FlowSpec Architect (Web / Chrome)
Same three workflows as architect, but uses Chrome browser tools instead of MCP — for users without the desktop app.
Core pattern: Generate JSON locally in Claude Code -> inject into flowspec.app via Chrome tools -> iterate.
Requires: Claude in Chrome extension (mcp__claude-in-chrome__* tools).
Prerequisites & Auth Check
- Call
mcp__claude-in-chrome__tabs_context_mcpto find an existing flowspec.app tab. - If no tab exists:
Navigate to https://flowspec.app/projects to browse and create projects from the dashboard, or go directly to https://flowspec.app/editor to open the canvas.mcp__claude-in-chrome__tabs_create_mcp(url: "https://flowspec.app/projects") - Use
mcp__claude-in-chrome__read_pageto check for auth state:- Look for
.signout-btnor user email display to confirm signed-in state - If not signed in: instruct user to sign in manually at flowspec.app/sign-in
- Look for
- Wait for the dashboard or editor to fully load before proceeding.
Interaction Model
| Action | Method |
|---|---|
| Generate JSON | Locally in Claude Code (same logic as architect) |
| Import JSON into project | javascript_tool with synthetic file injection (see chrome-recipes.md) |
| Export JSON from project | Click Export button via javascript_tool, then read_page |
| Create new project | Click "New Project" button via browser tools |
| Upload wireframe images | User does manually (browser file picker security) |
| Annotate screen regions | User does manually (drag-to-create is more ergonomic) |
| Auto-layout | Click toolbar button via browser tools |
Fallback: If any automation step fails, always provide manual instructions so the user can complete the step themselves.
UC1: Wireframe -> Dataflow (Web)
Step 1: Analyse wireframe images
Use Claude vision to examine wireframe images the user provides:
- Identify UI regions, data elements, forms, displays
- Note interactive elements and data relationships
Step 2: Interview user
Ask about data types, sources, business logic, constraints — same as architect UC1 Step 4.
Step 3: Build JSON spec
Assemble the JSON following the v1.2.0 schema.
Step 4: Navigate to flowspec.app
Ensure we have a flowspec.app tab open (preferably at /projects for the dashboard):
mcp__claude-in-chrome__tabs_context_mcp
If needed, navigate to https://flowspec.app/projects and create a new project from the dashboard, or go directly to https://flowspec.app/editor.
Step 5: Import JSON
Use the synthetic file injection recipe from chrome-recipes.md to inject the JSON into the web app's import handler.
Step 6: Guide wireframe upload
Instruct the user to:
- Open the Screens section in the sidebar
- Click "Create Screen" or use the image dropzone
- Upload their wireframe image files
- Name each screen
Step 7: Guide region annotation
Instruct the user to:
- Open a screen in the editor
- Draw regions over the wireframe by clicking and dragging
- Name each region and link data elements using the region panel
Step 8: Export and iterate
Use browser tools to click "Export JSON" and read the result. Review and refine the spec, then re-import if needed.
UC2: Codebase -> Dataflow (Web)
Steps 1-7: Analyse locally
All codebase analysis happens locally in Claude Code — identical to architect UC2 Steps 1-7. See the architect skill's codebase-analysis.md methodology (which is the same reference used by the MCP-powered skill).
Step 8: Create project and import
Navigate to https://flowspec.app/projects, create a new project from the dashboard, and import the JSON via Chrome tools.
Step 9: Guide auto-layout
Instruct the user to click the Auto Layout button in the toolbar, or use javascript_tool to trigger it.
Step 10: Export and validate
Export JSON via browser tools and validate the spec locally.
UC3: Design New Screens (Web)
Step 1: Obtain screen image
User designs in an external tool and provides screenshots, or starts with an empty screen.
Step 2: Guide screen creation
Instruct the user to create screens and upload images via the web UI.
Step 3: Build JSON for data architecture
Create dataPoints, components, transforms for the screen's data needs.
Step 4: Import JSON
Import via Chrome tools using the synthetic file injection recipe.
Step 5: Guide region annotation
Same as UC1 Step 7 — user annotates regions manually in the web UI.
Limitations
- Cannot automate authentication — user must sign in manually
- Image upload requires user interaction — browser file picker security prevents automation
- Region annotation is best done manually — the web UI's drag-to-create is more ergonomic than coordinate-based automation
- Large JSON files may need manual file import as fallback — copy JSON to clipboard, user pastes or uses file dialog
- DOM selectors may change — if the web app UI updates, the Chrome recipes in
references/chrome-recipes.mdneed updating
Chrome Tool Reference
| Tool | Purpose |
|---|---|
tabs_context_mcp |
Get current browser tab info |
tabs_create_mcp |
Open new tab |
read_page |
Read page content/DOM |
javascript_tool |
Execute JavaScript in page |
navigate |
Navigate to URL |
find |
Find elements on page |
form_input |
Fill form fields |
computer |
Click/interact with page |
See chrome-recipes.md for concrete JavaScript snippets.
Converted and distributed by TomeVault — claim your Tome and manage your conversions.