Design to Code

Convert visual designs into production-ready React components using a four-stage pipeline: Extract, Match, Adapt, Render.

design-to-code screenshot of hero section    # From description
design-to-code /tmp/mockup.png               # From screenshot
design-to-code https://example.com/pricing    # From URL

Pipeline Overview

Input (screenshot/description/URL)
  │
  ▼
┌─────────────────────────┐
│ Stage 1: EXTRACT         │  Stitch MCP → screen details + URLs
│ generate_screen_from_text│  Create screens from a text prompt
│ get_screen               │  Screen details with download URLs
└─────────┬───────────────┘
          │
          ▼
┌─────────────────────────┐
│ Stage 2: MATCH           │  1. Storybook MCP → check existing
│ Storybook-first lookup   │  2. 21st.dev → search public registry
│ Then 21st.dev fallback   │  3. Filesystem → grep codebase
└─────────┬───────────────┘
          │
          ▼
┌─────────────────────────┐
│ Stage 3: ADAPT           │  Merge extracted design + matched
│ Apply project tokens     │  components into final implementation
│ Customize to codebase    │  Tests + types included
└─────────┬───────────────┘
          │
          ▼
┌─────────────────────────┐
│ Stage 4: RENDER          │  Register as json-render catalog entry
│ Generate Zod schema      │  Same component → PDF, email, video
│ Add to defineCatalog()   │  Multi-surface reuse via MCP output
└─────────┬───────────────┘
          │
          ▼
┌─────────────────────────┐
│ Stage 4b: VERIFY         │  Storybook MCP → self-healing loop
│ run-story-tests(a11y)    │  Fix violations, retry (max 3)
│ preview-stories          │  Embed live preview in chat
└─────────────────────────┘

Argument Resolution

INPUT = ""  # Full argument string
# Detect input type:
# - Starts with "/" or "~" or contains ".png"/".jpg" → screenshot file path
# - Starts with "http" → URL to screenshot or live page
# - Otherwise → natural language description

Step 0: Detect Input Type and Project Context

# 1. Create main task IMMEDIATELY
TaskCreate(subject="Design to code: {INPUT}", description="Four-stage pipeline: extract, match, adapt, render", activeForm="Converting design to code")

# 2. Create subtasks for each stage
TaskCreate(subject="Extract design context", activeForm="Extracting design context")               # id=2
TaskCreate(subject="Match components (Storybook-first)", activeForm="Matching components")         # id=3
TaskCreate(subject="Adapt to project tokens and conventions", activeForm="Adapting to project")    # id=4
TaskCreate(subject="Register in json-render catalog", activeForm="Registering in catalog")         # id=5
TaskCreate(subject="Verify with Storybook self-healing", activeForm="Verifying component")         # id=6

# 3. Set dependencies for sequential stages
TaskUpdate(taskId="3", addBlockedBy=["2"])  # Match needs extracted design context
TaskUpdate(taskId="4", addBlockedBy=["3"])  # Adapt needs matched components
TaskUpdate(taskId="5", addBlockedBy=["4"])  # Render needs adapted component
TaskUpdate(taskId="6", addBlockedBy=["5"])  # Verify needs rendered component

# 4. Update status as you progress
TaskUpdate(taskId="2", status="in_progress")  # When starting
TaskUpdate(taskId="2", status="completed")    # When done — repeat for each subtask

# Detect project's design system
Grep("@theme", glob="**/*.css")   # Tailwind v4: theme lives in CSS, not a config file
Glob("**/tailwind.config.*")      # Tailwind v3 only (v4 ignores this file)
Glob("**/tokens.css")
Glob("**/.tokens.json")
# Read existing tokens if found → used in Stage 3

# Detect shadcn/ui style (v4 style system)
Glob("**/components.json")
# Read → style field (e.g., "radix-luma", "base-nova")
# Determines class names: Luma=rounded-4xl, Nova=compact, Lyra=sharp
# Store: SHADCN_STYLE for Stage 2 filtering + Stage 3 adaptation

Stage 1: Extract Design Context

If stitch MCP is available:

# Live Stitch MCP tools (https://stitch.googleapis.com/mcp), matching
# design-context-extractor / design-system-architect agent grants:
#   - list_projects() / get_project(projectId)
#   - list_screens(projectId)
#   - get_screen(name="projects/{projectId}/screens/{screenId}")
#       ONE required param: name. Returns screen details including
#       download URLs (fetch those URLs; no separate code/image export tool)
#   - generate_screen_from_text(projectId, prompt)
#       quota-costing write; can take a few minutes. DO NOT RETRY.
#       On timeout: poll get_screen(name=...) every 30 s, up to 10 times.
#
# For screenshot/URL input:
#   1. list_projects / get_project to locate the Stitch project
#   2. generate_screen_from_text with the extraction goal (or use an
#      existing screen if the project already has one). Never retry.
#   3. list_screens, then get_screen(name="projects/{projectId}/screens/{screenId}")
#   4. Fetch the download URLs (or Read multimodal) for token extraction
#
# For description input:
#   1. generate_screen_from_text(projectId, prompt=<description>) once
#   2. On timeout, poll get_screen(name=...) every 30 s up to 10 times
#   3. Otherwise get_screen(name="projects/{projectId}/screens/{screenId}")
#
# DESIGN.md import (Stitch Pro, Mar 2026+):
#   Prefer upload_design_md / create_design_system_from_design_md when
#   those tools are available on the live server; otherwise regenerate
#   via generate_screen_from_text from the DESIGN.md contents (once, no retry).

If stitch MCP is NOT available (fallback):

# For screenshot: Read the image file directly (Claude is multimodal)
# Analyze layout, colors, typography, spacing from the image
# For URL: WebFetch the page, extract HTML structure
# For description: Skip extraction, proceed to Stage 2 with description

Resolution budget (Opus 5.5 / CC 2.1.111+): Mockups up to 2,576 px on the long edge (~3.75 MP, 3× prior ceiling) produce better component boundaries and spacing extraction. Full-page desktop mockups at native resolution are now in-budget; previously they had to be resized down and lost fine detail. Only downscale inputs exceeding 2,576 px.

Extract and produce:

• Color palette (hex/oklch values)
• Typography (font families, sizes, weights)
• Spacing patterns (padding, margins, gaps)
• Component structure (headers, cards, buttons, etc.)
• Layout pattern (grid, flex, sidebar, etc.)

Stage 2: Match Components (Storybook-First)

Priority 1 — Check project's own Storybook (if storybook-mcp available):

# Search the project's existing component library first
inventory = list-all-documentation()  # Full component + docs manifest
for component in inventory.components:
    if component matches extracted_description:
        details = get-documentation(id=component.id)
        # Returns: props schema, stories, usage patterns
        # → Existing component found — skip external search

Priority 2 — Search 21st.dev (if no Storybook match and 21st-dev-magic available):

# Search 21st.dev for matching components
# Use the component descriptions from Stage 1
# Example: "animated pricing table with toggle"
# Filter: React, Tailwind CSS, shadcn/ui compatible
# If SHADCN_STYLE detected, prefer components matching style's visual language
# (e.g., Luma → rounded/pill-shaped, Nova → compact/dense, Lyra → sharp/boxy)

Priority 3 — Filesystem fallback (if no MCP servers available):

# Search for components in the project's existing codebase
Grep(pattern="export.*function|export.*const", glob="**/*.tsx")
# Check for shadcn/ui components
Glob("**/components/ui/*.tsx")
# Generate from scratch if no matches found

Present matches to user:

AskUserQuestion(questions=[{
  "question": "Which component approach for {component_name}?",
  "header": "Component",
  "options": [
    {"label": "Reuse from Storybook", "description": "{existing_component} — props: {prop_list}"},
    {"label": "Use 21st.dev match", "description": "{matched_component_name} — {match_score}% match"},
    {"label": "Adapt from codebase", "description": "Modify existing {existing_component}"},
    {"label": "Generate from scratch", "description": "Build new component from extracted design"}
  ],
  "multiSelect": false
}])

Stage 3: Adapt to Project

Merge the extracted design context with matched/generated components:

1. Apply project tokens — Replace hardcoded colors/spacing with project's design tokens
2. Apply shadcn style classes — If SHADCN_STYLE detected, use style-correct class names:
   • Luma: rounded-4xl buttons/cards, shadow-md + ring-1 ring-foreground/5, gap-6 py-6
   • Nova: compact px-2 py-1, reduced margins, tight spacing
   • Lyra: rounded-none, sharp edges, monospace-friendly
   • Mira: ultra-dense, minimal padding
3. Match naming conventions — Follow project's component naming patterns
4. Add TypeScript types — Full type safety with Zod validation for any data props
5. Include tests — MSW handlers for API-backed components, render tests for static
6. Responsive — Mobile-first with breakpoints matching project's system

Output Structure

src/components/
  └── {ComponentName}/
      ├── {ComponentName}.tsx       # Main component
      ├── {ComponentName}.test.tsx  # Tests
      └── index.ts                 # Barrel export

Stage 4: Register in json-render Catalog

After ADAPT produces a working React component, register it as a json-render catalog entry for multi-surface reuse.

1. Generate Zod schema — Derive a Zod schema from the component's TypeScript props
2. Add catalog entry — Register in the project's defineCatalog() call with props schema and children declaration
3. Verify rendering — Confirm the component renders correctly through <Render catalog={catalog} /> path

import { z } from 'zod'

// Zod schema derived from {ComponentName}Props
const componentSchema = z.object({
  title: z.string().max(100),
  variant: z.enum(['default', 'featured']).default('default'),
  // ... props from the adapted component
})

// Add to project catalog
import { defineCatalog } from '@json-render/core'
import { existingCatalog } from './catalog'

export const catalog = {
  ...existingCatalog,
  {ComponentName}: {
    props: componentSchema,
    children: true, // or false for leaf components
  },
}

Enables: same component output to PDF, email, video, or MCP response — no reimplementation needed.

Skip condition: If the project has no json-render dependency or catalog, inform the user and skip catalog registration. The component from Stage 3 is still fully usable standalone.

Stage 4b: Self-Healing Verification (if storybook-mcp available)

After generating the component, verify it with Storybook MCP:

# 1. Write CSF3 story for the new component
Write("src/components/{Name}/{Name}.stories.tsx", story_code)

# 2. Run tests via MCP (component + a11y)
results = run-story-tests(
    stories=[{ "storyId": "{name}--default" }],
    a11y=True
)

# 3. Handle failures — self-heal up to 3 attempts
if not results.all_passed:
    for failure in results.failures:
        # Read violation details, fix the component
        Edit("src/components/{Name}/{Name}.tsx", fix)
    # Re-run failing tests
    results = run-story-tests(stories=[...], a11y=True)

# 4. Preview — embed live story in chat for user confirmation
previews = preview-stories(stories=[
    { "absoluteStoryPath": "src/components/{Name}/{Name}.stories.tsx",
      "exportName": "Default" }
])
# Include preview URLs in response for visual confirmation

Skip condition: If storybook-mcp is not available, skip verification. The component is still usable — just not auto-verified.

Graceful Degradation

The skill ALWAYS produces output regardless of MCP availability. Storybook MCP adds component reuse (Stage 2) and self-healing verification (Stage 4b) — skipping them still yields a working component.

Anti-Patterns

• NEVER output components with hardcoded colors — use design tokens
• NEVER skip TypeScript types — all props must be typed
• NEVER generate without checking existing project patterns first
• NEVER ignore the project's existing component library structure

Quality Bar

Done means all of these hold:

• Generated components carry no hardcoded colors or literal spacing — every visual value resolves through project design tokens.
• Every prop is TypeScript-typed; data-bearing props carry Zod validation.
• Stage 2 checked existing components (Storybook, then 21st.dev, then filesystem) before generating from scratch, and the chosen match source is named.
• When storybook-mcp is available, the component passes run-story-tests (component + a11y) before handoff, self-healing within at most 3 attempts.
• Output follows the {ComponentName}/ structure — component, test, and barrel export.
• Each skipped stage (json-render catalog, Storybook verify) is taken ONLY when its dependency is absent, and the skip is stated with its reason.

Related Skills

• storybook-mcp-integration — Storybook MCP tools: component discovery, testing, previews
• component-search — Search 21st.dev registry standalone
• design-context-extract — Extract design DNA from screenshots
• design-stylecards — named aesthetic recipes (shadow stacks, glass surfaces, gradients, type scales) for the polish pass on generated components
• design-system-tokens — Token architecture and management
• storybook-testing — CSF3 patterns, Vitest integration, Chromatic TurboSnap
• json-render-catalog — Catalog definition, Zod schemas, defineCatalog patterns
• multi-surface-render — Render catalog entries to PDF, email, video, MCP