design-review

Imported from fusengine/agents (plugins/design-expert/skills/design-review/SKILL.md). Install upstream with npx skills add fusengine/agents --skill design-review. Copyright stays with the author.

Findings are tagged P0-P3 (Blocking / Major / Minor / Polish) and reported individually — there is NO aggregate score, health score, or letter grade, ever; an aggregated number invites self-grading theater.

The visual review is a bounded loop, maximum 2 fix cycles: the loop exits PASS only once the register-applicable positive floors are actually met, and a plateau (cycle 2 repeating cycle 1's finding) stops immediately even at cycle 1 of 2.

Before any "done" claim, a mandatory in-loop challenger gate judges the deliverable blind — fresh-context, off a rendered PNG only, never fed this procedure's own reasoning. A Block from the challenger must be resolved or owner-accepted before "done" is reported.

Design Review — Final Quality Gate

When

After design-motion (web/webapp) or directly after the mockup (design-ios/ design-android). The last step before reporting any design deliverable done — never skipped, never reported as passed if it wasn't actually run (references/review-procedure.md Failure Handling).

Input

• The complete rendered artifact: components/mockup with animations (if any) and interactive states.
• design-system.md as the audit baseline (tokens, forbidden fonts, contrast floors).
• The Je reproduis line written by design-web/design-webapp — three declared elements, two tagged [corpus: {reference}/{tokens section}] (measured procedure, the craft) and one tagged [sector: {url} — register] (register signal only; [sector: none — register from brief] is valid, an empty slot is not). A missing or untagged line is a blocking Part 1 finding — the binary verdict has no input without it.
• The corpus pages opened (Gate 0 artefact 3): at least two {reference}-recode/index.html, named. The [corpus: …] tags prove the corpus was read, never that it was seen — and a page sourced from markdown alone clears every mechanical check and reads dead. Absent ⇒ blocking Part 1 finding.
• The on-screen inventory (Gate 0 artefact 4): what each section physically carries, plus the screen-holding-floor verdict. Blocking when VISUAL_DENSITY ≤ 3 or DESIGN_VARIANCE ≥ 6; reported otherwise.
• Any declared premium pattern(s); the signature element (design-method Step 2) instead gets a Signature Dominance verdict.
• The plan's two structural picks — hero treatment (design-method/references/macrostructure-bank.md) and body sequence (design-method/references/body-sequence-bank.md). A plan naming neither is itself a detector signal (anti-slop entry 15). brand register; a product deliverable carries neither.
• The register (brand/product, from design-method Gate 0) — an input, never inferred mid-review: check 14's floors, Part 2 item 8's floors and the Lookalike Test (item 8b) all branch on it.

The two assessments — ISOLATED, never in the same head

This is the structural rule the whole procedure hangs on: the mechanical read and the subjective read must not anchor on each other.

1. Deterministic detection (references/anti-ai-slop-audit.md + references/pre-flight-checklist.md) — grep/count/compute, zero taste involved. Run by design-expert itself; there is nothing to isolate here because there is no judgment call to bias.
2. Subjective visual review — screenshots, named elicitation techniques (references/elicitation-visual.md), Nielsen heuristics (references/ux-nielsen.md), UX laws (references/ux-laws.md). This pass is run by design-expert, so it is not yet independent — independence is delivered by the mandatory challenger gate below, which judges blind, fresh-context, off the rendered PNG only, never this procedure's own reasoning. Do not treat the elicitation-technique pass alone as "isolated" — it isn't; the challenger is the isolation.

Never collapse these into one pass "to save a round" — that is exactly how the cream/serif/terracotta cluster shipped through this audit once already (documented in anti-ai-slop-audit.md Gate Semantics).

Part 1 — Deterministic Checks (mechanical, run first)

Full 15-item list is canonical in references/review-procedure.md Part 1 — contrast (4.5:1 text / 3:1 UI, light+dark), forbidden fonts, OKLCH-only color format in the produced code (corpus values arrive as hex/rgb — convert them, never drop them), em-dash crutch (2+ occurrences), token adherence, the anti-AI-slop audit (references/anti-ai-slop-audit.md, 15 detectors — including entry 15, the generic body sequence: canonical order verbatim AND uniform vertical rhythm, a conjunction because the canon treated is legitimate; clusters 9-11 are FLAG-with-justification, not a block, if declared per design-method Step 2), the mechanical pre-flight (references/pre-flight-checklist.md), WCAG beyond contrast (references/ux-wcag.md), cross-component consistency (references/consistency-checks.md), mobile nav functionality, doc↔code animation diff, integrity (no fabricated numbers/false urgency), no-JS baseline, Type-Scale/Body-Size/Measure/Focal-Block floors, and Dark-Elevation Direction. Any Critical/Major finding here is fixed before Part 2 runs.

Part 2 — Bounded Visual Review

Serve → screenshot per section + one full-page capture, light AND dark via colorScheme → cross-viewport (mobile/tablet/desktop) → compare declared elements (binary present/absent, no partial credit; signature element gets Signature Dominance) → localized critique only (name the exact section/element, never "improve the style") → motion verdict if any animation exists (references/motion-verdict.md, Before/After/Why table + tiered impact + explicit Block/Approve) → cite at least two named techniques from references/elicitation-visual.md (Squint / Subtraction / Competitor Line-up / 5-Second / Persona) → fix gaps → Lookalike Test, brand only / product exempt (item 8b — this procedure is the only executant of the test design-web declares; definition canonical in design-web/references/design-inspiration.md, never restated here; binary [distinguishable]/[indistinguishable], and an [indistinguishable] sends the page back to the body sequence pick, never to a cosmetic tweak. On product it does not apply — an app screen is supposed to share a silhouette with other app screens, Jakob's Law — and the Domain-Specificity Floor (design-method/references/register/product.md §2) runs in its slot instead, per design-webapp/SKILL.md). Full step-by-step: references/review-procedure.md Part 2.

Bounded loop — maximum 2 fix cycles. The loop exits PASS only once the register-applicable positive floors are actually met (not merely "0 flags"). Issues (or an unmet floor) remaining after cycle 2 stop the loop and get reported, not chased further. A plateau — cycle 2 repeating cycle 1's finding — stops immediately, even at cycle 1 of 2. This is a hard cap, not a target to exceed when "almost there."

Findings — P0-P3, NO aggregate score, ever

Tag every finding with severity, never sum or average them into a total:

This maps onto references/audit-checklist.md's Critical/Major/Minor column (Critical→P0, Major→P1, Minor→P2/P3 by impact) — but its "Scoring" section (letter grades A-D) is not used here: it is superseded by this rule. No aggregate score, no health score out of N, no letter grade — ever. Same Gate Semantics already enforced in references/anti-ai-slop-audit.md and in design-method's critique.md/audit.md moves: each finding stands PASS/FAIL or tier-tagged independently. An aggregated number invites self-grading theater and has already failed once in this exact pipeline — don't reintroduce it in the report format.

Challenger gate — mandatory, in-loop, not a trailing consultation

Before any "done" claim, design-expert invokes fuse-ai-pilot:challenger (it holds the Task tool) to judge blind: PNG + a short brief, named elicitation lenses, fresh-context — never this procedure's own reasoning fed in as the frame. A Block must be resolved or owner-accepted before "done" (consultative verdict, not a veto — challenger never overrides the owner). Fallback: only if Task/Agent is unavailable (agent at max nesting depth 5) → report "not judged" / escalate to the owner, never a silent "done". Full mechanics: references/review-procedure.md item 9.

Failure Handling

• All server ports 8899-8905 busy → stop, report the deliverable unreviewed, say so explicitly — never report a validation that wasn't executed.
• Screenshot tool fails → retry once; on a second failure, stop and report the gap rather than declaring the visual review passed.

Output

• Part 1 deterministic results, all Critical/Major resolved.
• Light/dark + 3-viewport screenshots (Part 2).
• Binary verdict per declared element/pattern, each cited with its [corpus: …]/[sector: …] tag; Signature Dominance for the signature element.
• Lookalike Test verdict (brand), with the hero treatment and body sequence named — or [n/a — product register; Domain-Specificity Floor run instead: {verdict}]. Never blank, never "passed" by default.
• Motion Block/Approve verdict if applicable.
• Findings list, P0-P3, no aggregate score.
• Challenger verdict: resolved/owner-accepted, or "not judged" on tool-unavailable fallback.
• Any remaining Minor/P2-P3 issues after the 2-cycle cap, reported, not hidden.

Next → report to the owner. A P0/P1 unresolved or challenger-Blocked finding means

the deliverable is not "done" — say so plainly, don't soften it into "mostly ready."

References