| name | design-critique |
| description | Structured critique in two modes — discovery mode for UX discovery documents (run by a fresh-context agent as the ux-discovery critique gate), visual mode for rendered UI, mockups, and React components. |
Design Critique
Two modes. Pick by artifact: a discovery/design document → discovery mode; a rendered screen/mockup/component → visual mode.
Both modes follow the same admissibility ordering — describe, then analyze, then judge. Judgment vocabulary ("clean", "intuitive", "confusing") is inadmissible until you've described the mechanism that produces it.
Discovery mode
Run with fresh context — never the authoring conversation. Inputs: the artifacts handed to you (typically requester-input.md + 00-framing.md + 03-synthesis.md + 04-design.md, or a final DISCOVERY.md / DISCOVERY-LITE.md) plus the canonical shared context, which is required reading, not contamination: under ../../core/ux-discovery/references/ read product-context.md (personas, principles), design-judgment.md, and domain.md. Rubric items that test a field living only in an artifact you weren't given are skipped, stated as skipped — not reported as missing.
Your brief is adversarial toward the recommendation, not the requester: build the strongest case that the problem isn't worth solving, the evidence is misread, or the chosen direction is wrong — then report what survives that case as honestly as what doesn't. The rubric below is your instrument set, not your deliverable.
Output findings as defects with severity (critical / major / minor), each with the evidence quoted and what it breaks for whom. No verdicts — no "ready for implementation", no "proceed"; readiness is the requester's call. No quotas either: a finding that wouldn't change the design if true is noise, and padding to look thorough is itself a defect — a short list of real findings beats three decorated ones.
1. DESCRIBE — before any evaluation, write down what the proposed design makes each persona actually do: the steps, the reading load, the decisions per screen, what they see first, how they leave. If you can't reconstruct the user's path from the doc, that's finding #1.
2. ANALYZE against each test:
- Problem framing: passes the binary test (names a role, a concrete moment, a cost — zero UI nouns)? Rests on cited incidents with tiers and provenance, or on assertion?
- Evidence honesty: are tiers stamped and consequential (assumed load-bearing claims carry a test or recorded bet)? Any external claim without a source? Any invented precision — formulas, thresholds, or targets with no rationale at the actual scale (real team size, real account count)? When
requester-input.md is among the inputs, spot-check 2–3 load-bearing claims against the verbatim answers: tier inflation, a dropped hedge ("I think" becoming fact), or a customer-proxy answer recorded as customer evidence are each a major defect.
- Gate honesty: the gate tick blocks live at the end of the artifacts (Gate F in
00-framing.md, Gate I in 03-synthesis.md, Gate D in 04-design.md, the tier check atop DISCOVERY-LITE.md) — spot-check two ticks against the lines they point to; a tick whose pointer doesn't hold is a major defect. A tripped lite escalation trigger without a recorded override is a critical finding that blocks presenting: the authoring session must escalate or record the override before continuing (readiness stays the requester's call — this is a defect report, not a verdict).
- Verdict integrity: was not-building / shrinking / solving-another-way genuinely considered, or does the doc design exactly what was asked with rationale decorating it? Does the discovery delta say what changed and why?
- Role disposal: every persona explicitly disposed (designed-for / out-of-scope-because / degraded-accepted / N/A)? Shared surfaces checked for the non-primary roles who will see them?
- Self-consistency: does the design honor the doc's own problem filter and the product principles — or violate them silently? (A principle deviation is fine argued, a defect silent.) Do mechanisms contradict stated stances (e.g. "no thresholds" plus a trend-drop detector)?
- Design judgment (per
design-judgment.md from the required reading): decision test per Primary element; queue completability; alert lifecycle and volume; recommendation five-pack; dual-audience altitude and metric identity; click economics on daily tasks; states designed (empty, re-entry, domain edge cases); "not supported on this channel" honest?
- Domain checks (per
domain.md from the required reading): peak/calendar stress test on trends and stock math; primary-language layout and copy; per-channel semantics not flattened into a generic abstraction; every action the design implies either sits inside the envelope domain.md's automation frontier currently states, or names its human handoff.
- Falsifiability: do key bets carry "we're wrong if…" lines pointing at watched signals? Steelman check (needs synthesis or the final doc among inputs): is the strongest rejected alternative represented in its own best form, or a weakened restatement? When a final document (
DISCOVERY.md or DISCOVERY-LITE.md) is among the inputs, also check: metrics ("How we'll know it worked" in lite) carry number + baseline + source + first check date; the discovery delta is stated; the open-questions ledger is complete (every question answered / deferred / assumed — no Primary element depending on an unanswered one).
3. JUDGE — rank findings by severity, each tied to what it breaks for whom. Separate "fix before build" from "carry as open question".
Visual mode
For rendered UI, mockups, HTML previews, React components.
- First impression (2 seconds): what draws the eye first — is that the right thing? Is the purpose immediately clear?
- Usability: can the user accomplish the goal? Unnecessary steps? Interactive elements obvious?
- Hierarchy: clear reading order; the right elements emphasized; whitespace and typography doing the work (color reserved for state and action).
- Consistency: follows the design system; spacing/colors/type consistent; similar elements behave similarly.
- Accessibility & i18n: contrast, touch targets, readability — and RTL/localization: does the layout survive text expansion, do numerals and directionality behave?
- States: empty, loading, error, and the domain's real edge cases — not just the populated happy path.
Giving feedback (both modes)
- Be specific: "the CTA competes with the navigation", not "the layout is confusing".
- Explain the mechanism: connect each finding to a principle or a user cost.
- Propose the alternative, not just the objection.
- Acknowledge what works — findings land better beside named strengths.
- Match the stage: early exploration gets direction-level feedback, final polish gets detail-level.
