Product design review

Product Design Review

Review a product the way a strong design lead would: walk its real tasks, judge it against explicit heuristics, and deliver findings a team can act on tomorrow. Prime directive: evidence over opinion. Every finding names where it occurs, what the user suffers, and the concrete fix — and never claims to have seen something you didn't actually render. A review that says "improve the hierarchy" has failed; one that says "the destructive Clear action sits one unconfirmed click from the primary CTA (checkout.html:18) — gate it behind a confirm" has done its job.

When to use / when not

• Use for critique of an existing product surface: a page, flow, app, prototype, or design system in use.
• If the user wants the fixes implemented, deliver the review, then route each fix per the table in step 5 — don't rebuild the UI inside this skill.
• If the product's problem is the promise, not the pixels (nobody wants the thing), hand off to skills/product-strategy/foundation-sprint.
• If the user wants a broad "make it better" engagement rather than a review artifact, hand off to skills/frontend-design/frontend-orchestrator.

Step 1 — Establish evidence mode

Before any judgment, determine what you can actually inspect. Pick the strongest mode available; state which one you're in at the top of the report.

Artifact	Procedure	Evidence tier
Live URL + browser automation available (claude-in-chrome, agent-browser, Playwright)	Render it. Screenshot at ~375px and ~1440px. Tab through the primary flow checking focus visibility. Trigger reachable error, empty, and loading states.	Observed
Live URL, no browser	Fetch the HTML and CSS. You can verify markup semantics, labels, real color/type values, and compute contrast ratios. You cannot see layout, motion, or states — say so.	Computed
Codebase	Read the flow's components, styles, and copy. If the project documents a way to run the app, run it and review rendered output (upgrades findings to observed).	Computed
Screenshots / mockups	Visual, copy, and hierarchy critique of what's shown. Keyboard behavior, states, and motion are unverifiable — list them under "needs live check", don't guess.	Observed (static only)
Figma link	Use Figma MCP tools (get_screenshot, get_design_context) if connected; otherwise ask for exported frames.	Observed (static only)

Tier discipline: tag each finding observed, computed, or inferred. Severity-3 findings require observed or computed evidence. Never present an inferred visual claim ("the spacing feels cramped") as something you saw — either render it or label the inference.

Step 2 — Intake (one batch, then move)

Ask only what's missing, in a single batch: who the primary user is, the one flow that matters most, and what "better" means (conversion, activation, task speed, trust). Infer platform and product type from the artifact itself. Default the accessibility target to WCAG 2.2 AA without asking. If the user is unavailable or the request is self-evident ("review my checkout page"), state your assumptions in the report and proceed — never stall the review on missing context.

Step 3 — Walk the tasks

Define 3–5 representative tasks for the surface under review (for a checkout: find the field, complete payment, recover from an error, back out safely; for a dashboard: orient, find today's anomaly, drill in, act). Attempt each task in your evidence mode as the target user would, and evaluate against the lenses in references/heuristics-and-checklists.md (read it now — it holds the heuristics, the first-use and trust lenses, the WCAG 2.2 quick checks with real thresholds, and the severity rubric).

Scope by request: a single screen gets all lenses on one surface; a flow gets task walkthroughs end-to-end including error paths; a full product audit gets the top 3 revenue- or activation-critical flows — tell the user what you scoped out.

For every issue capture three things at the moment you find it: location (file:line, selector, screen, or timestamp), what the user experiences, and which lens it violates. An issue without a location doesn't go in the report.

Step 4 — Score and synthesize

Severity scores impact if the issue is real — never discount severity because you're unsure; that's what confidence is for.

• 3 Critical — blocks task completion or destroys trust/data for many users (unrecoverable destructive action, broken primary CTA, checkout that can't be completed by keyboard).
• 2 Major — significant friction or drop-off on a core task; WCAG failures on core-task content.
• 1 Minor — noticeable friction; task completes.
• 0 Cosmetic — polish; no task impact.

Confidence (High/Med/Low) scores the evidence: High = observed or computed; Med = strong inference from source; Low = pattern-matched guess (severity ≤ 1 only). Then group findings by task, and split recommendations into quick wins (S effort, high confidence) vs strategic work.

Step 5 — Recommend and route

Every finding gets a concrete fix — the changed value, copy, or pattern, not "improve X". When a cluster of fixes falls into one category, name the sibling skill that implements it:

Fix category	Route to
Hierarchy, clutter, "nothing stands out"	skills/frontend-design/visual-hierarchy-refactoring
Contrast, palette drift, dark mode	skills/frontend-design/color-system
Type scale, readability	skills/frontend-design/typography-system
Spacing, grid, alignment	skills/frontend-design/layout-system
Accessibility remediation	skills/frontend-design/accessibility-excellence
Error messages, validation, recovery	skills/frontend-design/error-handling-recovery
Loading, empty, progress states	skills/frontend-design/loading-states
Dead-feeling interactions, janky motion	skills/frontend-design/interaction-physics
Component/state architecture debt	skills/frontend-design/component-architecture
Slow or slow-feeling surfaces	skills/frontend-design/performance-optimization

Propose an experiment only when a recommendation is genuinely contested or risks the product's identity (e.g., adding onboarding copy to a deliberately minimal product) — state hypothesis, test, and metric. Don't append boilerplate A/B tests to obvious fixes.

Required output format

# Design Review: [product/surface]
**Evidence mode:** [observed via rendered browser | computed from source | static screenshots] — [what this could and could not verify]
**Scope & assumptions:** [flows covered; assumptions made in place of missing context]

## Executive summary
[2–3 sentences: overall verdict + the single most likely cause of the user's complaint]
**Top 3 issues:** [one line each, worst first]
**Quick wins:** [1–3 S-effort/high-confidence fixes]
**Strategic:** [1–2 larger moves]

## Findings
| # | Issue | Evidence (location) | Tier | Sev | Fix | Effort | Conf |
[worst first; every row has a location and a concrete fix]

## Task walkthroughs
### [Task name]
[What happens when the target user attempts it; where it breaks; the fix]

## Accessibility (WCAG 2.2 AA)
[Blockers with criterion numbers and computed values, e.g. "#999 on #fff = 2.8:1, fails 1.4.3"]

## Needs live check
[Claims that require rendering/interaction you couldn't perform — omit section if none]

## Next steps
1–3 immediate actions, then fix clusters routed to skills per the table above.

Quality bar — check before delivering

• Every finding has a location; every severity ≥ 2 finding is observed or computed, not inferred.
• Every contrast claim states the computed ratio; every accessibility blocker cites its WCAG 2.2 criterion.
• Every finding's fix is concrete enough to implement without asking what you meant.
• The executive summary answers the user's actual complaint ("something feels off" → name the thing).
• Nothing in "Findings" contradicts the evidence-mode disclosure; anything unverifiable sits in "Needs live check".
• No recommendation introduces a dark pattern (fake urgency, confirm-shaming, forced continuity), even if it would lift the stated metric.

Integration

• Consumes: a Founding Hypothesis or positioning from skills/product-strategy/foundation-sprint when one exists — review against the promise the product makes.
• Produces: the findings report that skills/frontend-design/frontend-orchestrator or any routed sibling skill uses as its work queue; accessibility blockers feed skills/frontend-design/accessibility-excellence directly.

References

• references/heuristics-and-checklists.md — severity rubric, Nielsen heuristics, first-use and trust lenses, usability/visual/content checklists, WCAG 2.2 AA quick checks with thresholds, platform checks. Read during step 3.