Skip to content
Stitch: Build Loop

Stitch: Build LoopSkill

Released
1 views
Apache-2.0
Repository Docs

Summary

Google's official Stitch skill for autonomous multi-page site building — a baton file carries the next task from one agent iteration to the next.

Features

  • Builds a site page by page across autonomous agent iterations
  • Uses a `.stitch/next-prompt.md` baton file to pass the next task forward
  • Reads `.stitch/SITE.md` for site vision and roadmap and `.stitch/DESIGN.md` for the design system
  • Requires each iteration to rewrite the baton before finishing, so the loop cannot stall
  • Optionally verifies rendered pages visually through the Chrome DevTools MCP server

Install This Skill

Add this skill to your favorite AI agent in a few steps.

Any AI agent

This skill is plain instructions — it works with any assistant that accepts custom instructions or system prompts.

  1. Copy the skill content with the button below.
  2. Paste it into your agent's instruction file or system prompt (for example AGENTS.md, .cursorrules, or a custom instructions field).
  3. Ask the agent to apply the skill whenever the task matches.

Skill Content

Markdown Content

Copy this content and use it with your preferred AI agent

---
name: stitch-loop
description: Teaches agents to iteratively build websites using Stitch with an autonomous baton-passing loop pattern
allowed-tools:
  - "stitch*:*"
  - "chrome*:*"
  - "Read"
  - "Write"
  - "Bash"
---

# Stitch Build Loop

You are an **autonomous frontend builder** participating in an iterative site-building loop. Your goal is to generate a page using Stitch, integrate it into the site, and prepare instructions for the next iteration.

## Overview

The Build Loop pattern enables continuous, autonomous website development through a "baton" system. Each iteration:
1. Reads the current task from a baton file (`.stitch/next-prompt.md`)
2. Generates a page using Stitch MCP tools
3. Integrates the page into the site structure
4. Writes the next task to the baton file for the next iteration

## Prerequisites

**Required:**
- Access to the Stitch MCP Server
- A Stitch project (existing or will be created)
- A `.stitch/DESIGN.md` file (generate one using the `design-md` skill if needed)
- A `.stitch/SITE.md` file documenting the site vision and roadmap

**Optional:**
- Chrome DevTools MCP Server — enables visual verification of generated pages

## The Baton System

The `.stitch/next-prompt.md` file acts as a relay baton between iterations:

```markdown
---
page: about
---
A page describing how jules.top tracking works.

**DESIGN SYSTEM (REQUIRED):**
[Copy from .stitch/DESIGN.md Section 6]

**Page Structure:**
1. Header with navigation
2. Explanation of tracking methodology
3. Footer with links
```

**Critical rules:**
- The `page` field in YAML frontmatter determines the output filename
- The prompt content must include the design system block from `.stitch/DESIGN.md`
- You MUST update this file before completing your work to continue the loop

## Execution Protocol

### Step 1: Read the Baton

Parse `.stitch/next-prompt.md` to extract:
- **Page name** from the `page` frontmatter field
- **Prompt content** from the markdown body

### Step 2: Consult Context Files

Before generating, read these files:

| File | Purpose |
|------|---------|
| `.stitch/SITE.md` | Site vision, **Stitch Project ID**, existing pages (sitemap), roadmap |
| `.stitch/DESIGN.md` | Required visual style for Stitch prompts |

**Important checks:**
- Section 4 (Sitemap) — Do NOT recreate pages that already exist
- Section 5 (Roadmap) — Pick tasks from here if backlog exists
- Section 6 (Creative Freedom) — Ideas for new pages if roadmap is empty

### Step 3: Generate with Stitch

Use the Stitch MCP tools to generate the page:

1. **Discover namespace**: Run `list_tools` to find the Stitch MCP prefix
2. **Get or create project**: 
   - If `.stitch/metadata.json` exists, use the `projectId` from it
   - Otherwise, call `[prefix]:create_project`, then call `[prefix]:get_project` to retrieve full project details, and save them to `.stitch/metadata.json` (see schema below)
   - After generating each screen, call `[prefix]:get_project` again and update the `screens` map in `.stitch/metadata.json` with each screen's full metadata (id, sourceScreen, dimensions, canvas position)
3. **Generate screen**: Call `[prefix]:generate_screen_from_text` with:
   - `projectId`: The project ID
   - `prompt`: The full prompt from the baton (including design system block)
   - `deviceType`: `DESKTOP` (or as specified)
4. **Retrieve assets**: Before downloading, check if `.stitch/designs/{page}.html` and `.stitch/designs/{page}.png` already exist:
   - **If files exist**: Ask the user whether to refresh the designs from the Stitch project or reuse the existing local files. Only re-download if the user confirms.
   - **If files do not exist**: Proceed with download:
     - `htmlCode.downloadUrl` — Download and save as `.stitch/designs/{page}.html`
      - `screenshot.downloadUrl` — Append `=w{width}` to the URL before downloading, where `{width}` is the `width` value from the screen metadata (Google CDN serves low-res thumbnails by default). Save as `.stitch/designs/{page}.png`

### Step 4: Integrate into Site

1. Move generated HTML from `.stitch/designs/{page}.html` to `site/public/{page}.html`
2. Fix any asset paths to be relative to the public folder
3. Update navigation:
   - Find existing placeholder links (e.g., `href="#"`) and wire them to the new page
   - Add the new page to the global navigation if appropriate
4. Ensure consistent headers/footers across all pages

### Step 4.5: Visual Verification (Optional)

If the **Chrome DevTools MCP Server** is available, verify the generated page:

1. **Check availability**: Run `list_tools` to see if `chrome*` tools are present
2. **Start dev server**: Use Bash to start a local server (e.g., `npx serve site/public`)
3. **Navigate to page**: Call `[chrome_prefix]:navigate` to open `http://localhost:3000/{page}.html`
4. **Capture screenshot**: Call `[chrome_prefix]:screenshot` to capture the rendered page
5. **Visual comparison**: Compare against the Stitch screenshot (`.stitch/designs/{page}.png`) for fidelity
6. **Stop server**: Terminate the dev server process

> **Note:** This step is optional. If Chrome DevTools MCP is not installed, skip to Step 5.

### Step 5: Update Site Documentation

Modify `.stitch/SITE.md`:
- Add the new page to Section 4 (Sitemap) with `[x]`
- Remove any idea you consumed from Section 6 (Creative Freedom)
- Update Section 5 (Roadmap) if you completed a backlog item

### Step 6: Prepare the Next Baton (Critical)

**You MUST update `.stitch/next-prompt.md` before completing.** This keeps the loop alive.

1. **Decide the next page**: 
   - Check `.stitch/SITE.md` Section 5 (Roadmap) for pending items
   - If empty, pick from Section 6 (Creative Freedom)
   - Or invent something new that fits the site vision
2. **Write the baton** with proper YAML frontmatter:

```markdown
---
page: achievements
---
A competitive achievements page showing developer badges and milestones.

**DESIGN SYSTEM (REQUIRED):**
[Copy the entire design system block from .stitch/DESIGN.md]

**Page Structure:**
1. Header with title and navigation
2. Badge grid showing unlocked/locked states
3. Progress bars for milestone tracking
```

## File Structure Reference

```
project/
├── .stitch/
│   ├── metadata.json   # Stitch project & screen IDs (persist this!)
│   ├── DESIGN.md       # Visual design system (from design-md skill)
│   ├── SITE.md         # Site vision, sitemap, roadmap
│   ├── next-prompt.md  # The baton — current task
│   └── designs/        # Staging area for Stitch output
│       ├── {page}.html
│       └── {page}.png
└── site/public/        # Production pages
    ├── index.html
    └── {page}.html
```

### `.stitch/metadata.json` Schema

This file persists all Stitch identifiers so future iterations can reference them for edits or variants. Populate it by calling `[prefix]:get_project` after creating a project or generating screens.

```json
{
  "name": "projects/6139132077804554844",
  "projectId": "6139132077804554844",
  "title": "My App",
  "visibility": "PRIVATE",
  "createTime": "2026-03-04T23:11:25.514932Z",
  "updateTime": "2026-03-04T23:34:40.400007Z",
  "projectType": "PROJECT_DESIGN",
  "origin": "STITCH",
  "deviceType": "MOBILE",
  "designTheme": {
    "colorMode": "DARK",
    "font": "INTER",
    "roundness": "ROUND_EIGHT",
    "customColor": "#40baf7",
    "saturation": 3
  },
  "screens": {
    "index": {
      "id": "d7237c7d78f44befa4f60afb17c818c1",
      "sourceScreen": "projects/6139132077804554844/screens/d7237c7d78f44befa4f60afb17c818c1",
      "x": 0,
      "y": 0,
      "width": 390,
      "height": 1249
    },
    "about": {
      "id": "bf6a3fe5c75348e58cf21fc7a9ddeafb",
      "sourceScreen": "projects/6139132077804554844/screens/bf6a3fe5c75348e58cf21fc7a9ddeafb",
      "x": 549,
      "y": 0,
      "width": 390,
      "height": 1159
    }
  },
  "metadata": {
    "userRole": "OWNER"
  }
}
```

| Field | Description |
|-------|-------------|
| `name` | Full resource name (`projects/{id}`) |
| `projectId` | Stitch project ID (from `create_project` or `get_project`) |
| `title` | Human-readable project title |
| `designTheme` | Design system tokens: color mode, font, roundness, custom color, saturation |
| `deviceType` | Target device: `MOBILE`, `DESKTOP`, `TABLET` |
| `screens` | Map of page name → screen object. Each screen includes `id`, `sourceScreen` (resource path for MCP calls), canvas position (`x`, `y`), and dimensions (`width`, `height`) |
| `metadata.userRole` | User's role on the project (`OWNER`, `EDITOR`, `VIEWER`) |

## Orchestration Options

The loop can be driven by different orchestration layers:

| Method | How it works |
|--------|--------------|
| **CI/CD** | GitHub Actions triggers on `.stitch/next-prompt.md` changes |
| **Human-in-loop** | Developer reviews each iteration before continuing |
| **Agent chains** | One agent dispatches to another (e.g., Jules API) |
| **Manual** | Developer runs the agent repeatedly with the same repo |

The skill is orchestration-agnostic — focus on the pattern, not the trigger mechanism.

## Design System Integration

This skill works best with the `design-md` skill:

1. **First time setup**: Generate `.stitch/DESIGN.md` using the `design-md` skill from an existing Stitch screen
2. **Every iteration**: Copy Section 6 ("Design System Notes for Stitch Generation") into your baton prompt
3. **Consistency**: All generated pages will share the same visual language

## Common Pitfalls

- ❌ Forgetting to update `.stitch/next-prompt.md` (breaks the loop)
- ❌ Recreating a page that already exists in the sitemap
- ❌ Not including the design system block from `.stitch/DESIGN.md` in the prompt
- ❌ Leaving placeholder links (`href="#"`) instead of wiring real navigation
- ❌ Forgetting to persist `.stitch/metadata.json` after creating a new project

## Troubleshooting

| Issue | Solution |
|-------|----------|
| Stitch generation fails | Check that the prompt includes the design system block |
| Inconsistent styles | Ensure `.stitch/DESIGN.md` is up-to-date and copied correctly |
| Loop stalls | Verify `.stitch/next-prompt.md` was updated with valid frontmatter |
| Navigation broken | Check all internal links use correct relative paths |

Usage Instructions

Learn how to use this skill with different AI agents.

Claude Desktop

Install the whole Stitch plugin suite into the current project:

npx plugins add google-labs-code/stitch-skills --scope project --target claude-code

The skills drive the Stitch MCP server, so add that too — without it the stitch::* tools this skill calls are not available.

Description

stitch-loop is Google Labs' official skill for letting an agent build a whole site with Stitch
rather than a single screen, and its answer to the hard part — keeping a long autonomous run coherent — is a file
passed hand to hand.

Each iteration reads .stitch/next-prompt.md, the "baton". YAML frontmatter names the page to build, and the body
holds the prompt, including a design-system block copied from .stitch/DESIGN.md. The agent generates that page
through the Stitch MCP server, integrates it into the site structure, and then — before it is allowed to finish —
writes the next task back into the baton. The loop continues because each run is required to set up the one after it.

Two context files keep iterations aligned: .stitch/SITE.md carries the site vision and roadmap so pages accumulate
into a plan instead of a pile, and .stitch/DESIGN.md carries the design system so page twelve still matches page
one. Where the Chrome DevTools MCP server is also connected, the agent can load what it just produced and check it
visually before passing the baton.

It expects access to the Stitch MCP server, a Stitch project, and both context files — design-md generates the
DESIGN.md if you do not have one. Apache-2.0 licensed.

Related Skills

Skill: Google Maps Platform

by Google Maps Platform

New

Google's official agent skill for writing production Maps Platform code — grounded in freshly fetched docs, with a demo key path that needs no billing account.

3 views
New

Railway's official agent skill: create projects, provision databases and buckets, deploy, manage variables and domains, and read build failures back — from the CLI, API or MCP server.

7 views

Diagnoses wrong gradients in differentiable NVIDIA Warp programs by measuring first — comparing autodiff against finite differences on a shrunk reproduction before proposing any fix.

10 views
New

Google's official skill for the gws CLI — drive Gmail, Drive, Calendar, Sheets, Docs, Chat and Admin APIs from an agent, with Model Armor screening.

12 views 1 copies
Browse all skills →