Web Clone

Excluded:

• Backend API integration or data fetching
• Authentication flows or protected content
• Dynamic/personalized content (user-specific data)
• Multi-page crawling or route graph cloning
• Third-party widget functionality (maps, embeds, chat widgets)
• Image/asset replication (use placeholders for external images)

Legal notice: Only clone sites you own or have explicit permission to replicate. Respect copyright and trademarks. </Scope_Limits>

1. Before first tool use, call ToolSearch("browser") or ToolSearch("playwright") to discover available browser tools.
2. If no browser tools are found, instruct the user:

   Playwright MCP is required. Configure it:
   codex mcp add playwright npx "@playwright/mcp@latest"
   

3. Required tools: browser_navigate, browser_snapshot, browser_take_screenshot, browser_evaluate, browser_wait_for. Optional: browser_click, browser_network_requests. </Prerequisites>

<Tool_Usage>

• Before first MCP tool use, call ToolSearch("browser") or ToolSearch("playwright") to discover deferred Playwright MCP tools.
• If no browser tools are found, stop immediately and instruct the user to configure Playwright MCP.
• Use browser_snapshot (accessibility tree) for structural understanding — it is far more token-efficient than screenshots.
• Use browser_take_screenshot only when visual verification is needed (Pass 1 baseline, Pass 4 comparison).
• Use browser_evaluate for DOM/style extraction — pass the scripts from this skill EXACTLY as written (do not modify them).
• If running within ralph, use state_write / state_read for web-clone state persistence between iterations.
• Skip Codex consultation for straightforward extraction; use it only if verification repeatedly fails on the same issue. </Tool_Usage>

<State_Management> Persist extraction and progress data so the pipeline can resume if interrupted.

• After Pass 1 completes: Write extraction summary to .omx/state/{scope}/web-clone-extraction.json containing:
  • target_url, extracted_at timestamp
  • screenshot_path (path to target-full.png)
  • landmark_count (number of nav, main, footer, form elements)
  • interactive_count (number of detected interactive elements)
  • extraction_size_kb (approximate size of DOM extraction data)
• After each Pass 4 verification: Append the composite verdict to .omx/state/{scope}/web-clone-verdicts.json.
• When running within ralph: Also persist the visual portion of the composite verdict to .omx/state/{scope}/ralph-progress.json for ralph compatibility, mapping visual.score → top-level score and visual.verdict → top-level verdict.
• On completion or failure: Write final status with completed_at or failed_at timestamp. </State_Management>

<Context_Budget> Pass 1 extraction can produce very large data. Apply these limits proactively:

• DOM tree: If the serialized JSON exceeds ~30KB, reduce depth parameter from 8 to 4 and re-extract. Focus on top-level structure.
• Accessibility snapshot: If it exceeds ~20KB, this is normal for complex pages. Summarize key landmarks rather than keeping the full tree.
• Interactive elements: Cap at 50 elements. If more exist, keep only visible ones (isVisible: true).
• Total extraction context: Aim for under 60KB combined. If exceeded, prioritize: screenshot > accessibility snapshot > interactive elements > DOM styles.
• Image tokens: Full-page screenshots are expensive. Take one baseline in Pass 1 and one comparison in Pass 4. Do not take screenshots between iterations unless debugging a specific region. </Context_Budget>

Pass 1 — Extract

Capture the target page's structure, styles, interactions, and visual baseline.

1. Navigate: browser_navigate to target_url.
2. Wait for render: browser_wait_for with appropriate condition (network idle or timeout of 5s) to ensure full render including lazy-loaded content.
3. Accessibility snapshot: browser_snapshot — captures the semantic tree (roles, names, values, interactive states). This is your primary structural reference.
4. Full-page screenshot: browser_take_screenshot with fullPage: true — save as reference baseline target-full.png.
5. DOM + computed styles: browser_evaluate with the following script. COPY THIS SCRIPT EXACTLY — do not modify it:

   (() => {
     const walk = (el, depth = 0) => {
       if (depth > 8 || !el.tagName) return null;
       const cs = window.getComputedStyle(el);
       return {
         tag: el.tagName.toLowerCase(),
         id: el.id || undefined,
         classes: [...el.classList].slice(0, 5),
         styles: {
           display: cs.display, position: cs.position,
           width: cs.width, height: cs.height,
           padding: cs.padding, margin: cs.margin,
           fontSize: cs.fontSize, fontFamily: cs.fontFamily,
           fontWeight: cs.fontWeight, lineHeight: cs.lineHeight,
           color: cs.color, backgroundColor: cs.backgroundColor,
           border: cs.border, borderRadius: cs.borderRadius,
           flexDirection: cs.flexDirection, justifyContent: cs.justifyContent,
           alignItems: cs.alignItems, gap: cs.gap,
           gridTemplateColumns: cs.gridTemplateColumns,
         },
         text: el.childNodes.length === 1 && el.childNodes[0].nodeType === 3
           ? el.textContent?.trim().slice(0, 100) : undefined,
         children: [...el.children].map(c => walk(c, depth + 1)).filter(Boolean),
       };
     };
     return walk(document.body);
   })()
   

6. Interactive elements: browser_evaluate to catalog all interactable elements. COPY THIS SCRIPT EXACTLY — do not modify it:

   (() => {
     const results = [];
     document.querySelectorAll(
       'button, a[href], input, select, textarea, [role="button"], ' +
       '[onclick], [aria-haspopup], [aria-expanded], details, dialog'
     ).forEach(el => {
       results.push({
         tag: el.tagName.toLowerCase(),
         type: el.type || el.getAttribute('role') || 'interactive',
         text: (el.textContent || '').trim().slice(0, 80),
         href: el.href || undefined,
         ariaLabel: el.getAttribute('aria-label') || undefined,
         isVisible: el.offsetParent !== null,
       });
     });
     return results;
   })()
   

7. Network patterns (optional): browser_network_requests — note XHR/fetch calls for reference. Do not attempt to replicate backends.

Keep all extraction results in working memory for Pass 2.

Pass 2 — Build Plan

Analyze extraction results and decompose into a component plan.

1. Identify page regions: From DOM tree + accessibility snapshot, identify major sections:

   • Navigation bar / header
   • Hero / banner section
   • Main content area(s)
   • Sidebar (if present)
   • Footer
   • Overlay elements (modals, drawers)

2. Map components: For each region, define:

   • Component name and responsibility
   • Key style properties (from computed styles)
   • Content summary (headings, text, images)
   • Child components if nested

3. Create interaction map: From interactive elements list:

   • Navigation links → anchor tags with href
   • Form elements → proper <form> with inputs, labels, validation
   • Buttons → click handlers (toggle, submit, navigate)
   • Dropdowns/modals → show/hide toggle with transitions
   • Accordions/tabs → state-based visibility

4. Extract design tokens: Identify recurring values:

   • Color palette (primary, secondary, background, text colors)
   • Font stack (families, size scale, weight scale)
   • Spacing scale (padding/margin patterns)
   • Border radius values

5. Define file structure:

   {output_dir}/
   ├── index.html          (or App.tsx / App.vue)
   ├── styles/
   │   ├── globals.css      (reset + tokens)
   │   └── components.css   (or scoped styles)
   ├── scripts/
   │   └── interactions.js  (toggle, modal, dropdown logic)
   └── assets/              (placeholder images)
   

   Adapt to tech_stack if specified (React components, Vue SFCs, etc.).

Pass 3 — Generate Clone

Implement the clone from the plan. Work component-by-component.

1. Scaffold: Create the directory structure and base files.
2. Design tokens first: Implement CSS custom properties or Tailwind config from extracted tokens.
3. Layout shell: Build the page-level layout matching the original's flexbox/grid structure.
4. Components: Implement each region top-down:
   • Match DOM structure from extraction (semantic tags, landmark roles)
   • Apply computed styles — prioritize layout properties, then typography, then decorative
   • Use actual extracted text content; use placeholder <img> for external images
5. Interactions: Wire up detected behaviors:
   • Navigation: working <a> tags (to # anchors or stubs for v1)
   • Forms: proper structure with <label>, input types, placeholder text
   • Toggles: JavaScript for dropdowns, modals, accordions
   • Hover/focus states: CSS transitions matching original behavior
6. Responsive: If the original uses responsive breakpoints (detectable from media queries in computed styles or from viewport behavior), add basic responsive rules.

Pass 4 — Verify

Compare the clone against the original across three dimensions.

1. Serve the clone: Start a local server for the generated project:

   npx serve {output_dir} -l 3456 --no-clipboard
   

   If npx serve is unavailable, fall back to: python3 -m http.server 3456 -d {output_dir}. The clone will be accessible at http://localhost:3456.

2. Visual verification:

   • Navigate to the clone with Playwright: browser_navigate to clone URL.
   • Take full-page screenshot of clone.
   • Run $visual-verdict with: reference_images=["target-full.png"], generated_screenshot="clone-full.png", category_hint="web-clone".
   • The visual portion of the verdict feeds directly into the composite verdict below.
   • Visual pass threshold: score >= 85.

3. Structural verification: Compare landmark counts:

   • Count <nav>, <main>, <footer>, <form>, <button>, <a> in both original and clone.
   • Structure passes when all major landmarks exist (missing landmarks = fail).

4. Functional spot-check: Test 2–3 detected interactions via Playwright:

   • Click a navigation link → verify URL change or scroll behavior
   • Toggle a dropdown/modal → verify visibility change
   • Interact with a form field → verify it accepts input
   • Use browser_click and browser_snapshot to verify state changes.

5. Emit composite verdict:

{
  "visual": {
    "score": 82,
    "verdict": "revise",
    "category_match": true,
    "differences": ["Header spacing tighter than original"],
    "suggestions": ["Increase nav gap to 24px"]
  },
  "functional": {
    "tested": 3,
    "passed": 2,
    "failures": ["Dropdown does not open on click"]
  },
  "structure": {
    "landmark_match": true,
    "missing": [],
    "extra": []
  },
  "overall_verdict": "revise",
  "priority_fixes": [
    "Fix dropdown toggle interaction",
    "Increase header nav spacing"
  ]
}

Pass 5 — Iterate

Fix highest-impact issues and re-verify.

1. Prioritize fixes by impact: layout > interactions > spacing > typography > colors.
2. Apply targeted edits: Fix only the issues listed in priority_fixes. Do not refactor working code.
3. Re-verify: Repeat Pass 4.
4. Loop: Continue until overall_verdict is pass OR max 5 iterations reached.
5. Final report: Summarize what was successfully cloned, any remaining differences, and elements that could not be replicated.

<Output_Contract> After each verification pass, emit a composite web-clone verdict JSON:

{
  "visual": {
    "score": 0,
    "verdict": "revise",
    "category_match": false,
    "differences": ["..."],
    "suggestions": ["..."],
    "reasoning": "short explanation"
  },
  "functional": {
    "tested": 0,
    "passed": 0,
    "failures": ["..."]
  },
  "structure": {
    "landmark_match": false,
    "missing": ["..."],
    "extra": ["..."]
  },
  "overall_verdict": "revise",
  "priority_fixes": ["..."]
}

Rules:

• visual follows the VisualVerdict shape from $visual-verdict
• functional.tested/passed are counts; failures list specific interaction failures
• structure.landmark_match is true when all major HTML landmarks (nav, main, footer, forms) are present
• overall_verdict: pass when visual.score >= 85 AND functional.failures is empty AND structure.landmark_match is true
• priority_fixes: ordered by impact, drives the next iteration </Output_Contract>

<Iteration_Thresholds>

• Visual pass: score >= 85
• Functional pass: zero failures on tested interactions
• Structure pass: all major landmarks present
• Overall pass: all three dimensions pass
• Max iterations: 5 (report best achieved result if threshold not met) </Iteration_Thresholds>

<Error_Handling>

• Playwright MCP unavailable: Stop. Instruct user to configure it. Do not attempt to clone without browser tools.
• Page fails to load: Report the URL and HTTP status. Suggest the user verify the URL is accessible.
• browser_evaluate returns empty: The page may use heavy client-side rendering. Wait longer (browser_wait_for with extended timeout) and retry once.
• Visual score stuck below threshold after 3 iterations: Report the current state as best-effort. List the unresolved differences for the user.
• Extraction data too large for context: Truncate deep DOM branches (depth > 6). Focus on top-level structure and defer nested details to iteration fixes. </Error_Handling>

Pass 1: Navigate to HN. Extract: table-based layout, orange (#ff6600) nav bar, story list with links + points + comments, footer. Screenshot saved.

Pass 2: Regions: nav bar (logo + links), story table (30 rows × title + meta), footer. Tokens: orange #ff6600, gray #828282, Verdana font, 10pt base. Interaction map: story links (external), comment links, "more" pagination.

Pass 3: Generate index.html with HN-style table layout, CSS matching extracted colors/fonts, working <a> tags for stories.

Pass 4: Visual score=78 (font size off, spacing between stories too tight). Functional 2/2 (links work). Structure match=true.

Pass 5 iteration 1: Fix font to Verdana 10pt, increase row padding → score=88. Functional 2/2. Structure match. → overall_verdict: pass. Done. </Example>

<Final_Checklist>

• Pass 1 extraction completed and summarized (screenshot + accessibility tree + DOM styles + interactions)
• Pass 2 component plan created with file structure
• Pass 3 clone generated and files written to output_dir
• Clone serves locally without errors
• Pass 4 composite verdict emitted with all three dimensions
• overall_verdict is pass, or max 5 iterations reached with best-effort report
• When in ralph: visual verdict persisted to ralph-progress.json
• Extraction summary persisted to web-clone-extraction.json </Final_Checklist>