Clone Design Pipeline

Phase 2: ANALYZE — Extract Design Tokens

Read extraction-report.md and design-tokens.json. Extract:

• Color palette — every hex used, grouped by role (background, text, accent, border)
• Typography — font families, sizes, weights per element type
• Spacing — section gaps, card padding, element margins
• Components — identify cards, buttons, navs, hero sections, footers
• Layout — section order, background color per section, full-page height

Phase 3: VISUAL STUDY — Section-by-Section Analysis

Use the image tool to analyze each screenshots/section-N.png:

• Describe every element, its position, colors, typography
• Note exact content (text, labels, CTAs)
• Identify images, illustrations, icons that need compositing
• Map section boundaries and background transitions

Phase 4: GENERATE — Stitch Per Section (Multi-Pass)

For long pages, generate in 2-3 Stitch calls max (Stitch handles long prompts well). Group sections logically:

• Pass 1: Hero + above-fold sections
• Pass 2: Mid-page content sections
• Pass 3: Footer + city grid + final sections

Each prompt must include:

• Exact hex colors from extraction
• Exact font families and sizes
• Exact content text (copy from source.html or visual analysis)
• Exact layout structure and spacing
• References to extracted image positions ("170px tall image placeholder at position X")
• deviceType="MOBILE" for mobile clones

Use edit_screens for refinement passes rather than regenerating from scratch.

Phase 5: PULL — Get Generated Code

MCPORTER_CALL_TIMEOUT=60000 mcporter call stitch.get_screen \
  name="projects/{pid}/screens/{sid}" projectId="{pid}" screenId="{sid}" --output json

Download the HTML code via the downloadUrl. Save to stitch-pulls/<project>/.

Phase 6: COMPOSITE — Inject Real Assets

NODE_PATH=/data/.npm-global/lib/node_modules node skills/clone-design/scripts/compose-assets.js \
  <generated.html> <extract-dir> <output.html>

This script:

• Replaces placeholder/stock images with extracted real images
• Injects @font-face declarations for downloaded fonts
• Preserves inline SVGs from extraction

After auto-compositing, manually review and fix:

• Image ordering (script matches sequentially — may need reordering)
• SVG icons that need injection at specific positions
• Font family names in CSS that don't match extracted names
• Any remaining placeholder content

Phase 7: RENDER — Screenshot the Clone

NODE_PATH=/data/.npm-global/lib/node_modules node -e "
const { chromium } = require('playwright');
(async () => {
  const browser = await chromium.launch();
  const page = await browser.newPage({ viewport: { width: 390, height: 844, deviceScaleFactor: 2 } });
  await page.goto('file:///path/to/composed.html');
  await page.waitForTimeout(4000);
  await page.screenshot({ path: '/path/to/output.png', fullPage: true });
  await browser.close();
})();
"

Phase 8: COMPARE — Visual Diff

NODE_PATH=/data/.npm-global/lib/node_modules node skills/clone-design/scripts/visual-diff.js \
  <extract-dir> <composed.html> <diff-output-dir>

Then use the image tool to compare each section pair:

• Reference screenshots/section-N.png vs Clone clone-section-N.png
• Score each section on: layout, colors, typography, content, assets
• Identify specific pixel-level differences

Phase 9: ITERATE — Fix and Re-render

Based on comparison findings:

1. Edit the HTML directly for small fixes (colors, spacing, text)
2. Use edit_screens in Stitch for structural changes
3. Re-composite assets if images shifted
4. Re-render and re-compare

Max 3 iterations. After 3, assess if the approach needs rethinking.

Quality Targets

File Organization

stitch-pulls/<project>/
├── extract/                  # Phase 1 output
│   ├── screenshots/          # Reference screenshots
│   ├── assets/               # Downloaded images, SVGs, fonts
│   ├── design-tokens.json    # Structured tokens
│   ├── extraction-report.md  # Human-readable summary
│   └── source.html           # Raw source
├── generated.html            # Phase 5 Stitch output
├── composed.html             # Phase 6 with real assets
├── final-rendered.png        # Phase 7 screenshot
└── diff/                     # Phase 8 comparison
    ├── clone-section-N.png
    └── diff-report.md

Generating Original Designs (Non-Clone)

When designing original screens (not cloning an existing site), skip Phases 1-3 and 6. Instead:

1. Use brand tokens from SOUL.md / brand-systems skill
2. Craft a detailed Stitch prompt with exact specs (Phase 4)
3. Pull and render (Phases 5, 7)
4. Run the standard 10-point design review (from design-review-protocol skill)
5. Iterate on code directly for precise fixes
6. Max 3 iterations before reassessing approach

Tips

• Always set MCPORTER_CALL_TIMEOUT=180000 for Stitch generation calls
• Use modelId="GEMINI_3_1_PRO" for highest quality generation
• Copy exact text content from source.html — never paraphrase
• For emoji rendering, ensure fonts-noto-color-emoji is installed
• For dark themes, extract colors carefully — rgba(0,0,0,0) is transparent, not black
• Long pages (>3000px) should be generated in 2-3 passes for better fidelity