The author of a pull request is asking "did I make this worse?". A review that answers "here is everything wrong with this codebase" answers a question nobody asked and buries the two lines that matter. Scope the review to the change, then report in review-format's table with one extra column: status.

1. Resolve the scope

With no target named: if HEAD is ahead of git merge-base origin/<default> HEAD, review that range plus any uncommitted changes, and say how many commits and how many uncommitted files. Else, if the tree is dirty, review the uncommitted changes. Else there is no change — ask what to review rather than inventing a scope. Read the whole diff before forming an opinion; a skimmed diff produces findings the next hunk already fixed.

2. Expand to the blast radius

A changed file is evidence, not the subject. Its blast radius is the set of surfaces it renders in — every screen using a changed component, every consumer of a changed token. Review those surfaces, not just the hunk.

3. Read the removed side

Regressions are invisible in the after state. Search the - lines (git diff -U0 … | grep '^-') for signals that left:

Removed	Check
`aria-*`, `role=`, `alt=`, `<label`, `htmlFor`, `scope=`	Lost name, description, announcement or association (screen-reader-announcements)
`<button>`, `<a>`, `<nav>`, `<main>` turned `div` / `span`	Keyboard and assistive behaviour traded for styling (keyboard-patterns)
`:focus-visible`, `outline`, `tabindex`	The focus indicator or the tab stop is gone
`prefers-reduced-motion`, `prefers-contrast`	The user's system preference is now ignored
Logical properties swapped for `left` / `right`; `lang=`, `dir=`	Direction handling dropped (logical-properties-rtl)
`text-wrap`, `line-clamp`, `tabular-nums`	Wrapping or figure alignment silently changed (line-behavior)
A token swapped for a literal or a lighter token	The rendered pair may now fail — measure it
A user-facing string deleted or shortened	A label, error or empty state lost its information

A signal is a lead, not a finding. Equivalent replacements clear it: aria-label → aria-labelledby on visible text; a role dropped because the element became native; outline → a box-shadow ring on the same :focus-visible; a literal → a token of the same value; physical → logical (that is the fix); a string moved into the catalogue. Read the hunk around each removal before deciding — -U0 hides the element it came from.

4. Status by cause

Introduced — the change created it.
Regression — the change weakened something previously correct.
Pre-existing — present before, not made worse.

Status follows cause, not location. An untouched consumer of a changed token that now fails contrast is Introduced; a bug three lines from a hunk that the change did not touch is Pre-existing. Only Introduced and Regression findings count toward severity and the verdict, so a change whose only findings are pre-existing is an approve.

Reporting

| Severity | Status | Location | Before | After | Why |, Introduced and Regression only, ranked by severity then reach. Then at most three pre-existing findings, plainly marked not this change's responsibility. With nothing caused by the change: "No actionable interface findings in this change." Correctness, tests, security and performance belong to the general code review — name such a concern once and drop it.

Gotcha

The most-missed regression is a <button> that became a styled <div> in a "refactor to the new design system" PR. The after state looks identical, so only the - side shows it.

Sources

Jakub Krehel, interfaces skills (MIT, github.com/jakubkrehel/skills, HEAD 592aedc, read 2026-10-25): skills/interface-review/SKILL.md, scope-resolution.md, removed-signals.md. With the companion installed, route a diff review to /interface-review (interfaces-reconciliation).
Related: review-checklist, ui-reviewer, skill-router.