AI design review with Claude (the MCP recipe)
One ask — "review my checkout flow" — and expert critique lands as pinned, dismissible comments on the canvas, each anchored to the control it's about. Reviews comment; they never edit. The MCP server enforces it: review findings are comments, and any edit Claude ever makes lands as a candidate version a human accepts or rejects.
One-time setup
Same server as SKETCH-TO-WIREFRAME.md:
claude mcp add wireframes -- node /abs/path/to/scripts/mcp-server.mjs /abs/path/to/project.wfm.json
The recipe
Review the "Checkout" screen (or: every screen in the Pantry flow). 1. Call
get_review_contextfor the screen — it returns the rendered SVG, the control list, links in and out, project variables, and prior comments in ONE call. 2. Judge it against the checklist below. LOOK at the SVG; don't guess from the control list alone. 3. For each finding, calladd_commentwithkind: "review", anchored to the control id it concerns (screen-level only when it's truly global). One finding per comment, specific and actionable. 4. Do NOT propose edits. If asked "fix it", that's a separate request that goes throughpropose_editand human review.
Findings appear as pins on the canvas and grouped under Review → AI design review, with per-screen and bulk resolve/dismiss.
The checklist (keeps review quality repeatable)
Nielsen's ten heuristics, phrased for wireframes:
- Visibility of status — does the screen show where the user is and what's happening?
- Match to the real world — labels in the user's words, not the system's.
- User control — a way back / cancel / undo from everything destructive.
- Consistency — same action, same words, same placement across screens.
- Error prevention — confirmations before destructive actions; constraints before errors.
- Recognition over recall — options visible, not memorized (check ComboBox contents).
- Flexibility — shortcuts for repeat users where the flow is long.
- Minimalism — anything on the screen that doesn't serve the task?
- Error recovery — do error states exist at all? (wireframes usually forget them)
- Help — is guidance present where a first-timer will stall?
Mobile ergonomics (for phone-frame screens):
- Tap targets ≥ 44×44px — flag anything interactive smaller than that (the
control list carries exact w/h).
- Thumb reach: primary actions in the lower half of tall phone screens.
- Text under ~11px is unreadable on device.
Flow integrity (from linksOut / linksIn / deadEnd):
- Dead ends:
deadEnd: truescreens the user can enter but not leave. - Orphans: screens nothing links to (check
get_flowfor the project-wide view). - Back-less overlays: overlay links whose screen offers no dismissal.
- Missing states: lists with no empty state; forms with no error state;
actions with no confirmation where heuristic 5 demands one.
Copy tone: labels that are vague ("Submit", "OK"), inconsistent ("Sign in" vs "Login" — find/replace fixes this later), or jargon.
Scope patterns
- One screen: one
get_review_context, ~5–10 findings. - A flow:
get_flowfirst for the map, thenget_review_contextper
screen on the path, plus flow-integrity findings at the junctions.
- Whole project:
list_screens, then batch by area; lead each batch with
the worst three findings so the signal survives the volume.
What this becomes
Today this runs free via Claude Code/Desktop with your own key. The identical recipe is the one a hosted option would use later — same tools, same rail, a server-side key instead of yours.