UX critique

An audit is worth having only if its findings are real. The failure mode of this task is not missing problems — it is manufacturing them, because a review that returns little looks like a review that did little. Everything below exists to make that harder.

Run it in two stages

Stage 1 — find. Sweep the interface against the checks in references/sweeps.md for the surfaces present. Be generous here; collect more than you will report. Record evidence as you go, because you cannot reconstruct it later.

Stage 2 — filter, in a separate pass. Take each candidate finding and argue against it. Would a senior designer actually raise this? Is the evidence in hand, or inferred? Is it in scope? Drop anything that survives only because it sounded plausible.

Where a sub-agent is available, run stage 2 as its own agent given only the finding and the artifact — not your stage-1 reasoning. The finder has an interest in looking useful; the filter must not inherit it. If you cannot isolate the passes, say so in the report: ⚠️ single-pass audit — findings not independently filtered.

Before either stage: fix the scope

Write these down first. An audit without them produces arbitrary results.

• The task scenario. "A first-time visitor on mobile trying to buy one item as a guest." Not "the checkout".
• What you can actually observe — a static screenshot, a live URL, the source, the running app.
• What you cannot — hover and focus states, dark mode, logged-in views, error states, real content lengths, other breakpoints, anything behind an interaction you did not trigger.