---
name: web-clone
description: "[OMX] URL-driven website cloning with visual + functional verification"
---
Clone a target website from its URL, replicating both visual appearance and core interactive functionality. Uses Playwright MCP for live page extraction, LLM-driven code generation, and iterative verification with `$visual-verdict` for visual scoring.
- User provides a target URL and wants the site replicated as working code
- User says "clone site", "clone website", "copy webpage", or "web-clone"
- Task requires both visual fidelity AND functional parity with the original
- Reference is a live URL (not a static screenshot — use `$visual-verdict` for screenshot-only tasks)
- User only has screenshot references without a live URL — use `$visual-verdict` directly
- User wants to modify, redesign, or "improve" the site — use standard implementation flow
- Target requires authentication, payment flows, or backend API parity — out of scope for v1
- Multi-page / multi-route deep cloning — v1 handles single-page scope only
**v1 scope**: Single page clone of the provided URL.
Included:
- Layout structure (header, nav, content areas, sidebar, footer)
- Typography (font families, sizes, weights, line heights)
- Colors, spacing, borders, border-radius
- Core interactions: navigation links, buttons, form elements, dropdowns, modals, toggles
- Responsive hints from the extracted layout (flexbox/grid patterns)
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.
Playwright MCP server must be available for browser automation.
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`.
- `target_url` (required): The URL to clone
- `output_dir` (optional, default: current working directory): Where to generate the clone project
- `tech_stack` (optional, inferred from project context): HTML/CSS/JS, React, Vue, Svelte, etc.
- 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.
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.
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.
## 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**:
```javascript
(() => {
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**:
```javascript
(() => {
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 `
After each verification pass, emit a **composite web-clone verdict** JSON:
```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
- **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)
- **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.
**User**: "Clone https://news.ycombinator.com"
**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 `` 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.
- [ ] 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`