stitch-upload

name	stitch-upload
description	Upload a wireframe, screenshot, or sketch to Google Stitch for digitization into a clean UI. Use when user provides an image file and asks to convert it to UI or digitize a sketch.

Upload to Stitch

Overview

Upload an image (wireframe, screenshot, sketch) to Google Stitch to create an editable UI screen from it. The uploaded image serves as a starting point that Stitch interprets and converts into clean HTML with Tailwind CSS.

Prerequisites

• Stitch MCP tools must be available
• User must provide an image file (PNG, JPG, etc.)

Workflow

Step 1: Prepare Image

Identify the image the user wants to upload:

• If the user attached a file → use that path directly
• If the user described an image → ask them to provide the file path

Save a copy to ./stitch-design/references/ for tracking.

Step 2: Get Context (Optional)

Ask the user:

• "What style should the final UI be? (clean, Material Design, minimalist...)"
• "Any specific changes from the original? (different colors, add elements...)"
• "Device type?" (DESKTOP / MOBILE / TABLET)

Step 3: Project Setup & Upload (Fire-and-Poll Protocol)

Project management:

1. Read ./stitch-design/context-map.md — check for active project
2. If active project exists → reuse it, never create a new one
3. If no project → create_project, IMMEDIATELY write ID to context-map.md before upload

Tell the user: "Uploading and processing image (~1 min)..."

upload_screens_from_images(projectId: "...", images: [{ filePath: "<path>", mimeType: "image/png" }])

If the call succeeds → proceed to Step 4.

If ECONNRESET / timeout / fetch failed:

1. Tell user: "Connection reset — checking if upload completed server-side..."
2. Wait 90 seconds
3. Call list_screens(projectId) — check for new screens
4. If new screen found → proceed to Step 4
5. If not found → wait 60 more seconds, poll again
6. If found → Step 4
7. If not found → "Upload didn't complete. Retry? (attempt X/2)"
8. Max 2 retries. Always use the same projectId.

Step 4: Refine (Optional)

If the user wants modifications from the original, use edit with the same fire-and-poll protocol:

edit_screens(projectId: "...", screenIds: ["<screenId>"], prompt: "<refinement prompt>")

Step 5: Download & Display

Use human-readable paths:

node ${CLAUDE_PLUGIN_ROOT}/scripts/download.mjs "<html-url>" "./stitch-output/{project-slug}/{screen-slug}/index.html"
node ${CLAUDE_PLUGIN_ROOT}/scripts/download.mjs "<image-url>" "./stitch-output/{project-slug}/{screen-slug}/preview.png"

Inline preview: Read the PNG file via Read tool — displays the image inline in chat.

Open in browser:

open ./stitch-output/{project-slug}/{screen-slug}/index.html

Step 6: Update Tracking (MANDATORY — do not skip)

1. context-map.md — add screen entry with reference to uploaded image (upsert by screen ID)
2. prompts.md — log the upload context and any refinement prompts
3. usage.json — increment upload counter for today

Step 7: Offer Next Steps

• "Edit the design further?"
• "Generate variants?"
• "Upload another image?"

Error Handling

ECONNRESET / timeout: Use fire-and-poll protocol (Step 3). The upload usually completes server-side.

All MCP calls failing: Ask user to check API key and restart Claude Code. Do not attempt workarounds.

Max retries: 2 per upload/edit operation.