Docs · skills/heuristic-audit/SKILL.md

Heuristic audit

Evaluate a live product or a set of screens against usability heuristics and WCAG 2.2, producing severity-rated findings with a cost and a sequenced fix list. Use for a UX audit, a product overview of something that already exists, or a competitive teardown.

An audit that lists sins is worthless. An audit that ranks findings by frequency × severity, attaches a cost, and sequences the fixes is a plan.

When to use

  • A product exists and the question is whether it works
  • Track B of an evaluation — the product-overview entry point
  • Teardown of a competitor, for the same rubric applied consistently

Procedure

  1. Read project-brief.md first. Audit against the job the product is supposed to do, not against generic best practice.
  2. Pick the flows. Two or three critical paths, end to end. An audit of "the product" is an audit of nothing.
  3. Walk each flow in every state you can reach — first run, empty, populated, error, slow network, keyboard only.
  4. Score against reference/rubric.md. Every finding gets: heuristic violated, evidence (screen + what happened), severity, frequency, cost.
  5. Run the accessibility pass separately using the WCAG 2.2 checklist in a11y-spec/reference/wcag22-checklist.md. Do not fold it into the general pass — it has different evidence rules.
  6. Sequence the fixes. Group by where the fix lands (a primitive, a pattern, a screen). Fixing one primitive often closes ten findings — say so.
  7. Write the report using templates/audit-report.md.

Severity scale

LevelMeaning
1 — CosmeticNoticed, not blocking. Fix when nearby.
2 — MinorSlows people down or looks unconsidered.
3 — MajorPeople fail the task or work around it. Fix this release.
4 — CriticalData loss, a dead end, a compliance exposure, or an accessibility barrier. Fix now.

Frequency: how many users hit it, how often. Severity 2 at high frequency outranks severity 3 at the margins — say which you are recommending and why.

Rules

  • Every finding carries evidence. "Feels cluttered" is not evidence; "seven primary actions compete at equal weight in the header" is.
  • Never report a count you did not count.
  • Say what you could not test — no account, no data, no device.
  • Findings without a cost get cut, not softened.

Bundled files

  • reference/rubric.md — the heuristics, with what evidence each requires
  • templates/audit-report.md — the output format

The method, written down

Every page here is one file from the skill library. The folder is loadable as a Claude project as-is.