ux-guardrails

ux-guardrails

Prose design rules are followed probabilistically and dropped under iteration pressure; a deterministic check is not. This skill ships that check — scripts/ux_lint.mjs, a single-file, zero-dependency Node lint — plus the hook wiring that runs it on every UI write and refuses to finish a session with critical violations outstanding.

When to use / when NOT to use

Use to install or run mechanical UX enforcement in a landing-page project: spacing on the grid, arbitrary-value and raw-palette bans, WCAG contrast on design-token pairs, focus-visibility, slop tells (purple gradients, gradient text, glassmorphism, transition: all).

Not for: choosing the palette/type/tokens (visual-design-system), picking the overall look (landing-page-art-direction), a full WCAG 2.2 audit with manual passes (landing-page-accessibility), motion design (scroll-motion), or generic test/format hooks (that's ordinary repo tooling).

Severity model (two tiers, on purpose)

• critical — objectively wrong, blocks: disabled zoom, <img> without alt, outline-none with no focus-visible replacement, positive tabindex, token pairs below 4.5:1. The agent must fix these.
• advisory — drift and slop tells, informs but never blocks: off-scale spacing, arbitrary text sizes, raw palette classes, gradient text, slow transitions. Per-edit nagging on taste rules has been observed to make agents conservative and samey — so taste stays advisory, and the brief wins: an intentional exception is suppressed explicitly, never fought.

Escape hatches (use these; never weaken the script): a ux-guardrails-ignore <rule-id> comment on or above the line, or ignoreRules/ignoreFiles in .ux-guardrails.json.

Workflow

1. Install into the consumer project. Copy the script and wire both hooks:

   mkdir -p .claude/hooks
   cp "${CLAUDE_SKILL_DIR}/scripts/ux_lint.mjs" .claude/hooks/ux_lint.mjs
   node .claude/hooks/ux_lint.mjs --self-test   # must print SELF-TEST PASSED
   

   Merge into .claude/settings.json (merge — do not clobber existing hooks):

   {
     "hooks": {
       "PostToolUse": [
         { "matcher": "Edit|Write|MultiEdit",
           "hooks": [{ "type": "command",
             "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/ux_lint.mjs\"",
             "timeout": 10 }] }
       ],
       "Stop": [
         { "hooks": [{ "type": "command",
             "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/ux_lint.mjs\"",
             "timeout": 30 }] }
       ]
     }
   }
   

2. Configure only if the project's scale differs. Optional .ux-guardrails.json at the project root:

   { "spacingStep": 4, "contrastMin": 4.5, "maxFontFamilies": 3,
     "transitionMaxMs": 500, "ignoreRules": [], "ignoreFiles": [] }
   

   Defaults match this repo's visual-design-system budgets. WCAG floors are the one thing config should never be used to relax.

3. Run it in CI too (same script, exit 1 on criticals):

   node .claude/hooks/ux_lint.mjs src/ app/ styles/
   

4. Token contrast needs the pairing convention. Contrast is computed only for CSS custom properties following the shadcn --x / --x-foreground convention (plus --background/--foreground), with resolvable literal values (hex, rgb, hsl, oklch — no var() chains, color-mix, or alpha). Anything unresolvable is skipped silently by design — a lint that false-positives on values it can't parse gets uninstalled.

How the hooks behave (what to tell the user)

• PostToolUse lints only the file just written. Criticals → exit 2, stderr becomes the agent's repair instruction (the write already landed — PostToolUse steers the next step, it cannot undo). Advisories → injected as context, never block.
• Stop lints all git-changed UI files (capped at 20). Criticals → exit 2 blocks finishing, once — it honors stop_hook_active, so a stuck false positive can never loop the session.
• Cross-file budgets (font-family count) run only at Stop/CLI where the whole changeset is visible; per-file they'd be meaningless.

Output spec

Findings as file:line [SEVERITY] [rule-id] message, grouped by file, ending with a N critical, M advisory budget line and the ignore instruction. The full rule catalog with thresholds and rationale: references/rules.md.

Gotchas

• Never "fix" a finding by editing the lint script, the hook config, or by scattering ignore comments. An intentional design exception gets one documented ignore; three ignores of the same rule mean the config (or the design) is wrong.
• The Stop gate reads git diff — in a project without git it silently checks nothing (PostToolUse still covers every write).
• --self-test exercises every rule against embedded fixtures; run it after any update of the script copy.
• Advisory findings are signals, not noise: resolve each one by either fixing the code or ignoring it explicitly with a reason.
• The purple-gradient / gradient-text tells are heuristics about defaults — a brand that genuinely wants purple overrides them (ignore + a line in DESIGN.md), which is exactly the "brief wins" rule.

References

• references/rules.md — every rule id, severity, exact trigger/threshold, and the research it's derived from.