Contrast + Uniformity Sweep

Audit rendered UI across explicit breakpoints and theme modes to find legibility and visual-system consistency failures.

This skill is diagnostic only: identify, evidence, classify, and propose fix direction. Do not redesign or patch code during the sweep unless explicitly requested.

Relationship to Other Skills

• tools-web-breakpoint: responsive containment/reflow failures across breakpoints. Use together when contrast failures appear only after responsive stacking.
• lp-design-qa: static code/token/a11y audit without runtime screenshots. Use to trace root causes for repeated uniformity drift.
• meta-user-test: broad site-wide health audit. Use when route coverage should be sitemap-driven rather than operator-selected.
• lp-do-build: implement fixes after this report is approved.

Required First Prompt

If missing from operator input, ask exactly:

Which breakpoint widths (px) and which theme modes (light/dark/brand themes) should I test for contrast + uniformity?

Then collect required surfaces:

Which UI surfaces should be in scope (header/nav, primary CTAs, forms, modals, tables, marketing pages)?

Inputs

If operator gives only breakpoints:

• default to WCAG 2.x AA
• run light mode first
• record assumptions explicitly in the report

Standards and Thresholds

Use these defaults unless operator requests otherwise.

WCAG AA defaults

• Normal text (<18pt and <14pt bold): >= 4.5:1
• Large text (>=18pt or >=14pt bold): >= 3.0:1
• Non-text essential UI (icons, borders, controls): >= 3.0:1
• Focus indicators: >= 3.0:1 against adjacent colors and visibly discernible

If operator requests AAA

• Normal text: >= 7.0:1
• Large text: >= 4.5:1
• Non-text/focus checks remain at WCAG non-text contrast criteria (>= 3.0:1)

Sampling requirement for complex backgrounds

For gradients/images/translucent overlays/blurred surfaces, sample multiple points and report:

• worst-case ratio (required)
• average ratio (recommended)

Repo Context to Load

Before auditing, load the visual-system baseline:

• docs/design-system-handbook.md
• docs/typography-and-color.md
• packages/themes/<theme>/src/tokens.ts (or packages/themes/base/src/tokens.ts)
• packages/design-system/src/utils/style/overflowContainment.ts

For token drift triage, run static scans in affected app paths (examples):

• non-token colors: rg -n "#[0-9a-fA-F]{3,8}|rgb\\(|hsl\\(" apps packages
• default Tailwind palette drift: rg -n "\\b(bg|text|border)-(red|orange|amber|yellow|lime|green|emerald|teal|cyan|sky|blue|indigo|violet|purple|fuchsia|pink|rose)-(50|100|200|300|400|500|600|700|800|900)\\b" apps packages
• arbitrary sizing/spacing drift (triage only): rg -n "\\[[0-9.]+(px|rem|em|vh|vw|%)\\]" apps packages

Workflow

1) Intake and Matrix Setup

1. Confirm breakpoints, modes, surfaces, and standard (AA/AAA).
2. Resolve route list from operator surfaces (or infer best-effort routes and mark assumptions).
3. Set stable viewport height (default 900).
4. Prepare artifact folder and screenshot naming convention.
5. Detect JS theme toggle — scan page source for localStorage + data-theme (or equivalent cookie/class-based toggle). If found, expand the test matrix (see step 2a below).

2) Breakpoint Sweep Execution

For each breakpoint W and each mode:

1. Set viewport to W x 900.
2. Visit each target route/surface.
3. Capture baseline screenshot.
4. Run contrast checks and uniformity checks.
5. Trigger interaction states where feasible:
   • hover / active
   • disabled
   • focus-visible via keyboard tab flow
   • modal/drawer/popover open states
   • form error states (if low-friction to trigger)

2a) Split-State Testing (mandatory when JS theme toggle detected)

A JS toggle (localStorage, cookie, or data-theme attribute) and prefers-color-scheme are independent systems. They can disagree — e.g. OS in light mode but toggle previously set to dark. CSS variables controlled only by @media(prefers-color-scheme) are invisible to the toggle, causing broken mixed states that emulateMedia alone can never reproduce.

Always test all four combinations when a toggle exists:

How to inject localStorage state in Playwright:

await page.goto(url, { waitUntil: 'load' });
await page.evaluate(() => localStorage.setItem('sl-theme', 'dark')); // or whatever key
await page.reload({ waitUntil: 'load' });
// confirm: await page.evaluate(() => document.documentElement.getAttribute('data-theme'))

What to check in split states:

• CSS custom properties that appear in @media(prefers-color-scheme) but NOT in html[data-theme] — these will hold stale values in the mismatched state
• Text colour vars (--text, --text-muted) that come from data-theme paired against background/surface vars that come from @media only (or vice versa)
• Any component that hardcodes color:#fff or color:#000 while its background var is theme-switched

Flag as S1 if any split state produces unreadable text. The root cause is always: CSS variables split across two theming systems that don't fully overlap.

Recommended fix pattern: Make html[data-theme="dark/light"] the single authoritative source for ALL CSS vars. Keep @media(prefers-color-scheme) only as a FOUC fallback (it fires before JS runs). Since html[data-theme] has higher specificity than :root inside @media, it will always win once JS sets it.

3) Contrast Checks (priority order)

Audit:

• primary text blocks (body/headings/captions/helper text)
• interactive text (links, button labels, nav items)
• form affordances (placeholder, borders, error/success text)
• icons and badges (especially icon-only controls)
• focus rings/outlines during keyboard navigation
• overlay scenarios (scrims, translucency, image cards, gradients)

Flag contrast findings when:

• measured ratio is below threshold
• normal state is borderline and fails on hover/focus/disabled/error
• worst-case sample on gradient/image fails
• placeholder substitutes label with low contrast
• focus indicator exists but is visually indistinct

Evidence required per contrast finding:

• breakpoint + mode + route
• element label/selector/component best-effort
• state (default|hover|focus|disabled|error|active)
• measured ratio(s) with threshold (worst-case mandatory)
• screenshot with element clearly visible

4) Uniformity Checks (priority order)

Audit:

• color/token compliance and repeated non-token drift
• component consistency (buttons/links/inputs across routes)
• typography consistency (size/weight/line-height hierarchy)
• spacing/radius/shadow rhythm consistency
• dark-mode parity for hierarchy/affordances

Flag uniformity findings when:

• same component family differs materially in computed styles
• repeated non-token colors or inconsistent opacity values appear
• disabled/focus states differ across equivalent components
• same semantic meaning uses conflicting visual treatments
• hierarchy weakens (e.g., heading visually weaker than body)

Evidence required per uniformity finding:

• breakpoint + mode + route
• component family (best effort)
• exact observed difference (what differs and where)
• screenshot pair (reference vs drift)
• likely cause hypothesis (token bypass/local override/missing variant)

5) Severity Model

• S1 Blocker: text/controls not reliably readable; critical actions obscured; focus effectively invisible for keyboard users
• S2 Major: widespread AA/AAA failures on key UI or system inconsistency harming trust/comprehension
• S3 Minor: localized drift/cosmetic inconsistency without material task failure

6) Responsive-vs-Palette Attribution Rule

If a contrast failure appears only at specific breakpoints due to layout/image stacking, classify as responsive root cause (with contrast impact), not pure palette failure. Cross-reference tools-web-breakpoint for follow-up.

7) Output Artifacts

Write report + screenshots:

• Report: docs/audits/contrast-sweeps/YYYY-MM-DD-<slug>/contrast-uniformity-report.md
• Screenshots: docs/audits/contrast-sweeps/YYYY-MM-DD-<slug>/screenshots/

Optional machine-readable artifacts (recommended for regression tracking):

• docs/audits/contrast-sweeps/YYYY-MM-DD-<slug>/contrast-findings.json
• docs/audits/contrast-sweeps/YYYY-MM-DD-<slug>/uniformity-findings.json

Use .claude/skills/tools-ui-contrast-sweep/modules/report-template.md for required structure.

8) Completion Message

Return:

• breakpoints and modes tested
• surfaces/routes covered
• issue totals by severity
• report path + JSON artifact paths (if produced)
• assumptions and uncovered scope gaps

If no issues are found, state exactly:

No contrast or visual-uniformity failures detected across the tested breakpoint/mode matrix.

• If issues were found and fixed via /lp-do-build: re-run this sweep to confirm findings are resolved before routing to tools-refactor.
• If the page has a JS theme toggle: re-run must include all four split-state combinations, not just clean emulateMedia states. A fix that only passes clean states may still fail in the mismatched case.

Guardrails

• Do not propose color changes that break brand intent without noting tradeoffs.
• Prefer token-level/component-variant fixes over per-page overrides.
• Validate focus-visible states explicitly via keyboard path.
• Avoid reporting micro-differences unless they affect readability, interaction, or hierarchy.

Integration

• Upstream: lp-design-qa (optional trigger — contrast-sweep is often invoked after lp-design-qa flags color or accessibility concerns); lp-do-build (direct invocation for pre-launch QA pass).
• Downstream: tools-refactor (contrast and token findings feed the refactor entry criteria); lp-do-build (issues returned as structured findings for fix tasks).
• Loop position: S9B secondary skill, required before S9B→SIGNALS advance (GATE-UI-SWEEP-01). Also runs alongside tools-ui-breakpoint-sweep at S9C for ongoing quality passes.
• S9B advance requirement: When running this sweep for S9B→SIGNALS advance, the operator must manually set Business: <BIZ> in the report frontmatter (the template includes a placeholder — replace it with the actual business identifier). All application routes must be covered; Routes-Tested: 0 will block the advance gate. Both light and dark modes must be tested. All S1 blockers must be resolved before advancing.