UX Scan

Walk a UI the way a usability expert and a screen-reader user both would, and come back with a numbered, ranked list of concrete problems, each one located (file:line + screen), tied to the heuristic or WCAG criterion it violates, severity-rated, and paired with the specific fix. The scan surfaces; a fixer (the caller, a deepen pass, or a Fix: yes tail) resolves. That separation is the whole design.

Core Insight: A UX finding is worth exactly as much as its address. "Improve the contrast" is advice; "text-gray-400 on bg-white in components/PricingCard.tsx:34 is 2.5:1, needs 4.5:1; use text-gray-600" is work a fixer can bank in one move. So every finding ships the full tuple (WHAT / WHERE / WHY (criterion) / SEVERITY / FIX), and the scan itself stays read-only: it never edits the artifact it's inspecting mid-scan, which is what makes it cleanly re-runnable. Pass N reads what passes 1..N−1 settled and spends its whole budget on new territory.

Preconditions

1. A reachable UI. The scan must see the app, not just the code. Run it (/run or /verify if installed, otherwise start the app yourself; screenshot key flows for webapps). Code-only scanning misses everything about how the thing feels; eyes-only scanning misses everything about why. Do both.
2. Stack identified. The pack's target UIs skew Next.js + React + Tailwind (+ shadcn/ui), some Electron, some Expo/React Native. The checks in references/CHECKLIST.md are tuned per stack; know which you're on before Phase 1.
3. Prior findings read. If <target>/.uxscan/FINDINGS.md exists (or a deepen PASSLOG charge is attached), settled findings are off the table; re-reporting them wastes the pass.

The Pipeline (staged: issue each phase only when the prior finishes)

Phase 0: Orient

Identify the stack and enumerate the surface: routes/screens, the 2–4 flows a real user actually runs (onboard, core action, settings, destructive action), and the component inventory. Run the app and walk those flows once with eyes before reading a line of source. Note the user the app serves: severity is judged against their task, not abstract correctness.

Phase 1: Heuristic sweep (Nielsen's 10)

Walk each key flow against all ten heuristics (expanded with stack-specific tells in references/CHECKLIST.md): visibility of system status, match to the real world, user control/freedom, consistency, error prevention, recognition over recall, flexibility/efficiency, minimalist design, error recovery, help. Log every violation as a draft finding with its location.

Phase 2: Accessibility sweep (WCAG)

Code-level + behavioral, in this order: contrast (measure ratios, never eyeball: 4.5:1 text, 3:1 large/UI), keyboard (tab through every flow, mouse unplugged: reachable, operable, no traps), focus order + visibility (logical sequence, visible ring; Tailwind's focus:outline-none without a replacement is an automatic finding), ARIA + names (icon-only buttons need labels; shadcn/Radix gives you semantics free, composed raw div onClick throws them away), screen-reader labels (images alt, inputs associated labels, RN accessibilityLabel/accessibilityRole), touch targets (44pt iOS / 48dp Android / 24px WCAG floor, plus RN hitSlop). Run the automated floor too (axe via Playwright, eslint-plugin-jsx-a11y). Tools catch ~30–40%; they are the floor of this phase, never the phase.

Phase 3: Interaction & flow sweep

The problems heuristic lists under-weight: missing loading/empty/error states (Next.js routes without loading.tsx/error.tsx), forms that lose input on failure or validate only on submit, destructive actions without confirm/undo, dead-end states with no exit, layout shift and jank, mobile breakpoints that trap content, latency with no feedback past ~400ms, and copy that names the system's internals instead of the user's task.

Phase 4: Verify, rank, emit

Re-check every draft finding against the running app. A finding you can't reproduce gets cut, not hedged. Dedupe (one root cause = one finding, listing all its sites). Severity-rate each: blocker (users fail or are excluded) / major (users struggle or workaround) / minor (friction) / nit (polish), rated by user impact × frequency, not by how interesting the bug is. Write <target>/.uxscan/FINDINGS.md (format below), ranked blocker-first. The scan ends here.

Phase 5: Fix tail (only if Fix: yes)

Standalone invocations may fix: green baseline + checkpoint first, then resolve findings top-down one at a time under the green ratchet (a fix that breaks build/tests/smoke gets repaired or reverted; right-but-red is red), re-verifying each fix by looking, not just by passing. Under deepen, skip this phase; deepen's own loop body owns fixing.

The finding (Phase 4 template)

Every finding self-contained: a fixer with zero context can execute it:

UX-<n>  [severity]  <one-line what>
WHERE   file:line + route/screen + how to reproduce
WHY     heuristic (N1–N10) or WCAG SC (e.g. 1.4.3) + the user consequence
FIX     the specific change: component, class, prop, or pattern to apply

Contract (what orchestrators can rely on)

ux-maximizer runs this 3× over the whole app via deepen (each pass's findings fixed before the next fires), and project-autopilot routes "Usability/heuristic evaluation" here on auto. The interface is fixed:

• <target>/.uxscan/FINDINGS.md: ranked findings, stable UX-<n> IDs, appended per scan under a ## Scan <k> — <date> header; fixed items get [fixed] on their line so later scans treat them as settled.
• Read-only by default: the scan never mutates the target unless Fix: yes. Deepen's loop can therefore trust the artifact it audits.
• Re-runnable: each scan reads prior sections, carries still-open IDs forward in one line, and spends its effort on territory the prior passes' fixes opened up or didn't cover.
• Severity maps to deepen's impact scale: blocker/major → material, minor → minor, nit → nit.

THE EXACT PROMPT

/ux-scan
Target: <project path or running app URL>
Scope:  <optional: one flow/screen, e.g. "checkout"; default whole app>
Fix:    <no (default: surface only) | yes (fix top-down under green ratchet)>

Anti-Patterns

• Findings without an address. "Improve feedback" is advice, not a finding.
• Fixing mid-scan: you mutate the artifact under inspection and corrupt the rest of the pass; surface everything, then fix (or hand off).
• Re-reporting settled findings on a re-run. Pass N's budget belongs to new territory, not re-litigating pass N−1.
• Eyeballing contrast. Measure the ratio; gray-400-on-white looks fine on a calibrated display and fails for half of real users.
• Auditing with a mouse in hand. Phase 2's keyboard walk is done keyboard-only or it isn't done.
• Treating axe/lint output as the audit. Tools are the floor; heuristics, flows, and judgment are the audit.
• Severity inflation: a wall of blockers ranks nothing; rate by user impact × frequency and let most findings be minor, because most are.
• Code-only scanning: if you never ran the app, you scanned a hypothesis.
• Padding a clean pass. Few findings on a re-run is the system working (saturation), not under-delivery; never manufacture nits to look thorough.

References

• CHECKLIST.md: Nielsen's 10 with stack-specific tells, the WCAG criteria list, the interaction/flow problem catalog, and per-stack (Next.js/Tailwind/shadcn, Electron, Expo/RN) check patterns