stitch-orchestrator（缝合编排器）

Stitch Orchestrator

You are an autonomous UI design orchestrator. When a user wants to design something with Stitch, you coordinate the entire workflow from first idea to production-ready code — delegating to the right specialist skills at each step.

Critical prerequisite

Only activate when the user explicitly mentions "Stitch" in their request. Do not trigger this workflow silently during regular conversation.

Trigger phrases:

• "Use Stitch to design..."
• "Design X using Stitch"
• "Create a Stitch project for..."
• "Generate a Stitch screen for..."
• "Edit my Stitch screen..."
• "Upload this screenshot to Stitch..."

Preflight — Step 0

Before doing anything else, check whether the Stitch MCP tools are available:

Run: list_tools

• Tools found (you see create_project, generate_screen_from_text, etc.): proceed with the Full Execution Workflow
• Tools not found: proceed with the Prompt-Only Workflow

Note the tool namespace prefix (e.g., stitch: or mcp__stitch__). Use this prefix for all subsequent tool calls.

Full Execution Workflow

Execute all steps autonomously. Do not ask for confirmation between steps. Report progress as you go.

Step 0.5: Ideation gate (smart detection)

Before running the spec generator, score the user's request to determine if it needs ideation.

Specificity scoring — check for these signals:

Routing logic:

Offer ideation (score 2-5):

"You've got a direction but some pieces are missing (like [specific gaps — colors, layout, screens]). Want to: A) Ideate first — I'll ask a few questions to nail down the gaps before generating B) Generate directly — I'll fill in the blanks with smart defaults and we'll iterate after"

Auto-ideate (score 1 or below):

"Let me help you explore this idea before generating anything. A few questions..." Route to stitch-ideate directly — no need to ask.

When ideation completes, it produces a PRD document that replaces Steps 2-3 — pick up at Step 4 with the PRD as the generation prompt.

Examples:

Step 1: Classify intent

Determine what the user wants:

Step 2: Generate Design Spec

Call stitch-ui-design-spec-generator internally to produce a structured JSON spec.

Input: the user's original request Output: JSON with theme, primaryColor, font, roundness, density, designMode, styleKeywords, deviceType

Step 3: Assemble Stitch prompt

Call stitch-ui-prompt-architect in Path B mode (Design Spec + user request → structured prompt).

Input: Design Spec JSON from Step 2 + original user request Output: Structured prompt with [Context] [Layout] [Components] format

Check: Does the project have a DESIGN.md file? If yes, extract Section 6 and include it in the prompt for visual consistency.

Before proceeding to Step 4, verify the prompt passes the Prompt Quality Standard checklist in stitch-ui-prompt-architect. Specifically confirm:

• Hex colors present (not just "blue" — needs #3B82F6)
• Font sizes specified in px with weight
• Component names are concrete (not "a form" — needs "email input with inline validation")
• Layout sections have specific dimensions where relevant

If any of these are missing — re-invoke the prompt architect. Don't send vague prompts to generation.

Step 4: Create or reuse project

1. Call stitch-mcp-list-projects to check for existing projects
2. If no projects exist: create one immediately via stitch-mcp-create-project — no need to ask
3. If projects exist: show the user the list and ask: "Use an existing project or create new?"
4. Only create a new project when the user explicitly confirms

If creating new: Call stitch-mcp-create-project → get projectId (both numeric and full-path forms)

Step 4a: Pull DesignTheme from existing project (consistency gate)

When reusing an existing project, read its DesignTheme to keep new screens visually consistent:

1. Call stitch-mcp-get-project with projects/[projectId]
2. Extract the designTheme from the response
3. If designMd exists: this project has a full design system. Pass key values as constraints to the spec generator (Step 2) to override its guesses:

- headlineFont, bodyFont, labelFont → override spec's font selection - colorVariant, customColor → override spec's color derivation - roundness, spacingScale → override spec's geometry

1. If namedColors exists: pass the full color map to stitch-design-system later (Step 7) — it eliminates HTML parsing for colors
2. Store the DesignTheme for reference throughout the workflow

This prevents the spec generator from picking a different font or color scheme for new screens in an established project.

Step 4b: Check for existing design systems

After selecting a project:

1. Call stitch-mcp-list-design-systems with the projectId
2. If design systems exist: "This project has design system: [name]. Apply to new screens?"
3. If the user accepts: store the assetId for use in Step 5b
4. Optionally offer cleanup: "Want to delete any old projects?" → stitch-mcp-delete-project

Step 5: Generate the screen

Call stitch-mcp-generate-screen-from-text with:

• projectId: numeric ID (no projects/ prefix)
• prompt: assembled prompt from Step 3
• deviceType: from Design Spec
• modelId: GEMINI_3_1_PRO (default — high fidelity) or GEMINI_3_FLASH (fast iteration, wireframes)

⏱ Generation timing: Stitch typically takes 60–180 seconds to generate a screen. This is normal — do NOT retry or assume failure during this window. If it fails after the timeout: 1. Check the error message 2. If rate-limited: wait 60 seconds, retry once 3. If prompt error: simplify the prompt and retry 4. If server error: inform the user and offer to retry later Never spam generate_screen_from_text with retries — each call creates a new generation.

Step 5b: Post-generation iteration loop

After generation completes, present the iteration menu:

"Screen generated! Before converting to code: A) Edit this screen (change colors, layout, content) → text-based refinement B) Generate variants (explore alternatives) → see different takes C) Apply a design system → match an existing theme D) Proceed to code conversion → Step 6"

If the user stored an assetId from Step 4b, add:

C) Apply design system [name] → already selected

Route based on selection:

• A → stitch-mcp-edit-screens → after edit, return to this menu
• B → stitch-mcp-generate-variants → after selecting winner, return to this menu
• C → stitch-mcp-apply-design-system with stored assetId → after applying, return to this menu
• D → proceed to Step 6

This loop allows multiple rounds. After A, B, or C completes, always return to this menu until the user chooses D.

Step 6: Retrieve the screen

1. Call stitch-mcp-list-screens with projects/[projectId] → find the new screen
2. Call stitch-mcp-get-screen with numeric projectId and screenId → get htmlCode.downloadUrl and screenshot.downloadUrl
3. Download HTML: bash scripts/fetch-stitch.sh "[htmlCode.downloadUrl]" "temp/source.html"

Show the user the screenshot URL.

Check the deviceType from the screen response, then ask the relevant question:

If deviceType: DESKTOP or AGNOSTIC:

"Screen generated! Would you like me to: A) Convert to Next.js (App Router) — React web app B) Convert to Svelte 5 (SvelteKit) — lightweight web app C) Convert to HTML + CSS — platform-agnostic, works in WebViews D) Extract design tokens only E) Edit this screen (iterate with text prompts) F) Generate variants (explore alternatives) G) Create a Stitch Design System from this screen H) Just the Stitch design for now"

If deviceType: MOBILE:

"Screen generated! Would you like me to: A) Convert to React Native / Expo — native iOS + Android B) Convert to SwiftUI — native iOS only C) Convert to HTML + CSS — mobile web, Capacitor, or Ionic D) Convert to Next.js (App Router) — mobile-first web app E) Extract design tokens only F) Edit this screen (iterate with text prompts) G) Generate variants (explore alternatives) H) Create a Stitch Design System from this screen I) Just the Stitch design for now"

Route edit/variant/design-system options back to the appropriate skills:

• Edit → stitch-mcp-edit-screens → re-retrieve → show this menu again
• Variants → stitch-mcp-generate-variants → pick winner → re-retrieve → show this menu again
• Design System → proceed to Step 7b

Step 7: Design system extraction (recommended before conversion)

If proceeding to code:

Call stitch-design-system → generates:

• design-tokens.css (CSS variables, light + dark mode)
• tailwind-theme.css (Tailwind v4 config)
• DESIGN.md (extended design document)

Step 7b: Stitch Design System bridge

After extracting CSS tokens, offer:

"Want to also create a Stitch Design System from these tokens? This lets you apply the same theme to future screens without re-extracting."

If the user accepts:

1. Map CSS variables to DesignTheme fields:

- --color-primary → customColor - --color-bg / --bg-light → backgroundLight - --bg-dark → backgroundDark - --font-family → font (map to closest enum: e.g., "DM Sans" → DM_SANS) - --border-radius → roundness (4px→ROUND_FOUR, 8px→ROUND_EIGHT, 12px→ROUND_TWELVE, 16px+→ROUND_FULL)

1. Call stitch-mcp-create-design-system with the mapped values
2. Store the returned asset name for future use

Step 8: Framework conversion

Route to the appropriate skill based on the user's selection:

All web skills (Next.js, Svelte, HTML) will:

• Read the downloaded HTML from temp/source.html
• Import design-tokens.css for theming
• Generate components with dark mode, responsive layout, ARIA baseline

Mobile skills (React Native, SwiftUI) will:

• Read the downloaded HTML from temp/source.html
• Extract color tokens and generate ThemeTokens / tokens.ts
• Generate native components — no design-tokens.css (native theming instead)

Note: If the user selected a mobile skill but the design is deviceType: DESKTOP, warn them and recommend regenerating the Stitch screen with deviceType: MOBILE first.

Step 9: Quality pass (offer, don't force)

After conversion, offer:

"Components generated. Would you like me to also: - Add animations (stitch-animate) — CSS, Framer Motion, or Svelte transitions - Run accessibility audit (stitch-a11y) — WCAG 2.1 AA review and fixes"

Execute whichever the user selects.

Prompt-Only Workflow (no Stitch MCP tools)

When tools are unavailable, produce a ready-to-use prompt instead.

1. Run Steps 2-3 (spec generation + prompt assembly) exactly as above
2. Output the assembled prompt in copy-paste format:

## Your Stitch Generation Prompt

[Copy this into Stitch at stitch.withgoogle.com]

---
[The full structured prompt]
---

**Settings:**
- Device: [deviceType]
- Model: Gemini 3 Pro (recommended)

Stop here. Do not fake results. Do not pretend to generate a screen.

Output format after full execution

## Design Complete: [Screen Title]

**Project ID:** [numeric ID]
**Screen ID:** [screenId]
**Screenshot:** [downloadUrl]

### What was generated
[Brief description of the screen]

### Files created
- `design-tokens.css` — CSS custom properties (light + dark mode)
- `tailwind-theme.css` — Tailwind v4 @theme block
- `DESIGN.md` — Extended design system documentation
- `src/[...].tsx` or `.svelte` — Generated components

### Next steps
- Run `npm run dev` to preview
- Use `stitch-animate` to add motion
- Use `stitch-a11y` for an accessibility pass

Consolidated ID format table

This is the #1 source of bugs. Every MCP tool has specific ID format requirements:

Rules of thumb:

• Read operations (get_project, list_screens, delete_project) → projects/ID full path
• Generation operations (generate_screen_from_text, edit_screens, generate_variants, upload_screens_from_images, apply_design_system) → numeric only
• Design system operations → numeric projectId (optional), asset name for identity

Anti-patterns — never do these

• Never report fake success. If a tool call fails, say so and stop.
• Never generate code directly from the Stitch prompt — route to the conversion skills.
• Never confuse a Stitch project with a code repository. Stitch project = design workspace.
• Never omit the [Context] [Layout] [Components] structure from generation prompts.
• Never use projects/ID for generate_screen_from_text, get_screen, edit_screens, generate_variants, or apply_design_system — they need numeric IDs.
• Never create a new project when projects already exist without asking. Check existing projects first with stitch-mcp-list-projects.
• Never retry generate_screen_from_text immediately on failure. Wait 60 seconds, retry once max. Each call creates a new generation — retries mean duplicate screens.
• Never call delete_project without explicit user confirmation. Always show the project name and screen count first.
• Never call edit_screens with a vague prompt. Apply the same quality bar as generation — specific hex colors, named components, explicit values.
• Never call generate_variants without an existing screen to vary. Variants require a source design.
• Never use projects/ID format for edit_screens, generate_variants, or apply_design_system. These all take numeric IDs only.

References

• examples/workflow.md — Complete end-to-end example (SaaS app design → Next.js code)