ux-audit

An interaction-first audit that walks through a live web app as a real user to find usability and behavioural bugs that static reviews miss. It requires proof of interaction — typing, clicking, sending, observing, and screenshotting — before any verdict, and a sweep that did not actually interact terminates with the verdict Incomplete. The audit exercises threads and every element, runs a multi-pane stress matrix, a visual polish sweep, a component perfection checklist, automated accessibility via axe-core, a pragmatic performance budget (LCP, CLS, INP), a scenario battery of eleven scenarios, and stress recipes including a real-flavour data battery.

Verdicts resolve to exactly one of Pass, Conditional Pass, Fail, or Incomplete, gated by a set of hard gates that auto-fail and cannot be downgraded: zero console errors and warnings during the walkthrough, zero network 5xx, no 403/404 on authenticated pages, no layout collapse at any tested viewport or pane combination, zero axe-core Critical or Serious violations per page, and performance thresholds such as LCP under 4.0s, CLS under 0.25, and INP under 500ms on a representative route. An audit-the-audit meta-check rejects rushed reports, for example when manifest timestamps cluster under half a second apart, screenshots number fewer than twice the routes, or a phase was skipped, and such reports cannot be upgraded to Pass even if everything observed looked fine. A configurable allowlist can suppress known-noisy console or network categories from findings while keeping them in the manifest.

The work proceeds through ordered phases: pre-flight (persona lock, browser tool selection, URL, viewport, and capability tests), discovery, walkthrough, polish, stress, verdict, and fix-and-verify. Pre-flight locks a persona so findings do not drift toward generic impressions, and it mandates the first-time-user lens on every multi-page feature. Browser tooling uses Chrome MCP for authenticated apps to reuse a real logged-in session and Playwright MCP for public sites, with instructions to stop rather than silently fall back if login is impossible. Each finding carries reproduction steps, an evidence path, and a suspected code location, and deeper detail lives in reference files for the performance budget, a11y automation, personas, browser tools, audit config, and project adaptation.