Critique Design

Findings first; praise and summary after. A finding says what is wrong, where, why it matters to the user or the business, and what to do, with evidence someone else can check. Judge the rendered result against the brief and the surface contract. The author's intent is not evidence.

Modes

Prefer a fresh subagent even for your own work: authors over-rate what they made. In self-check, write Track A before opening any tool output; the report starts Method: single-context (<reason>). redesign-brief weighs structure (heuristics section 2) as heavily as detail: detail fixes become every direction's baseline, and equity is what users rely on (learned locations, shortcuts, scanned data), not the layout. Findings feed redesign.

Two tracks, in this order

Track A: judgement, in a fresh context. It gets the brief (and PRODUCT.md), the surface contract, a screenshot manifest and the rubric. It does not get the conversation, the reasoning in decisions.md, drafts, rejected options or the author's self-assessment. Decide what you see before any number anchors you.

Track B: mechanical evidence, opened only after Track A is written. In ../design-studio/scripts/ (run each with --help first): lint.mjs (the deterministic floor on a rendered page), color_tools.py (contrast, matrix --from tokens.css, cvd for charts and status sets), capture.mjs (sizes x themes x states; non-zero exit on walls, blanks, errors), shot.sh (fallback). Installed alone: use what the host has (computed styles, axe, a contrast calculator) and name it. B wins on what it measures; A stands on what B cannot measure. Record agreements, what only B caught, and B's false positives. Floor items no tool checked are not verified: not passed.

Procedure

1. Check the evidence. Open every file: right route, fully loaded (no spinner, wall or error page), intended font, declared size, theme, state and build stamp. Recapture a bad file yourself (step 2); if you cannot, the disposition is recapture: list what to capture and stop.
2. Capture what is missing at the declared sizes and themes (defaults, traps: render-and-look), in every state you can trigger: focus by tabbing, hover, error, empty, loading, longest content, 200% zoom, reduced motion. Walk the primary task. Never judge from code alone or one shot; with code, cite the cause's file:line.
3. State the yardstick from the brief: audience, job, attributes, platform, constraints. With no brief, infer the design read and label it inferred. Note what works; the fix plan keeps it. Brand work: .design/brand/strategy.md and the tests in brand section 4.
4. Coverage pass (Track A). Walk the heuristics in order; run two or three personas. Then rationale against render: for each contract block (thesis, own-world, story, first viewport, form, finish line) and each claim in the critic brief, name the shot and region that delivers it: kept, partial or missing. A rationale the render does not show is design theater; partial and missing are findings. Write down every candidate; do not filter.
5. Track B. Run or open the mechanical evidence and merge it with Track A.
6. Vet pass. Re-open every cited location (shot region, selector, file:line) before a candidate becomes a finding. Drop false positives, fix wrong locations, merge repeats into one systemic finding, then grade it. Keep a one-line reason for each rejected candidate.
7. Score with the rubric: floor, then ceiling, then the gate.
8. Decide, crop, report: disposition, one crop per finding, report.

Personas: pick two or three, walk the primary task as each, name the element that broke

Grading findings

Also tag effort (S/M/L) and type: quick win, systemic (fix once in tokens or components), structural (needs a redesign). Word evidence as evidence ("当前页面显示…"), inference as inference ("这可能导致…").

Dispositions and rounds

At most two rounds. Round 2 re-checks each round-1 finding against its own evidence shot, re-captured at the same size, theme and state: resolved, partial or unresolved, plus at most three regressions. Partial or unresolved never adds up to ship. After round 2, disclose what remains.

Accepting a build (acceptance)

Only screenshots of the build count, from any source (browser, simulator, device, golden test, a paste). With none, acceptance has not happened: say so.

1. Pair each handoff shot (shots in handoff.json) with a build shot at the same logical size, DPR, theme, state and content, carrying the build stamp. A mismatched pair is recapture.
2. Compare region by region: the named components of the spec's structure tree (data-component in the mockup), side by side, then overlaid (render-and-look section 8). Per region: match, adapted, missing, contradicted or added, with a crop. Spacing, type (face, size, weight, line height, tracking), colour tokens, radii, borders, shadows, each asset in its slot at its rendered size, content; then states, other sizes, motion. Tolerances: the spec's Acceptance section (default: one spacing step, exact tokens, fonts as specified). Anti-aliasing is noise.
3. Each deviation: expected -> actual, screen > component, crop, severity, one class: build bug (differs from a clear spec: fix the build), spec gap (the handoff did not say: fix the handoff, never guess), platform difference (a system component or font renders differently: usually accept, record it in the spec).
4. Say how each thing a still cannot show was checked (targets, focus order and visibility, screen-reader names, contrast, text scaling, reduced motion), then fill the acceptance report. ship needs every region match or adapted and no open P0 or P1; a region is fixed only on a re-captured build shot.

Conduct

• Judge the work against the goal, never the person; be direct. Prefer the smallest fix that works.
• Ten findings that matter beat forty. Scores track change, never people. Write in the user's language.
• Never rewrite an issued finding to match what was built. The implementer answers it: fixed, with the re-captured evidence shot (same size, theme and state), or a deviation recorded in decisions.md.

Output

.design/critique/<date>-<target>/: report.md plus one crop per finding (NN-<slug>.png, enough context to locate it), committed; bulk captures stay in .design/shots/ (gitignored). Acceptance writes .design/handoff/<feature>/acceptance.md. The last report on the same target is the baseline for score changes. Quick work: the table in chat.

Feedback

When this skill is corrected, fails or lacks something, follow design-studio's feedback: silent, lab inbox only, short retro.