Read-only judgment of existing UI. This spine routes; the judgment lives in the reference files — load only what the mode needs.
Contract (all modes)
- Read-only. Never edit product code, even for a one-character fix. Findings name the change; applying it is the pe-build skill's job — hand off by name.
- Engine rules govern every mode. Read
references/engine.mdfirst, always: its evidence bar, escalation triggers, cheaper-fix ladder, consolidation, and the verdict format inreferences/review-format.md. - Severity is HIGH / MEDIUM / LOW plus the engine's escalation triggers. Accessibility findings are never discarded.
Modes
| Mode | When | Load |
|---|---|---|
| change | A diff, branch, PR, or "review my changes" is named. Always wins over screen when both could apply. | change-review.md, scope-resolution.md, removed-signals.md |
| screen | A screen, page, site, or screenshot; "review this UI." | screen-review.md, experience-rubric.md, technical-rubric.md |
| improve | "Audit this and give me a roadmap", "improve the app" — whole-surface improvement, delivered as self-contained plans another agent executes. Explicit invocation only. | audit-method.md, plan-template.md |
| guidelines | "Check against best practices / the guidelines"; terse file:line lint of UI code. | web-interface-guidelines.md |
| stress | "Does this component survive?" — render one component in every hostile state. Explicit invocation only. | stress.md, stress-scenarios.md |
| motion | "Review the animations / motion." For a whole-codebase motion audit with plans, add motion-audit.md + motion-audit-signals.md + plan-template.md. |
motion-review.md, motion-standards.md |
| opportunities | "What could be animated here?" — read-only; proposes motion with exact values, never implements. Explicit invocation only. | motion-opportunities.md |
| a11y | "Is this accessible?" — audit, not implementation. | a11y-audit.md (+ engine triggers) |
| fidelity | "Did we stay true to the mock / design?" — the implementation against the approved artifact in .product/approved/; every deviation is a finding with severity. Implied whenever a reviewed surface has an approved record. |
fidelity.md |
Disambiguation: verb beats noun — "review the animation" is motion mode here; "fix the animation" is not this skill at all (pe-build). A request that mixes modes runs the engine once and consolidates, never two parallel reviews.
Name mapping
Some references route to skills by name. Those names resolve as follows — apply the mapping silently:
| Text says | Use |
|---|---|
| better-accessibility | a11y-audit.md for ordering; the engine's own triggers for severity; the pe-build skill's references/a11y/ for depth |
| better-ui, better-layout, better-typography, better-colors, better-writing | the pe-build skill's references/craft/{ui,layout,typography,colors,writing}/ |
| better-interface | this skill's engine mode itself |
House rules
- Autocomplete: on for purposeful fields;
offonly for non-auth-sensitive cases. - P0–P3 appears only inside screen mode's files; report severity as HIGH/MED/LOW (P0+P1 → HIGH, P2 → MEDIUM, P3 → LOW).
Handoffs
Fixes → pe-build (findings include exact file:line and the named change). Proof that behavior works end to end, with recordings → pe-verify; this skill judges quality, that one runs the checks. Documenting the system a review revealed → pe-design, understand mode. Never resolve a finding by editing anything yourself.