Pix

/pix: The Pixel-Perfect Autonomous Loop

User provides a Figma link. You implement it pixel-perfect. That's it.

Tool names may vary based on your MCP configuration (e.g., figma__get_screenshot or mcp__figma__get_screenshot). Run /mcp to see available tools.

Resource Strategy

Tool Costs

Core rule: Cheap calls first to build a map, then expensive calls only on the smallest necessary nodes.

Image Budget

Claude API limits images to 2000px per dimension when >20 images are in a conversation. Old images cannot be individually removed.

• ~20 screenshots per conversation before the resolution limit kicks in
• 3-image working set: Figma layout (overview), Figma detail (current section), Chrome state (rendered result)
• Never re-screenshot a static Figma node — designs don't change mid-session
• Use getComputedStyle for numerical properties — zero image cost
• Recovery: /compact drops old images from context (see Recovery section)

Phase 1: Reconnaissance

When the user provides a Figma link, extract fileKey and nodeId from the URL and immediately start working. If any MCP call or Chrome interaction fails, run the Doctor Check to diagnose and fix, then resume.

Step 1: Get the Node Tree

Call get_metadata(nodeId, fileKey):

<Frame id="45:1" name="Header" type="FRAME" x="0" y="0" width="1440" height="80">
  <Component id="45:10" name="Logo" x="20" y="16" width="120" height="48"/>
  <Frame id="45:15" name="NavLinks" x="400" y="20" width="600" height="40">
    <Text id="45:16" name="Home" x="0" y="0" width="60" height="40"/>
  </Frame>
</Frame>

Build a mental map:

• Identify every major section and note each nodeId
• Note positions and sizes to understand the layout grid
• Identify components vs frames vs text vs icons

Step 2: Extract All Design Tokens (Once)

Call get_variable_defs(nodeId, fileKey). Save the result — do NOT call this again.

Step 3: Check Existing Components

Call get_code_connect_map(nodeId, fileKey). For matched components, follow this priority:

1. Reuse as-is — if it covers the Figma design exactly
2. Extend minimally — add a prop or variant if close but not exact
3. Compose — combine existing components
4. Create new — only if nothing existing fits

Step 4: Visual Overview

Call get_screenshot(nodeId, fileKey) on the root selection. This is Image 1 of your budget — your layout reference. Do NOT retake it.

At this point you have NOT called get_design_context at all. You have a complete structural map, all tokens, reusable component info, and a visual reference — all from cheap calls.

Phase 2: Study & Implement (Code in the Dark)

Study the design deeply, memorize every detail, then code from memory.

Step 1: Layout Shell

Using the metadata tree, implement the outer layout:

• Container dimensions and positioning (flex, grid)
• Major section placement
• Background colors from tokens
• Borders and dividers

The metadata + visual overview are usually sufficient. Only call get_design_context if you need specific properties you can't infer.

Step 2: Study Every Section

For each major section:

1. Figma Detail: get_screenshot(sectionNodeId, fileKey) — zoomed-in visual reference.

2. Design Context: get_design_context(sectionNodeId, fileKey) — text only, no image cost. If truncated, use child node IDs from get_metadata and fetch children individually.

3. Absorb every detail: fonts, sizes, weights, colors, spacing, borders, shadows, icon shapes. Burn it into memory.

4. Color sanity check: Compare get_design_context colors against the Figma screenshot. If the screenshot reveals opacity layering, overlapping fills, or gradients — the raw token values will be wrong. Use the visual truth, not the raw token.

No Chrome screenshots during this phase. You are studying, not checking.

Key rule: NEVER call get_design_context on the root selection. Always target the smallest meaningful node.

Step 3: Code from Memory

Implement everything using what you memorized:

• Design context output for exact properties
• Tokens from Phase 1 (do not re-extract)
• Reusable components from Phase 1
• Respect project rules: Check for .claude/rules, CLAUDE.md, and project instruction files. Follow established patterns for component structure, file placement, naming, and styling. Figma MCP output is a design representation — translate it into your project's conventions, don't paste it verbatim.

Frontend only. Don't touch backend, API routes, or database unless the user explicitly asks. Use mock/placeholder data if APIs don't exist yet.

Minimize screenshots during coding. You studied the design — use what you memorized. But if you're missing crucial data for a specific element (exact icon shape, a nested layout you didn't drill into, a subtle gradient), take a targeted get_screenshot on that Figma node rather than guessing. Getting it right the first time is cheaper than a refinement round.

Step 4: Design System Sync

NEVER hardcode values. Sync to the project's design system:

• Tailwind: Update tailwind.config.* if a token is missing. Never text-[#f3f3f3].
• CSS-in-JS: Add tokens to the theme object. Never inline hex.
• Component Library: Map to existing theme tokens. Never bypass the theme.
• CSS/SCSS: Add custom properties to :root. Never scatter magic values.

Step 5: Icons & Assets

Icons: Find a match in the project's existing icon library first — by name, then by visual shape. Match stroke-width and size exactly.

Fallback: If the layer name doesn't map, get_screenshot(iconNodeId) and identify visually. Match by shape, not name.

Images/illustrations: Download from the Figma MCP assets endpoint and save to the project's public/static folder. Don't create inline SVG blobs for complex illustrations. Don't import new icon packages without asking.

Property Checklist

Before writing code for ANY element, verify ALL applicable properties:

• Text: font-family, font-size, font-weight, line-height, letter-spacing, color, opacity, text-align, text-decoration, text-transform
• Container: width, height, min/max-width, padding (all 4), margin, background-color, border-radius, border-width/color/style, box-shadow, opacity, overflow
• Icon: size, fill, stroke, stroke-width, color (independent from parent text)
• Button/Link: All text + container props + cursor, hover/active/disabled states
• Image: width, height, object-fit, border-radius, border, aspect-ratio
• Spacing: gap, row-gap, column-gap

Layout Principle: Avoid hardcoded sizes. With correct font-size, line-height, padding, and parent width, elements render correctly. Hardcoded dimensions are a symptom of wrong upstream layout.

Responsive: Consider how the component adapts to different screen sizes. Use the project's responsive approach (Tailwind breakpoints, media queries, container queries).

Never use approximate Tailwind classes (like text-zinc-500) when exact hex values are available from tokens.

Phase 3: Refinement Loop

You think you're done. Now prove it. Keep cycling until the result is perfect.

The Loop

Step 1: Chrome screenshot (1 image)

First look at what you built. Compare against Figma screenshots in context. Be extremely picky:

• Visual alignment issues
• Missing elements or wrong proportions
• Layout shifts, spacing that looks off
• Wrong icon shapes
• Color or weight mismatches
• Color comparison: Sample dominant colors from Figma vs Chrome. Flag anything that looks off — opacity/layering can cause perceived differences even when hex values match.

Step 2: Numerical audit (zero images)

Run getComputedStyle on every element and compare against Figma values:

const el = document.querySelector('.target-element');
const s = getComputedStyle(el);
JSON.stringify({
  padding: s.padding, margin: s.margin, gap: s.gap,
  backgroundColor: s.backgroundColor, color: s.color,
  fontSize: s.fontSize, fontWeight: s.fontWeight,
  lineHeight: s.lineHeight, letterSpacing: s.letterSpacing,
  borderRadius: s.borderRadius, borderWidth: s.borderWidth,
  borderColor: s.borderColor, boxShadow: s.boxShadow
});

12px is not 10px. #f97316 is not #ff611c.

CSS specificity check: Icons inside wrapper components (Button, Link) often get wrong colors from parent overrides. Run getComputedStyle on the SVG itself. If the computed color doesn't match, use the Tailwind important modifier (prefix with !) to force it.

Step 3: Picky mismatch list

Combine visual + numerical issues into one list. No issue is too small. 2px off? List it.

Step 4: Fix everything

Batch all fixes. No screenshots between individual fixes.

Step 5: Repeat from Step 1

Exit condition: Retake get_screenshot on the Figma root node. Place side-by-side with your latest Chrome screenshot. Compare colors directly. Only move on when both screenshots match AND numerical audit shows zero mismatches.

Icon & Shape Verification

If an icon looks wrong during any loop iteration:

• get_screenshot(iconNodeId, fileKey) on the Figma icon
• Screenshot Chrome icon at 3x zoom
• Compare the SILHOUETTE — stroke count, shape, proportions

Common mismatches:

• "filter" → often lines-with-circles, NOT a funnel
• "calendar" → many variants (with/without dots)
• "mail" → open vs closed envelope

If it doesn't match: try another icon from the library, or use the Figma SVG inline.

Anti-Pattern: "Looks Good Enough"

This ALWAYS misses: 2-4px spacing differences, wrong icon variant, slightly wrong color, missing shadow, font weight mismatch.

RULE: Not done until zero visual mismatches AND zero numerical mismatches.

Recovery: Compact & Resume

If you hit the image limit (or proactively after ~15 screenshots):

1. Run /compact — summarizes conversation, drops old images. Notes and progress are preserved.
2. Retake 3 reference images: Figma overview, Figma detail (current section), Chrome state.
3. Continue the refinement loop — you still know everything from the compacted history. ~17 more image slots available.

Repeatable: compact → retake 3 → continue → compact again if needed.

Phase 4: User Review

Stop and ask:

• "Here's the final result. Are you happy with it?"
• "If something looks off, paste a Figma 'Link to Selection' for the specific area."

If the user provides a new link, re-run Phases 1-3 scoped to that selection.

Doctor Check

Run this ONLY when something fails. Not at startup.

Figma MCP not working?

• Call whoami to check authentication
• Not connected → alert user to configure Figma MCP
• Not authenticated → guide through Figma OAuth

Chrome not responding?

• Ensure Claude Chrome extension is active
• Navigate to the correct localhost page

Dev server not running?

• Check lockfile: package-lock.json (npm), yarn.lock (yarn), pnpm-lock.yaml (pnpm), bun.lockb (bun)
• Read package.json for dev command and port
• Default ports: 5173 (Vite), 3000 (Next/CRA), 8080 (Vue CLI)
• lsof -i :<PORT> — if not running, start in background

Unknown design system or icon library?

• Scan package.json: tailwindcss, styled-components, @emotion/*, @chakra-ui/*, @mui/*, @mantine/*
• Icon libraries: lucide-react, @heroicons/react, react-icons, @radix-ui/react-icons, @fortawesome/*, @phosphor-icons/react, @tabler/icons-react

Examples

Invocation:

/pix
> Paste Figma link: https://figma.com/design/abc123/MyApp?node-id=42-100

What happens:

1. Recon: get_metadata + get_variable_defs + get_code_connect_map (0 images)
2. Figma overview: get_screenshot on root (image 1)
3. Study section 1: get_screenshot (image 2) + get_design_context (text). Memorize.
4. Study section 2: get_screenshot (image 3) + get_design_context (text). Memorize.
5. Code everything from memory (0 images)
6. Loop round 1: Chrome screenshot (image 4) + audit → 4 mismatches + 1 bad icon
7. Fix mismatches. Icon check: get_screenshot on Figma icon (image 5) — swap it.
8. Loop round 2: Chrome screenshot (image 6) + audit → 1 spacing issue
9. Fix spacing.
10. Loop round 3: Chrome screenshot (image 7) + audit → zero mismatches.
11. Final check: Retake Figma get_screenshot (image 8) side-by-side with Chrome → match.
12. Done. Total: 8 images.

Bad patterns:

• Screenshotting every element individually (budget killer)
• Re-screenshotting the same Figma node (it's static)
• Chrome screenshots after every small CSS tweak
• Using screenshots to verify what getComputedStyle gives for free
• Not running /compact when approaching image limit
• get_design_context on root (token waste)
• get_variable_defs more than once (redundant)
• Hardcoded hex in Tailwind (text-[#f3f3f3])
• Magic width numbers (w-[247px])
• Saying "looks close" without numerical verification