ux-audit

What it does

Walk through a live web app AS a real user to identify usability and behavioural bugs that static reviews miss. It requires proof of interaction before any verdict and terminates with Incomplete if interaction is not demonstrated. It covers extensive checks: multi-pane stress, visual polish, component perfection, automated a11y (axe-core), pragmatic performance budget, a scenario battery, and stress recipes. Each finding includes reproduction steps, evidence path, and suspected code location. Verdicts include Pass, Conditional Pass, Fail, or Incomplete, with strict hard gates and an audit-the-audit meta-check.

How it works

The skill defines a structured, phase-based workflow:

• Phase 1: Pre-flight with Persona Lock, Browser Tool, URL selection, Viewport setup, and capability tests (screenshot, console read, network inventory, element selector query).
• Phase 2: Discovery with sitemap crawl, thread inventory, and element inventory per route.
• Phase 3: Walkthrough including an Interaction Manifest (MANDATORY) that logs typed actions, clicks, verifications, screenshots, and network activity per page. It requires at least one input, a primary action, a modal/detail pane, a console read, and before/after screenshots. It prohibits a Pass verdict without a complete manifest.
• Phase 4–6: Polish, Stress, Verdict with hard gates and a wrapped findings report. It explicitly Trigger phrases include: 'ux audit', 'walkthrough', 'qa sweep', 'audit the app', 'dogfood this', 'check all pages', 'find what's broken', 'stress the UI'.

When to use it

Use when you need an interaction-first usability and behavior audit of a live web app, especially when you must prove real user interaction and adhere to strict hard gates (e.g., console errors, 5xx networks, layout collapse, axe violations, and performance budgets) before delivering a verdict. The skill requires a claude-code-only environment and is designed to reject rushed reports via the audit-the-audit meta-check.

What it can touch

The skill requires the claude-code tool for execution and mandates integration with real browser tooling to perform interactions, take screenshots, read console output, and inventory network requests. It specifies that each interaction step must log exact timestamps and selectors and that the manifest includes per-page evidence and reproduction details. It relies on a supported tool path for authentication and page interaction as part of Phase 1 capability tests.

Caveats

Hard gates are auto-fail conditions that cannot be downgraded, including any console errors or warnings, 5xx network responses, layout collapse, axe-core violations, and strict performance thresholds. The audit is incomplete if the Interaction Manifest is missing or if phase timings indicate insufficient interaction. The skill requires proof of interaction; otherwise, the verdict is Incomplete. The toolchain must be connected (e.g., Chrome MCP or Playwright MCP) and authenticated; otherwise, the audit cannot proceed.