, or the page may have minimal styling."

Add dom_extraction to phases_completed.

Phase 4: Structural Analysis (standard + full only)

Skip if depth is tokens.

Accessibility tree:

browser_snapshot

Store the snapshot output. Use it to identify:

• Navigation patterns (nav elements, landmark roles)
• Primary interactive elements (buttons, links, inputs)
• Page structure (headings hierarchy, sections)

Note: extractComponentInventory.js already ran in Phase 3. Use its output for the component list.

Add structural to phases_completed.

Phase 5: Visual Analysis (standard + full only)

Skip if depth is tokens.

Screenshots:

For standard depth:

browser_take_screenshot

Capture 1 desktop screenshot (viewport already set to 1440x900).

For full depth, capture 3 screenshots:

1. Desktop (1440x900) — already set
2. Tablet: browser_resize: { width: 768, height: 1024 } then browser_take_screenshot
3. Mobile: browser_resize: { width: 375, height: 812 } then browser_take_screenshot
4. Restore viewport: browser_resize: { width: 1440, height: 900 }

Claude vision analysis:

For each screenshot, analyze with this prompt:

"Analyze this website screenshot for design system characteristics:

1. Layout pattern — What layout pattern is used? (sidebar-main, holy grail, single column, grid, etc.)
2. Visual hierarchy — How does the page guide the eye? What's most prominent?
3. Color harmony — Are colors consistent? What palette strategy? (monochromatic, complementary, analogous)
4. Information density — Is the page sparse, moderate, or dense with information?
5. Responsive indicators — Any visible responsive design patterns?"

Store vision analysis as extraction_results.visualAnalysis.

Add visual to phases_completed.

Phase 6: Interaction States (full only)

Skip if depth is not full.

From the component inventory (Phase 3), identify up to 5 primary interactive elements:

• Buttons (highest priority)
• Links
• Input fields

For each interactive element, extract hover state changes:

1. Capture default styles via browser_evaluate:

(() => {
  const el = document.querySelector('<selector>');
  if (!el) return JSON.stringify({ found: false });
  const s = getComputedStyle(el);
  return JSON.stringify({
    found: true,
    backgroundColor: s.backgroundColor,
    color: s.color,
    borderColor: s.borderColor,
    boxShadow: s.boxShadow,
    transform: s.transform,
    opacity: s.opacity
  });
})()

2. Hover: browser_hover: <selector>

3. Capture hover styles with the same script above.

4. Diff: compare default vs hover styles and record changes.

Also extract focus ring patterns:

(() => {
  const focusable = document.querySelectorAll('button, a, input, select, textarea, [tabindex]');
  const rings = new Set();
  for (let i = 0; i < Math.min(focusable.length, 10); i++) {
    focusable[i].focus();
    const s = getComputedStyle(focusable[i]);
    const ring = s.outlineWidth + ' ' + s.outlineStyle + ' ' + s.outlineColor;
    if (s.outlineStyle !== 'none') rings.add(ring);
  }
  document.activeElement?.blur();
  return JSON.stringify({ focusRings: Array.from(rings) });
})()

Store as extraction_results.interactionStates.

Add interactions to phases_completed.

Phase 7: Synthesis

Merge all extraction results into W3C DTCG format.

Merge Algorithm

1. Custom properties are the primary token source. Each CSS custom property from extractCSSCustomProperties.js that represents a design token (color, spacing, typography) becomes a named token in the output. Use the resolvedValue field for the actual token value.

2. Computed colors fill gaps. Colors from extractColorTokens.js that don't match any custom property's resolvedValue are added as unnamed tokens (e.g., color.extracted-1, color.extracted-2).

3. Dembrandt provides token names. For any Dembrandt token whose value matches a DOM custom property's resolved value, use Dembrandt's semantic name as the token name. On value conflicts, DOM value wins (reflects runtime state).

4. Multi-page conflicts (when --pages is used): When the same CSS custom property name has different resolved values across pages, report BOTH values with page context in warnings. Use the value with higher total frequency.

5. Track completeness. Set intersight:meta.phases_completed and intersight:meta.dembrandt_available.

Build Output Structure

Read the schema template from ${CLAUDE_PLUGIN_ROOT}/scripts/extraction/schema.json.

Populate each section:

intersight:meta:

• source_url: the analyzed URL
• analyzed_at: current ISO timestamp
• analysis_depth: the depth mode used
• pages_analyzed: list of page paths analyzed
• tool_version: "0.1.0"
• content_hash: from Phase 1
• dembrandt_available: boolean
• phases_completed: list of phase names
• warnings: accumulated warnings

color group: Merge custom properties (color-type) + Dembrandt color tokens + extracted computed colors. Format each as:

{
  "color-name": {
    "$type": "color",
    "$value": "#hex or rgb(...)",
    "$description": "Frequency: N, source: custom-property/dembrandt/computed"
  }
}

dimension group: From spacing extraction. Format each as:

{
  "space-Npx": {
    "$type": "dimension",
    "$value": "Npx",
    "$description": "Frequency: N"
  }
}

typography group: From typography extraction. Each unique type scale entry:

{
  "type-scale-N": {
    "$type": "typography",
    "$value": {
      "fontFamily": "...",
      "fontSize": "...",
      "fontWeight": N,
      "lineHeight": "...",
      "letterSpacing": "..."
    },
    "$description": "Frequency: N, used in: tag1, tag2"
  }
}

shadow group: From shadows extraction.

border group: From borders extraction.

intersight:components: Array from component inventory.

intersight:ux_flow: (standard + full only) From structural analysis — navigation patterns, primary actions, information density assessment.

intersight:visual_analysis: (standard + full only) From Claude vision analysis — layout pattern, visual hierarchy, responsive behavior.

Format Output

Based on --format:

json (default): Output the full DTCG JSON with all extensions.

markdown: Generate a human-readable report:

# Design System Analysis: <domain>

**URL:** <url>
**Analyzed:** <timestamp>
**Depth:** <depth>

## Color Palette
[Table of colors with hex, frequency, source]

## Typography Scale
[Table of type entries with font, size, weight, frequency]

## Spacing Scale
[Table of spacing values with frequency]

## Shadows & Borders
[Tables]

## Breakpoints
[List of responsive breakpoints]

## Component Inventory (if standard+)
[Table of components with frequency, variants]

## Layout Analysis (if standard+)
[Vision analysis summary]

## Interaction States (if full)
[Hover/focus state changes]

---
*Generated by intersight v0.1.0 — [phases_completed]*

tokens-only: Output bare DTCG tokens (color, dimension, typography, shadow, border groups only — no $extensions).

Write output file

Always write the JSON output to a local file:

./intersight-analysis-<domain>-<timestamp>.json

where <domain> is the URL's hostname with dots replaced by dashes, and <timestamp> is YYYYMMDD-HHmmss.

Tell the user: "Output written to: [filepath]"

Cache result (if intercache available)

If intercache_available is true:

cache_store(key="intersight:" + url + ":" + depth, value=<json output>, ttl=604800)

(7-day TTL = 604800 seconds)

Add synthesis to phases_completed.

Multi-Page Analysis

If --pages was specified:

1. Parse comma-separated paths into a list.
2. For each page path: a. Navigate to <origin><path> b. Run Phases 1-6 for this page (skip Phase 0 — already done) c. Store per-page results d. If not localhost: wait 10 seconds between pages (rate limiting for third-party sites)
3. Merge results across pages:
   • Tokens: union, with higher-frequency value winning on name conflicts
   • Components: merge inventories, accumulate frequencies
   • Warnings: include per-page context for conflicts
4. Proceed to Phase 7 synthesis with merged results.

Error Handling

If any phase fails:

1. Log which phase failed and the error message
2. Add to warnings
3. Continue to the next phase if possible (Phases 2-6 can fail independently)
4. Phase 7 (synthesis) works with whatever data was collected

Only Phase 0 (Playwright MCP missing, robots.txt blocked) and Phase 1 (navigation failure, challenge page) are hard failures that abort the entire skill.

Report at the end:

• Which phases completed successfully
• Any warnings accumulated
• What data is available vs missing