Text → screens with Claude (the MCP recipe)
Ask for a screen in plain words — "a checkout for a coffee app" — and get an editable wireframe as a candidate version you review. Master is never touched: generation proposes, a human accepts. This is the third recipe of the AI trilogy, next to review and sketch conversion.
One-time setup
Same server as the other recipes:
claude mcp add wireframes -- node /abs/path/to/scripts/mcp-server.mjs /abs/path/to/project.wfm.json
Add --library /abs/path/library.wfm.json (exported from the app's Library panel) and Claude can place_library_component — place the team's nav bar, don't redraw it.
The recipe
Generate a two-screen checkout flow for a phone app. 1. Call
describe_control_type(no argument) for the vocabulary, then per-type for anything you use — never guess prop names. 2. Create screens withpropose_screen— fullprops, real coordinates, canvas sized for the device. Create target screens first, then the screens that link to them, giving each CTA alink: {target_screen, mode}so the flow is clickable immediately. 3. Callrender_screenon each proposal, LOOK at the SVG, and fix overlaps/misalignment withpropose_edit. 4. Calllint_flowwith the candidate version id — a generated flow should lint clean (no broken links, no accidental dead ends; mark true endpoints final later). 5. Optionally run the review recipe on your own output:get_review_contextworks on the fresh candidate, so generate → self-review → fix is one loop.
Layout constraints (generation, not just review)
The review checklist (Nielsen 10, ergonomics) applies at creation time:
- Canvas sizes: phone 360×700 · desktop 1200×800. Include an
iPhoneor
BrowserWindow control at (0,0) as the device frame.
- 8px rhythm: coordinates and sizes in multiples of 8 where possible;
a consistent left margin (24 on phone, 40 on desktop).
- Tap targets ≥ 44×44 on phone for anything interactive; primary actions
in the lower half of tall screens.
- Text sizes: titles 32–40 tall, body ≥ 20 tall boxes; nothing under 11px
reads on a device.
- One primary action per screen; give it a
link. Back affordances get
link: {mode: "back"}.
- Real-looking content: use
{placeholders}({customerName},{price})
instead of lorem — the app's data fill resolves them in play mode.
- States: if the flow has a list, consider proposing its empty state too.
Verified example (this exact conversation shape is the proof artifact)
The two-call sequence below was run against the real stdio server; the resulting candidate rendered correctly and lint_flow on the candidate reported 0 broken links:
propose_screen"Order confirmed" (phone, final screen content).propose_screen"Checkout" with the pay Button carrying
link: {target_screen: "Order confirmed", mode: "navigate"} and a back Icon carrying link: {mode: "back"}.
render_screenboth ·lint_flow {version}→ no findings on the pair.
What this becomes
Free today via Claude Code/Desktop with your own key; the identical recipe is the one a hosted option would use later, same rail, same tools.