Stitchflow — AI UI Design Automation

Automates the full UI design pipeline: project understanding → prompt engineering → Stitch generation → result capture. Connects to your already-logged-in Chrome browser via CDP, so no API keys needed.

Core principle: Understand the project first, then write the prompt. No generic one-size-fits-all prompts.

Prerequisites (User Must Do)

⚠ If the browser isn't logged into Google or hasn't visited Stitch, the script will fail. Tell the user: "Please log into your Google account in Chrome, then open stitch.withgoogle.com once to authorize."

Full Workflow (AI Execution Steps)

Phase 1: Understand the Project

Before touching Stitch, read the project's context files to understand what this project actually is:

Priority order:
1. CLAUDE.md                    → Brand identity, project architecture, behavior rules
2. Product data / knowledge base → SKU data, feature specs
3. Platform configs             → Deployment platforms (e.g., Shopify, JD, Douyin)
4. Brand assets                 → Logos, colors, design tokens (if available)
5. Current UI code              → Existing dashboard/index.html (if redesign)

Extract these dimensions for the prompt:

• Brand name / Project codename
• Industry / Vertical (e-commerce? SaaS? hardware?)
• Brand colors / Design tokens (hex values if available)
• Target user persona
• Core functional needs (KPI dashboard? Content management? Charts?)
• Current UI pain points (if redesign)

Phase 2: Craft a Tailored Prompt

Write the prompt based on actual project data. Never use a generic template — every project's prompt should differ significantly because domains differ.

Prompt structure (must cover all dimensions):

Project Identity: [Brand/Codename], [Industry/Vertical], [one-line positioning]
Design Goal: [New / Redesign / Exploration], what page to design
Brand Tone: [3-5 adjectives], key visual elements
Color Scheme: Primary #XXXXXX, Accent #XXXXXX, [Dark/Light/Mixed]
Target User: [Who uses it], [Usage scenario]
Page Structure:
  - Screen 1: [Name] — [Function]
  - Screen 2: [Name] — [Function]
  - ...
Functional Requirements: [Specific metrics, chart types, interactions]
Language: All UI text in [Chinese/English/Japanese/etc.]
Style Reference: [Optional, 1-2 design directions]

Example — E-commerce Operations Dashboard:

Project Identity: CatCaptain (Lingyun), storage hardware e-commerce, high-value Micro SD / SSD brand
Design Goal: Redesign the AI operations dashboard for daily KPI monitoring + content management
Brand Tone: Professional, tech-driven, data-first, dark theme with vibrant orange accents
Color Scheme: Dark theme, Primary #1a1f36 (deep navy), Accent #FF6B35 (brand orange), Data Green #00C897
Target User: AI operations agent (daily monitoring) + Brand founder (weekly reports)
Page Structure:
  - Screen 1 — Main Dashboard: KPI cards (GMV/Orders/ROI/CTR) + trend charts + platform breakdown
  - Screen 2 — Content Factory: Video asset list + A/B test status + content calendar
  - Screen 3 — Competitor Watch: Price comparison table + competitor activity timeline
  - Screen 4 — Auto Reports: Daily/weekly report preview + key metric change highlights
Functional Requirements: Real-time data cards + Chart.js line/bar charts + filterable tables + left sidebar nav + dark mode
Language: All UI text in Chinese
Style Reference: Vercel Analytics data cards + Linear dark sidebar

Phase 3: Launch Chrome CDP

Call stitch.py --launch-chrome to start Chrome in CDP mode.

macOS:

killall "Google Chrome" 2>/dev/null; sleep 2
cp -r "$HOME/Library/Application Support/Google/Chrome" "/tmp/chrome-profile-clone"
nohup /Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
  --remote-debugging-port=9222 \
  --user-data-dir="/tmp/chrome-profile-clone" \
  > /tmp/chrome-cdp.log 2>&1 &

Windows:

taskkill /F /IM chrome.exe /T
xcopy "%LOCALAPPDATA%\Google\Chrome\User Data" "%TEMP%\chrome-profile-clone" /E /I /H /Y
start "" "C:\Program Files\Google\Chrome\Application\chrome.exe" ^
  --remote-debugging-port=9222 ^
  --user-data-dir="%TEMP%\chrome-profile-clone" ^
  --disable-features=DevToolsDebuggingRestrictions

Windows note: Chrome 136+ (March 2025) silently ignores --remote-debugging-port on default profile paths. The --disable-features=DevToolsDebuggingRestrictions flag is required. macOS/Linux naturally bypass this by using /tmp cloned profiles.

Linux:

pkill -9 chrome 2>/dev/null; sleep 2
cp -r ~/.config/google-chrome /tmp/chrome-profile-clone
nohup google-chrome \
  --remote-debugging-port=9222 \
  --user-data-dir="/tmp/chrome-profile-clone" \
  > /tmp/chrome-cdp.log 2>&1 &

Phase 4: Run Stitch Design

python3 ~/.claude/skills/stitchflow/stitch.py "<prompt from Phase 2>" --output <project-dir>/stitch-design-v1.png

The script automatically:

1. Connects to Chrome via CDP
2. Opens the Stitch home page (stitch.withgoogle.com)
3. Enters the Stitch iframe (app-companion-430619.appspot.com)
4. Selects the "Web" platform on the home page (platform selector only visible on home page, NOT on project pages)
5. Types the prompt into the contenteditable editor using keyboard input
6. Presses Enter to create a new project (Stitch auto-starts generation — no need to click a generate button)
7. Polls until generation completes (checks for preview frame or completion cues)
8. Saves a full-page screenshot

Phase 5: Model Selection

Stitch defaults to a standard model. Always switch to the most capable model available in the Stitch UI — this changes over time as Google releases new models. At time of writing (May 2026), the top model is "Gemini 3.1 Pro", but always check the model selector for the newest option.

The rule: pick the model with the highest version number or the one labeled "Pro" / "Ultra" / "Max" — whichever is the most powerful in the dropdown list.

• Script now automates model selection: opens the model dropdown, scores available models, and picks the one with the highest version number + "Pro"/"Thinking" label
• Selects the top-tier model (more parameters → deeper design reasoning → more nuanced, creative output)
• Trade-off: longer generation time (~60-120s vs 30-60s for standard models)

Phase 6: Review, Export + Handoff

Use the Read tool to show the screenshot to the user. Once direction is confirmed:

Approved → Export design to code:

python3 ~/.claude/skills/stitchflow/stitch.py \
  "<same prompt>" \
  --export .stitch/designs/

The --export flag triggers stitch_export() which uses 4 fallback strategies:

1. Click Stitch's "Download" / "Export" button in the UI
2. Try the "..." more menu → find export options
3. Extract rendered HTML directly from the Stitch preview iframe
4. Capture browser download events for any triggered file downloads

Files are saved to .stitch/designs/ (Google stitch-skills compatible directory):

.stitch/designs/
├── home-dashboard.html     # Exported HTML
├── home-dashboard.png      # Design screenshot
└── index.html              # Full page HTML (fallback extraction)

AI then converts to project code:

• Read the exported HTML/CSS files from .stitch/designs/
• Map the design to the project's actual framework (React/Vue/static HTML)
• Match design tokens to the project's CSS variables / Tailwind config
• Write clean, production-ready components
• Respect the project's existing file structure and conventions

**Needs revision → Refine prompt based on feedba