Web Accessibility (WCAG 2 AA))

Apply WCAG 2 Level AA accessibility standards to design systems, UI components, and web pages — contrast math with a runnable validator, design-token patterns that make color accessible by construction, typography and focus-state rules, and a semantics checklist. Use when building or reviewing UI, choosing colors or type for a brand/design system, setting up design tokens, wiring accessibility checks into a build, or when the user mentions accessibility, a11y, WCAG, contrast, or screen readers.

Make interfaces conform to WCAG 2 Level AA by building accessibility into the design system, not auditing it in afterwards. A color, text style, or component that enters the system already passing AA scales to every page that uses it; one that doesn't becomes a bug on every page at once.

Conformance target: WCAG 2.2 Level AA (quick reference).

The order of operations

1. Tokens first — validate the palette as fg/bg pairs before any component uses it.
2. Components second — states, focus, target size, semantics per component.
3. Pages last — structure, headings, alt text, keyboard walk-through.

Fixing a failing token fixes every component; fixing a failing page fixes one page. Work top-down.

1. Color contrast is math, not judgment

Never eyeball contrast. The thresholds (WCAG 2.2):

What renders	Minimum	Success criterion
Normal text (below 24px / below 18.66px bold)	4.5:1	1.4.3
Large text (≥24px, or ≥18.66px bold)	3:1	1.4.3
Meaningful non-text (icons, chart marks, input borders, focus indicators)	3:1	1.4.11
Disabled controls, decorative elements, logos	exempt	—

Run the bundled validator (no dependencies, Node ≥14):

bash

# ad-hoc pairs: fg,bg,minimum,label
node scripts/check-contrast.mjs --pair "#0f766e,#ffffff,4.5,link on white"

# a whole palette from a JSON config (see script header for the format)
node scripts/check-contrast.mjs contrast-pairs.json

Wire it into the build. A contrast check that runs once is an audit; one that runs in npm run build or CI and exits non-zero is a guarantee. Add every token pair that renders as text or meaningful graphics, in every theme.

2. Token patterns that make color accessible by construction

Three patterns cover most real-world palette failures:

• Split fill roles from text roles. Vibrant brand colors (a teal, an orange, a mid-blue) usually pass as a fill behind dark text but fail as text on a light background — often by a factor of two. Define both and never let components cross them:

  css

  --color-primary:      #14b8a6;  /* fills, borders, focus rings — never text on light */
  --color-primary-text: #0f766e;  /* the same hue, dark enough to be text (≥4.5:1)   */
  

  Filled buttons then choose text per theme: dark text on a bright fill, or light text on a dark fill — whichever pair passes, verified by the script, not by taste.

• Theme-tune semantic colors. Success/warning/error values that read well on near-black (#f59e0b amber, #ef4444 red) fail on white — amber is ~2.2:1 there. Map --color-success/warning/error per theme instead of sharing one set; keep the bright ramp for dark themes and a darker ramp (e.g. #b45309, #dc2626) for light.

• Subtle text has a floor. Placeholder, hint, and caption text is still text — 4.5:1 applies. Mid-grays like #9ca3af on white (~2.5:1) are the single most common violation on the web. The lightest AA gray on white is around #767676; pick your "subtle" token at or below that lightness.

3. Typography that scales

• Relative units (rem/em) for font sizes, line heights, and containers that hold text — browser zoom and text-resize must work to 200% without loss of content (SC 1.4.4).
• Body text ≥ 1rem (16px); don't fight user font-size settings with html { font-size: 62.5% } tricks that bake in pixel assumptions.
• Line height ≥ 1.5 for body copy; paragraph spacing larger than line spacing (SC 1.4.12 expects content to survive these being user-forced).
• Don't convey meaning by color alone (SC 1.4.1): pair color with an icon, weight, underline, or text label — links inside prose need more than a hue change.

4. Interactive states

• :focus-visible on every interactive element — a 2px+ outline with 3:1 contrast against the surrounding surface (SC 2.4.7, 2.4.13). Never outline: none without a replacement.
• Hover/active recolors must also pass. A link that darkens on hover passes twice; one that brightens toward the background fails exactly when the user is looking at it. Add state colors to the contrast config too.
• Target size ≥ 24×24 CSS px (SC 2.5.8) — buttons, nav links, icon buttons. 44px stays the comfortable default for primary controls.
• Respect prefers-reduced-motion — gate non-essential animation behind the media query.

5. Semantics checklist (the part tokens can't fix)

Run through this per component/page; details and fix patterns in references/wcag-checklist.md:

• Images: alt text that says what the image does here; alt="" for decorative (1.1.1)
• One <h1>, headings in order, no skipping levels for styling (1.3.1)
• Real elements: <button>, <a href>, <label for> — not styled <div>s (4.1.2)
• Every input labelled; errors identified in text, not color alone (3.3.1, 3.3.2)
• Full keyboard walk-through: reach, operate, and leave everything; logical tab order; skip link on page-level nav (2.1.1, 2.1.2, 2.4.1)
• Current page/state marked (aria-current="page", aria-expanded, …)
• Page lang attribute and a descriptive <title> (3.1.1, 2.4.2)

Honest boundaries

• Automated checks (this script, axe, Lighthouse) catch roughly a third to half of WCAG failures — the mechanical ones. Keyboard walk-throughs and a screen-reader pass (VoiceOver/NVDA) are still required for real conformance. Say "the token layer is AA-verified", not "the product is accessible", until both have happened.
• A design system guarantees the visual layer: contrast, focus visibility, scalable type, target sizes. Content semantics — alt text, heading order, labels, keyboard flows — live in each product and must be checked there.
• Contrast ratios apply to the colors as rendered — opacity, gradients, and text over images need checking at their worst point, not their average.