Review design

Design Review

Comprehensive UX/UI/design quality analysis across 8 audit dimensions, each mapped to authoritative standards. This skill performs analysis only - it identifies design issues, explains findings against a cited criterion, and suggests fix approaches without making code changes.

It supports two input modes:

• Live mode — audits a running URL via the agent-browser CLI (screenshots + DOM/accessibility snapshot).
• Static mode — audits a codebase (CSS/components/design tokens) via Grep.

When both a URL and a codebase target are supplied, it runs both and emits two separate reports.

Audit Taxonomy (8 Dimensions)

#	Dimension	Key Standards
1	Visual Hierarchy & Layout	NN/g, Refactoring UI, Laws of UX, MD3 8pt grid
2	Typography	Butterick, Refactoring UI, MD3, Apple HIG
3	Color & Theming (incl. Dark Mode)	MD3 color roles, 60-30-10, WCAG 1.4.1/1.4.3
4	Depth & Elevation (Shadows)	MD3 elevation, Refactoring UI, Josh Comeau
5	Components & Affordance (Buttons/Icons)	MD3 buttons/icons, Apple HIG, NN/g
6	Feedback & States	NN/g visibility-of-status, MD3 state layers
7	Motion & Microinteractions	NN/g animation, Laws of UX (Doherty), MD3 motion
8	Accessibility (cross-cutting)	WCAG 2.2 AA

Measurable criteria for dimensions 1–4 are in references/criteria-foundations.md; dimensions 5–8 are in references/criteria-interaction.md.

Anti-Hallucination Guidelines

CRITICAL: Design reviews must be based on ACTUAL evidence, never assumptions:

1. Observe before claiming - Never report an issue without reading the code (static) or viewing the screenshot/snapshot (live)
2. Evidence-based findings - Every finding cites a file path + line number (static) OR a screenshot region + DOM ref (live)
3. Cite a criterion - Every finding maps to a criterion ID and an authoritative citation (WCAG SC, MD3 spec, etc.)
4. Measure, don't estimate - Report actual values (contrast ratio, px size, ms duration), not guesses. In static mode there is no rendered page to sample from — compute the WCAG relative-luminance contrast ratio directly from the two hex/rgb values found in the CSS/tokens (formula in references/agent-prompts.md); never eyeball a ratio
5. Applicable-only scoring - Only score dimensions that apply to the target; never penalize what cannot be observed
6. State what was NOT checked - Every report ends with an explicit coverage gap section
7. No invented standards - Only reference real WCAG SCs, MD3 specs, and HIG guidance

Review Workflow

Phase 0: Parse Arguments & Resolve Mode

Parse arguments to determine target and mode:

Arguments:
- <url>: A live URL to audit (http/https) → live mode
- <pr_number>: Files changed in a PR (e.g., "123", "#123") → static mode
- <commit_sha>: Files changed in a commit (e.g., "abc123") → static mode
- "--all" or a path or no args: Entire codebase → static mode
- "--mode [live|static|both]": Force a mode
- "--scope [dimension]": Focus on one dimension (e.g., typography, color, accessibility)

Mode resolution:

• A URL argument (or --mode live) → live audit.
• A PR / commit / --all / path (or --mode static) → static audit.
• Both a URL and a codebase target present (or --mode both) → run both, emit two reports.

For static PR/commit scope, get changed files:

# For PR
gh pr view <pr_number> --json files --jq '.files[].path'
# For commit
git diff-tree --no-commit-id --name-only -r <commit_sha>

Phase 1: Discovery

Use an Explore agent (model: haiku) to detect the design system and tooling:

Use Task tool with Explore agent (model: haiku):
- prompt: "Discover the project's design system and front-end stack:
    1. Detect CSS approach: Tailwind, CSS Modules, styled-components, vanilla CSS, Sass
    2. Detect component/design system: Material UI, Chakra, Radix, shadcn/ui, Ant Design, custom
    3. Find design tokens: tailwind.config.*, theme files, CSS custom properties (:root { --* }), tokens.json
    4. Find existing a11y tooling: eslint-plugin-jsx-a11y, axe, pa11y, Storybook a11y addon
    5. Detect dark-mode support: prefers-color-scheme, .dark class, data-theme
    6. List the key UI directories (components/, styles/, app/, pages/)
    Return: a stack + design-system summary with the token source of truth."
- subagent_type: "Explore"

Without a Task tool, run this discovery yourself inline using the same 6 steps before proceeding.

For live mode, confirm the URL is reachable and agent-browser is installed:

agent-browser --version || echo "agent-browser not installed (npm install -g agent-browser)"

If agent-browser is missing, tell the user how to install it (npm install -g agent-browser && agent-browser install) and offer to fall back to static mode.

Phase 2: Initialize Progress Tracking

Use TodoWrite to track progress across the 8 dimensions plus capture, consolidation, and report generation. One todo per dimension. Without TodoWrite, narrate phase/dimension transitions in your responses instead.

Phase 3: Capture, then Spawn Parallel Audit Agents

Live mode — capture first (the agents analyze these artifacts):

mkdir -p /tmp/review-design
agent-browser open <url>
agent-browser snapshot -i > /tmp/review-design/snapshot.txt   # interactive elements + a11y tree
agent-browser screenshot /tmp/review-design/page.png --full     # full-page screenshot
# Optional, for inspected elements:
# agent-browser get styles @<ref>   # computed CSS
# agent-browser get box @<ref>      # bounding box (for touch-target size)

Capture additional viewports/pages if the user names them. Always agent-browser close when done.

Then spawn 6 parallel Explore agents (model: sonnet) covering the 8 dimensions. For the full per-agent prompts (live screenshot/snapshot analysis AND static grep patterns), see references/agent-prompts.md. Without a Task tool, run the 6 dimension analyses sequentially inline yourself, one after another, using the same per-agent prompts and criteria references.

Agent assignments:

• Agent 1: Visual Hierarchy + Layout & Spacing (Dimension 1)
• Agent 2: Typography (Dimension 2)
• Agent 3: Color + Dark Mode (Dimension 3)
• Agent 4: Depth/Shadows + Components/Affordance (Dimensions 4, 5)
• Agent 5: Feedback & States + Motion/Microinteractions (Dimensions 6, 7)
• Agent 6: Accessibility — WCAG 2.2 AA (Dimension 8, cross-cutting)

Each agent must:

1. Read the relevant criteria reference for its dimensions (criteria-foundations.md or criteria-interaction.md)
2. Live: visually analyze the screenshot and cross-reference the DOM/a11y snapshot Static: Grep CSS/components/tokens for measurable failures, then Read each match to verify
3. Record the measured value (contrast ratio, px, ms, dp) as evidence
4. Map each finding to a criterion ID + citation (WCAG SC / MD3 / HIG / NN/g)
5. Classify severity (Critical/High/Medium/Low)
6. Provide a concrete fix (2-3 approaches with the target value)

Phase 4: Consolidate & Analyze Findings

After all agents complete:

1. Collect findings from the 6 agents
2. Deduplicate across agents (same element/criterion = one finding)
3. Prioritize by severity:
   • Critical: WCAG AA failure blocking use (contrast < 3:1 on text, no keyboard focus, missing form labels), unusable touch targets
   • High: WCAG AA contrast failures (< 4.5:1 body text), missing focus-visible, no reduced-motion support, broken hierarchy
   • Medium: Off-grid spacing, un-tinted shadows, weak typographic scale, missing hover/active states
   • Low: Polish — minor inconsistency, sub-optimal line length, icon-label spacing
4. Map to dimension: Group findings under the 8 dimensions
5. Applicable-only score: For each dimension that applies, score = passed criteria / applicable criteria. Skip dimensions that cannot be observed and say so.
6. Statistics: total findings, by severity, by dimension; elements/files reviewed vs. those with issues

Phase 5: Generate Report(s)

Generate a markdown report per the template in references/report-template.md.

• Live audit → design-report-live.md
• Static audit → design-report-static.md
• Both modes → produce both files.

Phase 6: Verification & Quality Gate

Before presenting, verify every finding has:

1. Evidence — file:line (static) or screenshot region + DOM ref (live)
2. A measured value where one applies (ratio / px / ms / dp)
3. A criterion ID + authoritative citation
4. A concrete fix with the target value
5. A justified severity
6. No duplicates, no placeholder text ("TODO", "[example]", "lorem")
7. An explicit "What Was NOT Checked" section
8. Applicable-only scoring (no penalty for unobservable dimensions)

Usage

# Live audit of a running site
review-design --url https://example.com
review-design https://example.com

# Static audit of the codebase
review-design --all --mode static
review-design 123                  # PR #123 changed files
review-design abc123def            # a commit

# Both: live + static (two reports)
review-design https://staging.example.com --all --mode both

# Focus a single dimension
review-design --url https://example.com --scope accessibility
review-design --all --scope typography

Scope Options

--scope focuses the audit on one dimension: hierarchy, layout, typography, color, dark-mode, depth, components, feedback, motion, accessibility. If omitted, all 8 dimensions are audited.

Additional Resources

• references/criteria-foundations.md — measurable criteria for dimensions 1–4 (hierarchy, layout, typography, color, depth)
• references/criteria-interaction.md — measurable criteria for dimensions 5–8 (components, feedback, motion, accessibility)
• references/agent-prompts.md — per-agent prompts (live + static) and grep patterns
• references/report-template.md — live and static report templates

What This Skill Does

• Audits UX/UI/design quality across 8 standards-mapped dimensions
• Works against a live URL (agent-browser) or a static codebase (grep)
• Measures real values (contrast, sizes, durations) as evidence
• Maps every finding to a criterion ID + authoritative citation
• Produces severity-ranked, fix-oriented report(s)

What This Skill Does NOT Do

• Does not modify any code or design files
• Does not auto-fix issues or commit changes
• Does not run a full automated a11y scanner (axe/pa11y) — it reasons from evidence
• Does not evaluate native mobile (SwiftUI/Compose/Flutter) static patterns (web-first; future extension)
• Does not guarantee 100% issue detection

Limitations

• Heuristic + measured: combines tool measurements with expert heuristics; some judgment calls remain
• Live mode needs a reachable URL and the agent-browser CLI installed
• Static mode is pattern-based: dynamic/runtime states may be missed
• Screenshot analysis depends on render fidelity at the captured viewport
• Expert review recommended for high-stakes or regulated interfaces

Standards References

• WCAG 2.2
• Material Design 3
• Apple Human Interface Guidelines
• Nielsen Norman Group
• Refactoring UI
• Laws of UX