Product Audit

Ask an agent to audit a product and you get a confident list where a contrast ratio it measured sits next to a server bug it imagined. Your team then fixes what isn't broken and misses what is.

This skill audits a live web product the way a careful senior reviewer does, and it never changes anything on production.

Three layers.

1. The product as a user meets it: key flows at desktop and phone width, every theme, every state (empty, error, success, the way back), with accessibility measured on every page: contrast pairs, names, labels, targets, headings, landmarks.
2. What the client ships: security headers and cookie flags, caching, robots and sitemap, the 404, structured data against what the page renders, the design system as it is actually used (rare values, arbitrary classes, two scales for one thing), and the public JavaScript read as text.
3. The lifecycle of your key scenario: states and transitions, what leaves the browser during a real run on your own account, and every rule the client enforces turned into a hypothesis with a safe test for staging.

Every finding is marked. Confirmed (measured on production, fix it), hypothesis (grounded in client behaviour, test it on staging first), not reproduced (re-test before filing), or cleared (suspected, measured, found sound). Severity is separate: P0, P1, P2.

Numbers are re-measured before they count. Every figure in the summary and every P0/P1 is checked a second way: recomputed independently, re-measured in another theme or viewport, or given to read-only verifier agents. The report shows what the cross-check corrected.

Two reports, ready to send:

• a summary for whoever decides: what you lose now, what to test, what is debt, one paragraph to read if nothing else, a fact-versus-hypothesis register and a decision matrix by effort, with owners;
• a teardown for specialists: every finding with its measurement, why it matters, the smallest fix and how it was re-checked.

Two scripts ship with it: a passive headers check and an in-page measurement that returns contrast failures, design values in use and where each rare one is. The reports are self-contained HTML in light and dark, in your language.

Safe by design. It asks first whose product it is and which account to use. No orders, no form submissions, no replayed requests, no other users' data, no load. When a check would be unsafe, it writes the safe test instead and carries on.