visual-audit-ca1b2f— agent skill

Configuration (fill in before first run)

Replace these placeholders: - [SITEMAP_OR_ROUTE_LIST] - e.g. docs/sitemap.md, docs/routes.md, src/router/index.ts - [DEV_URL] - e.g. http://localhost:3000, http://localhost:5173 - [DEV_COMMAND] - e.g. npm run dev, pnpm dev - [TEST_ACCOUNTS] - reference to test credentials (e.g. "see CLAUDE.md Environment section") - [LOGIN_ROUTE] - e.g. /login, /sign-in, /auth

>

If [SITEMAP_OR_ROUTE_LIST] or [DEV_URL] is not filled, the skill reports an error and exits.

Scope boundary: this skill covers aesthetic quality only. Accessibility contrast (APCA measurement, WCAG contrast thresholds, axe-core scan) lives in /accessibility-audit - run it for any contrast or accessibility concern. Design-token compliance → /ui-audit. Viewport fit → /responsive-audit.

Step 0 - Mode + target detection

Parse $ARGUMENTS:

Mode (controls page roster breadth):

• quick → 5 key pages only
• full → all routes from sitemap, both themes
• No mode keyword → standard - all routes from [SITEMAP_OR_ROUTE_LIST] (light + dark)
• critique → deep-dive on a single target:page: (required). Skips scoring table; produces Gestalt + hierarchy diagnosis + concrete redesign proposals per Critical/Major finding. Must be paired with target:page:<route>.

Target (filters the page roster):

• target:role:<role_name> → routes accessible to that role only
• target:section:<name> → routes whose path contains <name>
• No target → NO filter - use ALL pages from the mode roster per [SITEMAP_OR_ROUTE_LIST]

Mode and target are independent - quick target:role:<role_name> means "run quick mode but limit to that role's routes".

STRICT PARSING - mandatory: derive mode and target ONLY from the explicit text in $ARGUMENTS. Do NOT infer target from conversation context, recent work, active block names, or project memory. If $ARGUMENTS contains no target: token → apply NO filter (use all pages from the mode roster per [SITEMAP_OR_ROUTE_LIST]).

Announce at start: Running visual-audit in [STANDARD | QUICK | FULL] mode - scope: [FULL | target: <resolved description>]

Step 1 - Load context (parallel)

Read in parallel:

1. [SITEMAP_OR_ROUTE_LIST] - route inventory, role mapping, key components per page
2. Global styles file (e.g., [GLOBAL_STYLES_FILE]) - token definitions (brand colour, base scale, semantic tokens)
3. Theme styles file (e.g., [THEME_STYLES_FILE]) - color scale for light/dark

Hold in working memory:

• Brand color token and its approximate hue (for V4 colour discipline checks)
• Base scale: from lightest to darkest
• Semantic token map: surface, secondary surface, text, secondary text, border, brand accent - using the names from Stack adaptation
• List of all routes with their roles and key component files

Apply target filter from Step 0 to the route list from sitemap. The filtered list is the working roster for all subsequent steps.

Step 2 - Pre-flight check

Navigate to [DEV_URL]. If not reachable:

❌ Dev server not running. Start with [DEV_COMMAND], then re-run /visual-audit.

Record the base URL that responded - use it for all subsequent navigations.

Step 3 - Visual evaluation framework

Read ${CLAUDE_SKILL_DIR}/DIMENSIONS.md for the full scoring rubric: 10 dimensions (V1-V7, V9-V11), score anchors, scoring rules, and interpretation notes. Contrast is scored by /accessibility-audit, not here.

Apply all dimensions to every screenshot captured. Key scoring thresholds:

• Score 1-2 → Critical · Score 3 → Minor · Score ≥ 4 → no finding
• Every observation must be concrete and actionable - no "good overall"
• Code-grounded: flag code-level violations even if the screenshot looks acceptable

Step 4 - Page roster

Standard mode (default) and Full mode

Both derive the working roster from [SITEMAP_OR_ROUTE_LIST] already loaded in Step 1. Use all routes with a page file (exclude auth callback routes - they are route handlers with no renderable UI). Group routes by access level for organized reporting.

How to pick the role/credential for each route: use the "Roles" column from the route inventory. For routes accessible by multiple roles, default to the role that sees the most UI. For routes with explicit role variants, screenshot each variant separately under its own label.

Announce total route count at start: "Standard mode - N routes from route inventory."

full keyword is retained as an explicit synonym for scripting/pipeline use - behaviour is identical to standard.

Quick mode (4-5 representative pages - use when speed matters)

Pick 4-5 routes that cover: (1) first impression / auth page, (2) primary dashboard, (3) key data table or list, (4) admin or restricted area. Use [SITEMAP_OR_ROUTE_LIST] to select.

Critique mode

Single route only - requires target:page:<route>. See Step 0.

Step 5 - Screenshot session

5a - Per-page code inspection (MANDATORY before each screenshot)

Specifically note:

• Which design tokens or style declarations are used on the main container, headings, primary CTA buttons
• Whether the brand accent token appears on any row-level or repetitive button (→ V4 flag)
• Whether the project's semantic tokens (as defined in Stack adaptation) are used consistently
• Whether empty state, loading, and error states are handled in the component
• Empty state quality (→ V7/V11): does the empty state have a CTA guiding to the next action (not just a generic "no records" message)? Is the empty state text role-aware? Does the error state name the specific problem and suggest a recovery action?
• Interaction states (→ V11): for every custom interactive element, verify hover/focus/active/disabled/loading states are explicitly defined in the component code

This code context MUST inform the scoring in Step 6 - reference it explicitly when writing observations.

5b - Login helper

1. browser_navigate [base_url]/[LOGIN_ROUTE]
2. browser_wait_for: login form visible
3. browser_type email/username field → [credential]
4. browser_type password field → [test password from TEST_ACCOUNTS]
5. browser_click submit button
6. browser_wait_for: url no longer contains [LOGIN_ROUTE]
7. Confirm the authenticated role is visible in the app UI (navigation, header, or profile indicator)

Credentials: use test accounts from [TEST_ACCOUNTS] in CLAUDE.md (Key Commands or Environment section).

5c - Per-page capture loop

For each page in roster:

1. Read page + component files (Step 5a) BEFORE navigating
2. Switch role if needed (navigate to login page, sign out, re-login with the appropriate credential)
3. browser_navigate to route
4. browser_wait_for network idle or main content visible (2000ms max)
5. DOM preflight validation - run this evaluate before any screenshot:

   ({
     loaded: document.readyState === 'complete',
     hasMain: (document.querySelector('main')?.innerText?.length ?? 0) > 30,
     noError: !document.title.toLowerCase().includes('error') &&
               !document.body.innerText.includes('Application error') &&
               !document.body.innerText.includes('500')
   })

If hasMain === false OR noError === false:

⚠️ Page [route] failed preflight: [which flag was false]. Skipping screenshot - investigate page state before retrying.

Do NOT analyse an empty or errored page - flag it and continue to the next page.

1. Computed checks (run in light mode after preflight passes):

   // Typography scale - count distinct font sizes
   const fontSizes = [...new Set(
     [...document.querySelectorAll('h1,h2,h3,h4,p,td,th,label,span,a')]
       .map(el => parseFloat(getComputedStyle(el).fontSize))
       .filter(Boolean)
   )].sort((a, b) => a - b);

   // Spacing grid - spot check 3 heterogeneous elements (V2 extended)
   // Selectors come from [SPACING_SAMPLE_SELECTORS] in Stack adaptation.
   // Fallback: generic selectors that work across most frameworks.
   const spacingSelectors = ['article, .card, [role="article"]', 'td', 'form, [role="dialog"]']; // override with Stack adaptation [SPACING_SAMPLE_SELECTORS]
   const checkEls = spacingSelectors
     .map(sel => document.querySelector(sel))
     .filter(Boolean);
   const paddings = checkEls.map(el => ({
     tag: el.tagName, pad: parseInt(getComputedStyle(el).paddingTop)
   }));
   const misaligned = paddings.filter(({pad}) => pad % 4 !== 0);

   // Transition timing - check interactive elements
   const transitions = [...document.querySelectorAll('button, a[href], [role="button"]')]
     .map(el => getComputedStyle(el).transitionDuration)
     .filter(d => d && d !== '0s')
     .map(d => parseFloat(d) * 1000); // convert to ms

   return {
     fontSizeCount: fontSizes.length,
     fontSizes,
     paddingSpotCheck: paddings,
     paddingMisaligned: misaligned,
     transitionCount: transitions.length,
     transitionsOutOfRange: transitions.filter(ms => ms > 0 && (ms < 100 || ms > 400))
   };

Record results. Use to inform V1 (font scale), V2 (padding spot-check + misaligned), V7 (transition timing). (Contrast measurement moved to `/accessibility-audit` Step 3.)

1. Light mode screenshot: ensure current theme is light (detect with the snippet in Step 5d; apply [THEME_TOGGLE_ACTION] if the page loaded in dark mode)

1. Dark mode: apply [THEME_TOGGLE_ACTION] → browser_wait_for 500ms → capture screenshot. (Dark-mode contrast measurement moved to `/accessibility-audit`.)

1. Immediately analyse both screenshots + computed data against V1-V7, V9-V11 using the code context from Step 5a

Do not batch all screenshots then analyse. Analyse immediately after each pair - avoids context overload and produces sharper observations.

5d - Theme toggle interaction

Theme switching is project-specific. Use [THEME_TOGGLE_ACTION] from Stack adaptation (see bottom of this skill). Detect the current theme with:

// Robust theme detection across frameworks
const isDark = document.documentElement.classList.contains('dark') ||
               document.documentElement.getAttribute('data-theme') === 'dark' ||
               window.matchMedia('(prefers-color-scheme: dark)').matches;

[THEME_TOGGLE_ACTION] must be a DOM action (preferred - click on the theme control) or a code snippet (fallback - direct class manipulation). If the project has no in-app theme toggle, flip OS-level dark mode or rely on prefers-color-scheme simulation via browser_evaluate.

Step 6 - Per-page analysis

For each page, produce:

### [P##] - [Route] ([Role])

**Code context**: [key classes observed in code that informed scoring - 2-3 bullet points]

**Computed data**:
- Font sizes: [N distinct values - list]
- Padding spot-check: [{tag: tagName, pad: Npx}] · Misaligned: [list or "none"]
- Transitions out of range: [list ms values, or "none"]

| Dim | Score | Observation | Action needed |
|---|---|---|---|
| V1 Typographic hierarchy | N/5 | [specific observation - reference fontSizeCount from computed data] | [fix or "none"] |
| V2 Spatial rhythm | N/5 | [specific observation - reference paddingMisaligned from computed data] | [fix or "none"] |
| V3 Visual focal point | N/5 | [specific observation] | [fix or "none"] |
| V4 Colour discipline | N/5 | [specific observation - note if brand accent on row buttons] | [fix or "none"] |
| V5 Information density | N/5 | [specific observation] | [fix or "none"] |
| V6 Dark-mode polish | N/5 | [specific observation from dark screenshot - reference score anchors; note OKLCH hue shifts, washed-out tones, badge legibility] | [fix or "none"] |
| V7 Micro-polish | N/5 | [specific observation - reference transitionsOutOfRange; note empty state CTA quality from Step 5a] | [fix or "none"] |
| V9 Gestalt compliance | N/5 | [proximity grouping, figure/ground distinction, similarity across sections, continuity of lists/grids - specific observation] | [fix or "none"] |
| V10 Typographic quality | N/5 | [line-height ratio, letter-spacing on small fonts, line length on text/td - reference computed data if available] | [fix or "none"] |
| V11 Interaction state design | N/5 | [hover/focus/active/disabled/loading states - reference code inspection from Step 5a] | [fix or "none"] |

**Page score**: [sum]/50 - [label: Excellent ≥40 | Good 30-39 | Needs work 20-29 | Poor <20]
**Critical findings on this page**: [list or "none"]

Step 7 - Cross-page pattern analysis

After all pages, identify:

1. Recurring weaknesses: dimensions that score ≤ 3 on ≥ 3 pages → systemic issue
2. Best-in-class pages: the 2 pages with highest scores → what patterns make them work?
3. Worst offenders: the 2 pages with lowest scores → highest ROI for improvement
4. Theme gap: average V6 score vs average of V1–V5 → if V6 is ≥ 1 point lower, dark mode needs dedicated attention
5. Typography discipline: pages where fontSizeCount > 5 → systemic type scale violation; pages where V10 scores ≤ 3 → line-height or line-length issues
6. Gestalt patterns: pages where V9 scores ≤ 3 → proximity or similarity violations across sections
7. Interaction debt: pages where V11 scores ≤ 3 → missing hover/focus/loading states (cross-reference with V7)
8. Code pattern violations: list any systematic code patterns observed across pages that cause visual problems (e.g., "brand accent on row buttons appears on 4 pages: [list]")

Step 8 - Report

## Visual Audit - [DATE] - [MODE] - [TARGET]
### Reference: [SITEMAP_OR_ROUTE_LIST] · [GLOBAL_STYLES_FILE] · [THEME_STYLES_FILE]
### Scope: aesthetic quality · typography · spacing · visual hierarchy · colour · dark mode · polish
### Out of scope: token compliance → /ui-audit | Accessibility + contrast → /accessibility-audit | UX flows → /ux-audit | Responsiveness → /responsive-audit

---

### Overall score

| Dimension | Avg score | Trend |
|---|---|---|
| V1 Typographic hierarchy | N.N/5 | ↑/→/↓ |
| V2 Spatial rhythm | N.N/5 | ↑/→/↓ |
| V3 Visual focal point | N.N/5 | ↑/→/↓ |
| V4 Colour discipline | N.N/5 | ↑/→/↓ |
| V5 Information density | N.N/5 | ↑/→/↓ |
| V6 Dark-mode polish | N.N/5 | ↑/→/↓ |
| V7 Micro-polish | N.N/5 | ↑/→/↓ |
| V9 Gestalt compliance | N.N/5 | ↑/→/↓ |
| V10 Typographic quality | N.N/5 | ↑/→/↓ |
| V11 Interaction state design | N.N/5 | ↑/→/↓ |
| **Total** | **N.N/5** | |

---

### Per-page results

| Page | Route | Score | Weakest dim | Critical? |
|---|---|---|---|---|
| P01 Login | /login | N/50 | V[N] | yes/no |
...

---

### [Per-page analysis blocks from Step 6]

---

### Systemic issues (recurring across ≥ 3 pages)

| Issue | Affected pages | Dimension | Code cause | Suggested global fix |
|---|---|---|---|---|
| [e.g. H1 and subtitle same visual weight] | P02, P04, P07 | V1 | [class observed in code] | [fix] |

---

### Systematic code violations (patterns found in code across multiple pages)

| Violation | Pages | Design rule | Fix |
|---|---|---|---|
| [e.g. brand accent on repetitive buttons] | [list] | Repetitive action buttons use neutral/secondary style per design system | Replace with design system's secondary button variant |

---

### Roster gaps (routes in sitemap not in standard roster)
[list new routes flagged in Step 4 roster check, or "None - roster current"]

---

### Priority improvement list

Critical first, then by impact (pages affected × dimension weight):

1. **[CRITICAL]** [page] - [dimension] - [concrete fix]
2. **[MAJOR]** [page] - [dimension] - [concrete fix]
...

---

### Best-in-class patterns (preserve these)

1. [Page]: [what works and why - pattern to replicate elsewhere]
2. [Page]: [what works and why]

---

### Worst offenders (highest ROI)

1. [Page]: score N/50 - [top 3 fixes that would most improve it]
2. [Page]: score N/50 - [top 3 fixes]

Step 9 - Improvement offer

After the report:

"Do you want me to go deeper or generate concrete visual proposals? I can:

>

- Critique mode: /visual-audit critique target:page:<route> - deep-dive on a single page: Gestalt + hierarchy diagnosis + concrete redesign proposals - Standalone mockup: generate an improved standalone HTML alternative faithful to the project's design tokens - Targeted fix: apply improvements that don't require design decisions (e.g., padding spacing, font weights, missing tokens) - Dark mode pass: focus exclusively on improving the dark theme across all pages - Gestalt pass: in-depth V9 analysis on all pages with concrete fixes for proximity and similarity violations

For contrast verification and WCAG coverage, run /accessibility-audit separately - it owns both APCA measurement and the full WCAG 2.2 scan."

Do NOT apply visual changes without confirmation. Many of the fixes touch spacing and typography - they affect multiple components and require design decision, not just code.

Step 10 - Screenshot cleanup

After the report is delivered and the improvement offer is presented, clean up the temp directory:

rm -rf "${SCREENSHOT_TEMP_DIR:-/tmp/visual-audit-screenshots}"

Run this unconditionally at session end - screenshots are only needed during analysis and must not persist. Override the default path via [SCREENSHOT_TEMP_DIR] in Stack adaptation.

Notes for interpretation

See ${CLAUDE_SKILL_DIR}/DIMENSIONS.md → Interpretation notes for full guidance on V4 brand colour overuse, preflight failure thresholds, and score denominator (/50). Score bucket labels: Excellent ≥40 | Good 30-39 | Needs work 20-29 | Poor <20.

Stack adaptation (fill after scaffold)

<!-- Fill these values for your stack, then remove this comment block. --> <!-- Values left as placeholders will trigger warnings in tierward doctor. -->