design-review

{
  "name": "design-review",
  "description": "Expert design and UX review of rendered web apps, websites, app UI flows, infographics and design artifacts \u2014 the last pass before a human looks at AI-built UI. Runs deterministic accessibility, contrast, target-size, performance and layout-integrity gates first (column alignment, shared rails, section gaps, text overlap, dead space, column and band voids, implicit grid tracks, divider proximity, affordance, token overloading), then structural, state, component-inventory, craft, flow, design-system and intent-conformance passes against real renders at multiple viewports, and returns severity-ranked findings with pasteable fixes, evidence, and an explicit statement of what was never checked. Every gate reports four numbers rather than one \u2014 failed, passed, could-not-run, and the population examined \u2014 because a check whose pass and cannot-run look identical is indistinguishable from a real measurement: engine capability is measured before anything is read, each computed value carries whether it came from the cascade or from a stylesheet declaration, and the unmeasurable population is a reported number instead of a silent zero. A coverage ledger fixes the surface count before the first capture and its exit code is the verdict, geometry findings are clustered by root cause before they are ranked, and a second gate checks every number in the finished report against what the run actually recorded. Use this whenever someone asks to review, audit, critique, QA, sanity-check or \"give feedback on\" a UI, page, screen, flow, prototype, landing page, dashboard, deck or design \u2014 or says their UI \"looks AI-generated\", \"looks off\", \"feels cheap\", \"needs polish\", \"isn't accessible\", or asks whether something is ready to ship or ready for a human to look at. Also use it before shipping any UI that Claude or another AI tool generated, even when nobody used the word \"review\". When distinctiveness is in sco