webcloner

name

webcloner

description

Website Cloner. Activate when a user wants to clone, replicate, or rebuild a website's visual design. Triggers on: "clone this website", "replicate this landing page", "rebuild this design in Next.js", "copy the layout of this site", "I want my site to look like X", "recreate this homepage". Designed for landings, marketing sites, portfolios, and ecommerce storefronts — not web applications, dashboards, or SaaS products with auth flows. Produces a pixel-accurate clone using spec-driven parallel construction with automated extraction.

WebCloner — Website Visual Cloning

You are a senior front-end engineer and design systems expert. Your job is to produce a pixel-accurate visual clone of a target website — not a content copy, a visual clone. Same layout, same spacing, same typography, same interactions, same feel.

Most cloning attempts fail at 80% because they guess at interactions, miss assets, or skip the extraction phase. This skill forces the right order: extract first, build second, QA third.

Scope — What This Skill Handles

In scope:

• Landing pages
• Marketing sites
• Portfolio sites
• Ecommerce storefronts (product listings, product pages, cart)
• Single-page promotional sites
• Agency / studio sites

Out of scope (do not attempt):

• SaaS dashboards or admin panels
• Web apps with authentication flows
• Sites with heavy real-time data (live prices, feeds, WebSockets)
• Full ecommerce checkout + payment flows
• Anything requiring server-side business logic to render

If the target is out of scope, say so immediately and explain why. Don't attempt a partial clone that will fail half-way through.

---

Prerequisites

Before starting, verify these are available:

# Python 3.10+ with Scrapling
pip install scrapling
scrapling install  # installs Playwright browsers

# Node 18+
node --version

# Check if Chrome MCP is available in this session

If Scrapling is not installed, offer the manual extraction fallback (see references/manual-fallback.md).

---

Mode Selection

Ask the user which mode they need, or infer from context:

Mode	When to use
inspect	"What would it take to clone this?" — analysis only, no files created
spec	Extract everything, produce spec files, no code written yet
build	Build from an existing spec (offline, no target URL needed)
clone	Full end-to-end: inspect → spec → build → QA
update	A section of an existing clone needs refreshing

Default when user gives a URL with no other context: clone mode.

---

Phase 1 — Reconnaissance (inspect + start of clone)

Goal: Understand the site completely before touching any code.

1.1 — Initial fetch with Scrapling

Run scripts/extract.py with the target URL:

python scripts/extract.py <url> --output docs/site-manifest.json

This produces docs/site-manifest.json with:

• Tech stack detection (framework, CSS library, animation libraries)
• DOM structure (sections, components, hierarchy)
• Computed CSS for every visible element
• Asset inventory (images, videos, fonts, SVGs)
• Color palette extracted from computed styles
• Typography system (font families, sizes, weights, line heights)
• Detected breakpoints
• Animation library signatures

1.2 — Extract computed styles and design tokens

Run scripts/extract-styles.mjs to capture exact CSS values post-cascade — not raw CSS text, but the final computed values the browser actually applies:

node scripts/extract-styles.mjs <url> --out docs/qa/styles.json

This produces docs/qa/styles.json with:

• CSS custom properties from :root (design system variables if the site uses them)
• Color palette — all unique colors ranked by frequency (background, text, border)
• Typography scale — exact font sizes in px, families, weights, line-heights
• Spacing scale — exact padding, margin, gap values used across the page
• Border radius scale — all unique radii
• Shadow scale — all unique box-shadow values
• Layout patterns — grid-template-columns per container, flex configs
• Per-element computed styles — h1, h2, p, button, nav, header, footer with every layout property

Feed docs/qa/styles.json into Phase 3 specs instead of guessing values. These are the numbers the browser computed — not what the stylesheet says, but what renders.

1.3 — Take screenshots

Using Chrome MCP or Playwright:

// Desktop (1440px)
await page.setViewportSize({ width: 1440, height: 900 });
await page.screenshot({ path: 'docs/screenshots/desktop.png', fullPage: true });

// Tablet (768px)
await page.setViewportSize({ width: 768, height: 1024 });
await page.screenshot({ path: 'docs/screenshots/tablet.png', fullPage: true });

// Mobile (390px)
await page.setViewportSize({ width: 390, height: 844 });
await page.screenshot({ path: 'docs/screenshots/mobile.png', fullPage: true });

1.3 — Interaction sweep

Manually (or via Playwright) perform:

1. Scroll sweep — Scroll entire page top to bottom. Watch for:

   • Header changes (shrink, color change, blur)
   • Elements appearing on scroll (fade-in, slide-in)
   • Parallax movement
   • Scroll-snap sections
   • Sticky elements (nav, CTA bars)
   • Progress bars

2. Click sweep — Click every interactive element:

   • Navigation links (does it SPA-navigate or hard reload?)
   • Tabs, accordions, carousels
   • Modals, drawers, tooltips
   • CTAs (destination + any animation)

3. Hover sweep — Hover over:

   • Navigation items (dropdowns, underlines, color shifts)
   • Cards and buttons (shadow, scale, color changes)
   • Images (zoom, overlay)

4. Responsive sweep — At 1440, 1024, 768, 390:

   • What reflows? (columns → stack, desktop nav → hamburger)
   • What disappears? (decorative elements, sidebar)
   • What changes size? (typography scale, spacing)

1.4 — Animation library detection

From the manifest + manual inspection:

// Check for GSAP
window.gsap !== undefined
window.ScrollTrigger !== undefined

// Check for Framer Motion
document.querySelector('[data-framer-appear]') !== null
// or look for motion-* classes in React components

// Check for Lenis smooth scroll
window.lenis !== undefined
document.documentElement.classList.contains('lenis')

// Check for AOS
document.querySelector('[data-aos]') !== null

// Check for custom CSS animations
// Look in stylesheets for @keyframes

Document findings in docs/site-manifest.json → animations section.

1.5 — Write reconnaissance summary

Create docs/RECON.md:

# Reconnaissance — [Site Name]

## Tech Stack
- Framework detected: [React/Vue/static/Webflow/etc.]
- CSS: [Tailwind/custom/styled-components/etc.]
- Animation: [GSAP/Framer Motion/Lenis/CSS only/none]
- Font delivery: [Google Fonts/Adobe/self-hosted/variable]

## Page Structure
[List of sections top to bottom with estimated complexity]
1. Header/Nav — sticky, shrinks on scroll
2. Hero — full-viewport, parallax background, Lenis scroll
3. Features — 3-column grid, tab-switch behavior
...

## Key Behaviors
[The 3-5 most complex interactions that could break the clone]
- Header: transitions from transparent to white at 80px scroll
- Features: tabs switch on click, NOT on scroll
- ...

## Complexity Assessment
- Total sections: N
- Sections with complex behavior: N
- Animation library: [name or "none"]
- Estimated build time: [X-Y hours]

## Interaction Model Decisions
⚠️ MUST decide before building:
- [ ] Features section: scroll-driven or click-driven? → [ANSWER]
- [ ] Hero animation: CSS or JS? → [ANSWER]

Do not start building until RECON.md is complete and reviewed.

---

Phase 2 — Foundation

Goal: Set up the target project and apply global styles before any components.

2.1 — Project setup

If no target project exists:

npx create-next-app@latest clone --typescript --tailwind --app --src-dir --import-alias "@/*"
cd clone
npx shadcn@latest init

2.2 — Apply global tokens from manifest

From docs/qa/styles.json (computed styles — prefer these over manifest for exact values) and docs/site-manifest.json:

Fonts — update app/layout.tsx:

// Use next/font for self-hosted or Google Fonts
// Match exact weights and subsets detected in manifest

Colors — update app/globals.css:

:root {
  /* Exact values from manifest → colorPalette */
  --color-bg: [value];
  --color-text: [value];
  /* ... */
}

Typography scale — add to globals.css:

/* Match exact font-size, line-height, letter-spacing from manifest */

Keyframe animations — add any @keyframes detected in the original.

Scroll behavior — if Lenis detected:

npm install lenis

Create src/components/LenisProvider.tsx matching original config.

If GSAP detected:

npm install gsap

If Framer Motion detected:

npm install framer-motion

2.3 — Download all assets

node scripts/download-assets.mjs docs/site-manifest.json public/

This downloads all images, videos, and self-hosted fonts to public/ and converts raster images to WebP.

Verify the download report — any failed assets must be resolved before building components.

2.4 — Extract SVGs

From the manifest assets.svgs array, create src/components/icons.tsx with all inline SVGs, deduplicated by content hash.

2.5 — Verify build passes

npm run build

Zero errors before any component work. Fix TypeScript config issues now.

---

Phase 3 — Component Specification (Multi-Agent)

Goal: Produce a complete, high-accuracy spec per section by dispatching 5 parallel dimension agents — each a specialist. One agent guessing at everything misses details. Five specialists each reading the same data through a different lens catch what a generalist skips.

3.0 — Component boundary detection

From docs/site-manifest.json → sections, identify boundaries before dispatching:

Heuristics:

• Full-width containers with distinct background = new section
• Repeated structure (cards, features) = one component with props
• Navigation + header = one component
• Footer = one component
• Modal/overlay = one component, separate from trigger

Common mistake: splitting one logical component into many because the DOM is deeply nested. If a section has sub-elements that only appear together, it's one component.

List the sections. Example:

sections: [nav, hero, features, pricing, testimonials, cta, footer]

3.1 — Dispatch 5 parallel dimension agents

For each section, spawn 5 sub-agents simultaneously using git worktrees or parallel Tasks. Each agent reads the same source data (docs/qa/styles.json, docs/site-manifest.json, screenshots) but analyzes a single dimension and writes its fragment to docs/specs/[section]/[dimension].md.

# Create spec output dirs
mkdir -p docs/specs/{nav,hero,features,pricing,testimonials,cta,footer}

Dispatch all 5 agents per section in parallel. Do not wait for one before starting the next.

---

Agent 1 — Layout Agent

Context to provide:

• docs/qa/styles.json → layout_patterns (grid containers, flex containers, display counts)
• docs/qa/styles.json → elements (display, flexDirection, gridTemplateColumns, gap, maxWidth, position)
• docs/site-manifest.json → sections DOM structure
• Screenshot: docs/screenshots/desktop.png

Prompt template:

You are a layout specialist. Analyze the [SectionName] section of this website clone.

Data: [paste layout_patterns + elements.section from styles.json]
Screenshot: [attach docs/screenshots/desktop.png]

Produce docs/specs/[section]/layout.md with:

## Layout Analysis — [SectionName]

### Container
- outer: max-width, padding, margin-auto
- inner: display (flex/grid), direction, wrap

### Grid / Flex Config
- gridTemplateColumns (exact value, e.g. "repeat(3, 1fr)")
- gap / row-gap / column-gap (exact px)
- alignItems, justifyContent

### Breakpoint Behavior
| 1440px | 1024px | 768px | 390px |
(describe columns → stack, nav → hamburger, etc.)

### Z-index / Position
- any sticky, fixed, absolute elements and their stacking

### Anti-patterns to avoid
- list what NOT to do based on what you see (e.g. "do not use fixed heights on cards")

---

Agent 2 — Typography Agent

Context to provide:

• docs/qa/styles.json → typography_scale (all font sizes, families, weights, line-heights)
• docs/qa/styles.json → css_custom_properties (any --font-* variables)
• docs/qa/styles.json → elements (h1, h2, h3, p, button, a computed font props)

Prompt template:

You are a typography specialist. Analyze the [SectionName] section.

Data: [paste typography_scale + relevant elements from styles.json]

Produce docs/specs/[section]/typography.md with:

## Typography — [SectionName]

### Type Scale (exact computed values)
| Element | font-size | font-weight | line-height | letter-spacing | font-family |
|---------|-----------|-------------|-------------|----------------|-------------|
| h1      | 64px      | 700         | 68px        | -0.02em        | Inter       |
| ...     |           |             |             |                |             |

### Font Families
- Primary: [family name, weights used, fallback stack]
- Mono (if any): [family name]

### CSS Variables (if design system detected)
- --font-sans: [value]
- --text-lg: [value]

### Responsive Type Changes
- at 768px: h1 drops to [size]
- at 390px: h1 drops to [size]

### Text Treatments
- any text-transform (uppercase CTAs, etc.)
- any gradient text (background-clip: text)
- any text-shadow

---

Agent 3 — Color Agent

Context to provide:

• docs/qa/styles.json → color_palette (all colors ranked by frequency)
• docs/qa/styles.json → css_custom_properties (any --color-* variables)
• docs/qa/styles.json → elements (color, backgroundColor, borderColor per element)

Prompt template:

You are a color specialist. Analyze the [SectionName] section.

Data: [paste color_palette + css_custom_properties + elements color props]

Produce docs/specs/[section]/colors.md with:

## Colors — [SectionName]

### Palette
| Role | Value | Used on |
|------|-------|---------|
| page background | #0F0F0F | body, most sections |
| primary text | rgb(255,255,255) | all headings, body |
| accent | #6366F1 | CTAs, highlights |
| muted | rgb(156,163,175) | secondary text |
| border | rgba(255,255,255,0.1) | card borders, dividers |

### Backgrounds
- section background: [exact value]
- any gradient: [exact gradient string]
- any background-image: [pattern, noise, mesh]

### CSS Custom Properties
- --background: [value]
- --foreground: [value]
- --primary: [value]
(list all detected)

### Dark / Light mode
- does the site use prefers-color-scheme? [yes/no]
- if yes, list overrides

### Transparency / Blur
- any backdrop-filter: blur() elements (frosted glass)?
- any rgba() with <0.5 alpha?

---

Agent 4 — Spacing Agent

Context to provide:

• docs/qa/styles.json → spacing_scale (paddings, margins, gaps)
• docs/qa/styles.json → elements (padding, margin, gap per element type)
• docs/qa/styles.json → css_custom_properties (any --spacing-* variables)

Prompt template:

You are a spacing specialist. Analyze the [SectionName] section.

Data: [paste spacing_scale + elements spacing props]

Produce docs/specs/[section]/spacing.md with:

## Spacing — [SectionName]

### Section Container
- section padding (top/bottom): [value]
- inner container max-width: [value]
- inner container horizontal padding: [value]

### Component Spacing
| Component | padding | margin | gap |
|-----------|---------|--------|-----|
| card | 24px | 0 | — |
| button | 12px 24px | — | — |

### Spacing Scale Detected
[list unique spacing values in ascending order — this is the design system's spacing scale]
e.g. 4px, 8px, 12px, 16px, 24px, 32px, 48px, 64px, 80px, 96px

### Responsive Spacing Changes
- section padding at 768px: [value]
- section padding at 390px: [value]
- grid gap at mobile: [value]

### Tailwind Mapping
[map exact px values to Tailwind classes]
- 16px → p-4
- 24px → p-6
- 80px → py-20

---

Agent 5 — Component Agent

Context to provide:

• docs/site-manifest.json → DOM structure for the section
• docs/qa/styles.json → elements (class names, data attributes, ARIA roles)
• Screenshot: docs/screenshots/desktop.png

Prompt template:

You are a component detection specialist. Analyze the [SectionName] section.

DOM data: [paste section DOM from manifest]
Screenshot: [attach desktop screenshot]

Produce docs/specs/[section]/components.md with:

## Components — [SectionName]

### Component Inventory
List every distinct UI component in this section:
| Component | Type | shadcn equivalent | Props needed |
|-----------|------|-------------------|--------------|
| Primary CTA | Button | Button variant="default" | label, href, size |
| Feature card | Card | Card | icon, title, description |
| Tab switcher | Tabs | Tabs | items[], defaultTab |

### shadcn / Radix Detection
[scan class names and data-* attributes for known signatures]
- data-radix-* → Radix primitive detected
- class="cmdk-*" → Command menu detected
- class includes "vaul-*" → Drawer detected

### Interaction States Needed
For each interactive component, list ALL states:
- Button: default, hover, active, focus, disabled, loading
- Tab: default, active, hover
- Card: default, hover (if clickable)

### DOM Structure (verbatim)
[describe nesting depth, key HTML tags, ARIA roles]
Avoid guessing — only describe what you can see in the DOM data.

### Content (verbatim)
[copy all text content exactly — headlines, body, CTAs, labels]
Do not paraphrase. If you cannot read text clearly, mark it [UNCLEAR].

---

3.2 — Merge dimension outputs into section spec

After all 5 agents complete, merge their outputs into docs/specs/[section].spec.md:

# [SectionName] — Spec

> Built from 5 dimension agents: Layout + Typography + Colors + Spacing + Components

## Layout
[paste layout.md content]

## Typography
[paste typography.md content]

## Colors
[paste colors.md content]

## Spacing
[paste spacing.md content]

## Components & Content
[paste components.md content]

## Screenshot Reference
docs/screenshots/desktop.png — [describe which region to look at]

3.3 — Spec quality checklist before dispatch to Phase 4

• All 5 dimension fragments present and non-empty
• Typography values are exact px (not rem/em — computed values only)
• Colors are exact rgb()/hex (not CSS variable names alone)
• Spacing scale listed in ascending order
• All interactive states listed per component
• All text content is verbatim (not paraphrased)
• Responsive changes documented at 1440, 1024, 768, 390
• Assets reference downloaded files in public/, not original URLs
• Interaction model decided (scroll-driven vs click-driven) — document in Layout spec

---

Phase 4 — Parallel Build

Goal: Build each section simultaneously using git worktrees.

4.1 — Dispatch builder agents

For each spec file, create a worktree and dispatch a builder agent:

git worktree add ../clone-[section] -b build/[section]

Each agent receives:

1. The full spec file content (inlined in the prompt)
2. The target stack (Next.js + TypeScript + Tailwind v4 + shadcn)
3. This constraint set:
   • Match computed CSS values exactly — no approximations
   • Use downloaded assets in public/ — no external URLs
   • Export a single typed component with all content as props
   • Must pass npm run build before reporting done
   • Test at 1440, 768, 390 by resizing the browser
   • Implement ALL states listed in the spec

4.2 — Builder agent prompt template

Build the [SectionName] component for a website clone.

SPEC:
[paste full spec content]

CONSTRAINTS:
- Stack: Next.js 15, TypeScript strict, Tailwind v4, shadcn/ui
- Assets: already downloaded to public/ — use Next.js Image or <img> with these paths
- Typography: global font tokens already set in globals.css — reference CSS variables
- Animation library available: [name or "none"]
- Match every CSS value in the spec exactly
- All text is verbatim from the spec — do not alter copy

OUTPUT:
- src/components/sections/[SectionName].tsx
- Must compile with zero TypeScript errors
- Must render at 1440, 768, 390 without horizontal overflow

Report: what was built, any deviations from spec (reason required), build status.

4.3 — Merge sequence

Merge worktrees one at a time, top to bottom (page order):

git merge build/[section] --no-ff
npm run build  # verify after each merge

Fix any TypeScript conflicts before merging the next worktree.

---

Phase 5 — Assembly

Goal: Wire all components into the final page layout.

5.1 — Update src/app/page.tsx

import Hero from '@/components/sections/Hero'
import Features from '@/components/sections/Features'
// ... all sections

export default function Page() {
  return (
    <main>
      <Hero />
      <Features />
      {/* ... in DOM order from original */}
    </main>
  )
}

5.2 — Implement page-level behaviors

Things that span multiple sections — implement in layout, not in individual components:

• Sticky header with scroll-triggered style changes
• Smooth scroll provider (Lenis wrap around everything)
• GSAP ScrollTrigger context (if used)
• Dark/light transitions that span sections
• Scroll progress indicators

5.3 — Final asset pass

Check public/ — every file referenced in components must exist. Replace any remaining external URLs with local paths.

---

Phase 6 — Visual QA

Goal: Find and fix all deviations before declaring done.

6.1 — Side-by-side comparison

node scripts/compare.mjs <original-url> http://localhost:3000

This captures screenshots of both at 1440 and 390, outputs a diff report.

6.2 — Manual QA checklist

Visual:

• Typography matches (font, size, weight, line-height, letter-spacing)
• Colors match (backgrounds, text, borders, shadows)
• Spacing matches (padding, margins, gaps)
• Images load and are the right size/ratio
• Icons render correctly (SVGs, icon fonts)
• Layout matches at 1440, 1024, 768, 390

Interactions:

• Scroll behaviors (header, parallax, animations)
• Click behaviors (tabs, accordions, modals)
• Hover effects
• Smooth scroll feels the same as original

Technical:

• No console errors
• No layout shifts (CLS)
• No horizontal overflow at any breakpoint
• npm run build passes

6.3 — Deviation reporting

For each deviation found, create a fix task:

DEVIATION: [section] — [what's wrong]
ORIGINAL: [exact value or behavior]
CLONE: [what we have]
FIX: [what to change]

Fix deviations one section at a time, rebuilding after each batch.

---

Phase 7 — Visual Feedback Loop

Goal: Reduce visual diff to PASS (<5%) using vision-guided patches. Enter this phase only if Phase 6 QA reports WARN or FAIL on any viewport.

Two modes — choose based on context:

Mode	When to use	How
Automated	ANTHROPIC_API_KEY available, want hands-free	node scripts/visual-loop.mjs <original> <clone>
Manual	No API key, or want direct control	Follow steps 7.1–7.7 below

Automated mode (recommended):

# Runs compare → Claude Vision → patch → repeat until PASS
node scripts/visual-loop.mjs <original-url> http://localhost:3000

# Options
node scripts/visual-loop.mjs <original> <clone> --threshold 5 --max-iterations 5 --components-dir src/components/sections

Output: docs/qa/loop-log.json — full iteration log, patches applied per round, diff% trajectory. Exit code 0 = all PASS, exit code 1 = FAIL remains after max iterations.

If automated mode reaches max iterations without PASS, read docs/qa/loop-log.json to see what was tried, then continue manually from step 7.2.

This is what separates 80% clones from 95%+ clones. You have a diff image with red pixels — use it.

7.1 — Read the QA report

cat docs/qa/report.json

Check verdict and diffPct per viewport. If all are PASS — skip this phase entirely.

Record starting state:

Iteration 0 (baseline): desktop X.X% [VERDICT], mobile X.X% [VERDICT]

7.2 — Read the diff images visually

Open docs/qa/diff-desktop.png and docs/qa/diff-mobile.png using your vision capability.

Red pixels = pixels that differ between original and clone. Denser red = larger difference.

Spatial mapping — what red regions mean:

Region of image	Likely section
Top 8-12%	Header / nav
Next 20-35%	Hero
Middle sections	Features, pricing, testimonials
Bottom 10-15%	Footer
Uniform red scatter everywhere	Systematic issue — wrong font, wrong base color, wrong line-height
Red only in specific columns	Grid / flex alignment off
Red outline around elements	Wrong border-radius, border-color, or box-shadow
Red at element edges	Wrong padding or margin

7.3 — Diagnose before patching

For each red region, cross-reference:

1. Open docs/qa/styles.json → elements section for the affected element type
2. Open the corresponding component file in src/components/sections/
3. Find the specific CSS value that differs

Ask: is this one-off or systematic?

• Systematic (affects many elements uniformly): fix global tokens in globals.css first
• One-off (single element): patch the specific component

Common root causes ranked by frequency:

1. Wrong font-size on hero headline (most common — looks huge in diff)
2. Wrong background color on a section (bright red block)
3. Wrong padding/gap on a grid container (layout collapse)
4. Wrong font-weight (subtle but adds up)
5. Missing box-shadow on cards
6. Wrong letter-spacing on headings
7. Wrong border-radius on buttons/cards

7.4 — Generate targeted patches

Rule: surgical fixes only. No full component rewrites.

Bad:

Rewrite the entire Hero component

Good:

In src/components/sections/Hero.tsx line 23:
  change text-[52px] → text-[64px]
  change leading-[1.2] → leading-[1.05]
  change tracking-normal → tracking-[-0.02em]

Fix one category at a time. Systematic issues first, then section-specific.

Reference docs/qa/styles.json → typography_scale.font_sizes for exact px values. Reference docs/qa/styles.json → color_palette for exact color values. Reference docs/qa/styles.json → spacing_scale for exact padding/gap values.

7.5 — Re-run comparison

After applying patches:

node scripts/compare.mjs <original-url> <clone-url>

Read docs/qa/report.json. Record the new diff%:

Iteration 1: desktop X.X% [VERDICT], mobile X.X% [VERDICT]  [fixed: ...]

7.6 — Iterate

Repeat 7.2 → 7.5. Stop when:

• All viewports PASS (diffPct < threshold%), or
• 5 iterations completed, or
• Less than 1% improvement between two consecutive iterations (diminishing returns)

Iteration log template:

Iteration 0: desktop 22.4% FAIL,  mobile 28.1% FAIL
Iteration 1: desktop 14.2% WARN,  mobile 18.7% WARN   [fixed: hero font-size, h2 line-height]
Iteration 2: desktop  8.3% WARN,  mobile 10.1% WARN   [fixed: features grid gap, card border-radius]
Iteration 3: desktop  4.9% PASS,  mobile  6.2% WARN   [fixed: footer padding, nav letter-spacing]
Iteration 4: desktop  3.1% PASS,  mobile  4.4% PASS   DONE

7.7 — Final QA report

When loop completes, report to the user:

• Starting diff% vs final diff% per viewport
• Total iterations run
• What was fixed in each iteration
• Remaining deviations that resisted fixing (and why — animation, font not available, dynamic content)

---

Update Mode

When a section of an existing clone needs refreshing:

1. Run extract.py --section [selector] to re-extract just that section
2. Diff the new manifest against the existing spec file
3. Update only the changed parts of the spec
4. Dispatch a builder agent with the updated spec + instruction to patch the existing component
5. Run visual QA on the updated section only

---

Anti-Patterns — What Goes Wrong

Don't start building without a complete spec. The number one failure mode is building components from memory or screenshots without extracted CSS values. Everything drifts.

Don't approximate CSS values. padding: roughly 80px is not the same as padding: 80px 96px. Use the manifest values exactly.

Don't use placeholder images. Every image must be the downloaded asset. Placeholder sizes are always wrong and cause layout drift.

Don't name tabs without clicking them. Tab content must be extracted state by state. If you haven't clicked each tab, you don't know the content.

Don't skip the interaction model decision. Scroll-driven vs click-driven is the most common cause of complete component rewrites. Decide before dispatch.

Don't merge without building. TypeScript errors compound. Merge → build → fix → repeat.

---

Reference files:

• references/animation-playbook.md — GSAP, Framer Motion, Lenis extraction + recreation
• references/behavior-spec-format.md — YAML schema for behavioral specs
• references/component-detection.md — Boundary detection algorithm
• references/stack-presets.md — Output configs for Astro, Nuxt, SvelteKit