Version

Current release: v2.5.1 · See the Changelog · All releases

No build tools, dependencies, or runtime required — this is a pure instruction & knowledge layer for AI agents.

What It Does

Quick Start

Option A — Install with npx (recommended)

Drop the kit into any project, no clone needed:

npx ux-ui-agent-skills init          # full kit into the current folder
npx ux-ui-agent-skills add tokens taste design-systems   # just some areas
npx ux-ui-agent-skills list          # see all areas

Flags: --force (overwrite existing files) · --dry (preview, change nothing).

Option B — Clone

git clone https://github.com/plugin87/ux-ui-agent-skills.git
cp -r ux-ui-agent-skills/ your-project/

Then start using — open the project in Claude Code or any Claude-powered IDE. CLAUDE.md loads automatically, activating the agent persona with full access to every tokens / components / taste / design-system file and the runnable /skills.

"Design a notification component with all states and accessibility"
"Review this login page against WCAG 2.2 and Nielsen's heuristics"
"Generate React + Tailwind code for a data table with sorting and pagination"
"Create a color token palette for a fintech brand using blue as the primary"
"Audit this form for accessibility issues — give me a prioritized findings table"
"Write the empty state and error copy for the onboarding flow"
"Spec the motion for the modal open/close with reduced-motion fallback"

Start a New Design Project

The kit is the engine. A product repo that uses it wants its own lean layout, and templates/product-design/ is that starter, shipped ready to copy:

your-product/
  CLAUDE.md              the brief Claude reads every session (lean, placeholders to fill)
  CLAUDE.local.md        your personal prefs, gitignored
  .mcp.json              Figma / Notion connections, env-expanded, no secrets
  .claude/
    rules/               components.md · tokens.md · accessibility.md — load only when relevant
    skills/              repeatable workflows your team adds
    commands/            /gate — the project's own all-or-nothing check
    settings.json        shared permissions, checked into git
  design-tokens.json     source of truth: color, type, spacing (light + dark, WCAG-verified)
  src/components/        the real UI Claude reads and edits
  public/images/         real images so prototypes do not break
  reference/             real screens Claude studies for context

In Claude Code, from a clone of this repo:

/scaffold-project ../your-product

It copies the template, installs the engine areas next to it (npx ux-ui-agent-skills add tokens components taste accessibility workflows content frameworks design-systems scripts skills), and walks the placeholders with you. The project brief stays short on purpose: it loads on every turn, so everything that is not needed every turn lives in .claude/rules/ or a skill.

The seeded theme is not a guess. python3 scripts/validate_template.py proves the layout is complete, every alias resolves, and the required contrast pairs pass WCAG 2.2 AA in both light and dark before a project starts from it.

How to Use

New here? Read docs/WORKFLOW.md for the full end-to-end picture — how the Request Router loads layers on demand, real usage scenarios, and the automated release pipeline.

There are three ways to drive the kit. Use whichever fits the moment.

1. Just ask (zero commands)

CLAUDE.md loads automatically, so plain requests already route to the right knowledge. Describe what you want and the agent self-routes via the built-in Request Router:

"Generate a Svelte button with all states and dark mode"
"Make this landing page feel like Linear"
"Migrate our Material 3 colors into this token system"

2. Run a skill explicitly (/skill)

Type a slash command to invoke a capability directly. Each skill loads only the files it needs and can run its own scripts.

Each SKILL.md carries an invocation field, so it is clear which ones you drive and which ones the agent reaches for on its own. Both kinds can be typed as a slash command.

User-invoked — you start them, they orchestrate a whole job

Model-invoked — the discipline the agent applies while it works

Three slash commands round it out: /gate runs the whole gate and reports the real N/N, /ship adds the release checklist, and /scaffold-project starts a new product repo from the template.

/design-code  a pricing card in Vue, dark-mode aware
/apply-aesthetic  stripe   →  make the dashboard feel like Stripe
/a11y-audit  this checkout form
/migrate-design-system  from Material 3 to our tokens

3. Run the scripts (real, no dependencies)

Plain python3 — useful in the terminal or CI:

python3 scripts/validate_tokens.py                 # validate token JSON + alias refs
python3 scripts/validate_contrast.py               # batch WCAG gate: token pairs, light + dark
python3 scripts/contrast.py "#1d1d1f" "#ffffff"    # WCAG contrast ratio for one pair
python3 scripts/validate_component_spec.py         # every component spec is complete
python3 scripts/lint_hardcodes.py src/             # no off-theme hex/px/timing (consistency)
python3 scripts/lint_taste.py page.html            # heuristic anti-slop taste check
python3 scripts/design_systems.py list             # browse the 138-system library
python3 scripts/scaffold_component.py "Date Picker" # emit a component spec stub
python3 scripts/validate_template.py               # the starter template stays sound
python3 scripts/validate_instruction_surface.py    # no always-on rule got demoted
node    evals/run.mjs --self-test                  # the cold-start scorer still works

These are the same gates CI runs (.github/workflows/ci.yml) — token validity, WCAG contrast in light + dark, spec completeness, and zero hardcoded values — so theme/color stays consistent across every page and accessibility is enforced, not assumed.

Typical flow

1. /apply-aesthetic linear      → set the visual direction (tokens re-pointed)
2. /design-component Combobox    → spec it with states + a11y
3. /design-code  Combobox in React + Tailwind   → production code
4. /a11y-audit                   → verify contrast, keyboard, focus
5. /design-review                → score + findings before ship

Tip: skills compose. apply-aesthetic always re-verifies contrast through a11y-audit; redesign calls design-review + a11y-audit automatically.

Proving It, and Admitting What Cannot Be Proven

The kit ships 37 objective gates behind one command:

node scripts/accuracy_report.mjs     # 35/35 or it fails — no partial credit

Token validity, WCAG contrast on a real headless render in light and dark, every element in default/hover/focus, axe roles and names, focus traps, RTL, responsive at 280/320/414, target size, keyboard operability, reduced motion (including content that only an animation reveals), silent text clipping, token-by-intent, and zero emoji anywhere in the output or the instruction surface.

That is correctness. It is not quality, and the kit says so out loud:

/critique exists because a passing gate is never evidence of taste. It refuses to review from source alone, screenshots at 1280 and 390 in both themes, clicks every control, and returns a verdict with the three reasons a senior designer would send the work back.

Runs are recorded in evals/RESULTS.md with their provenance attached — who built the output and whether they could see the kit while doing it — because a run without that context is not evidence of anything.

The eval suite exists because "the kit's own examples pass" is a weaker claim than "an agent given only this kit and a brief produces work that passes". Building it caught two real defects the 34-check gate had missed. See evals/README.md.

Project Structure

.
├── CLAUDE.md                  # Agent persona, gates, router — the always-on brief (~270 lines)
├── CONTEXT.md                 # Ubiquitous language — shared domain vocabulary
├── CLAUDE.local.md            # Personal prefs (gitignored, per-machine)
├── .mcp.json                  # Project MCP servers (Figma) — no secrets, env-expanded
│
├── .claude/rules/             # Depth split out of CLAUDE.md, loaded only when relevant
│   └── tokens-and-color · typography-and-spacing · components · accessibility
│       frameworks · review-and-research · brand-and-operations
├── .claude/skills/            # Runnable skills — invoke via /name
│   └── design-tokens · design-component · design-code · design-review · a11y-audit
│       apply-aesthetic · redesign · migrate-design-system · prototype · ux-writing
├── .claude/commands/          # Custom slash commands — /gate · /ship · /scaffold-project
├── .claude/settings.json      # Shared permissions (scripts allowlist), checked into git
├── .claude/agents/            # design-critic — the adversarial reviewer behind /critique
│
├── evals/                     # Cold-start briefs + run.mjs — 14 objective gates on produced work
│
├── reference/                 # Real screens the agent studies before designing/reviewing
│
├── templates/product-design/  # Starter layout for a NEW product repo — /scaffold-project
│   ├── CLAUDE.md · CLAUDE.local.md.template · .mcp.json · design-tokens.json
│   └── .claude/{rules,skills,commands,settings.json} · src/components · public/images · reference
│
├── scripts/                   # Real helper scripts (python3, no deps)
│   ├── validate_tokens.py     # JSON + alias validation for tokens/
│   ├── contrast.py            # WCAG 2.2 contrast-ratio checker
│   ├── design_systems.py      # Browse/search the 138-system library
│   └── scaffold_component.py  # Emit a component spec stub
│
├── tokens/                    # Design tokens (DTCG format) — 13 files
│   ├── colors · typography · spacing · shadows · borders · breakpoints · motion
│   └── gradients · opacity · blur · sizing · states · theming
│
├── taste/                     # Aesthetic judgment layer
│   ├── design-taste.md        # Anti-slop doctrine, banned defaults, pre-flight check
│   ├── aesthetic-systems.md   # Archetypes + catalog of 138 design systems
│   └── motion-choreography.md # Motion grammar + reduced-motion parity
│
├── design-systems/            # Interop + brand library
│   ├── interop-protocol.md    # Map to/from ANY design system
│   ├── crosswalk.md           # Material 3 · Apple HIG · Fluent · Carbon · shadcn · Radix
│   │                          # · Ant · Polaris · Primer · Atlassian · Bootstrap
│   └── library/<name>/        # 138 brand-grade DESIGN.md specs
│
├── content/                   # UX writing & content design
│   └── voice-tone.md          # Voice & tone, error/empty-state copy, microcopy, inclusive language
│
├── components/                # Component specs (Atomic Design) — 50 components
│   ├── atoms · molecules · organisms · templates
│   ├── navigation · feedback · forms-advanced · overlays
│   └── data-display · data-viz · icon-system
│
├── accessibility/             # WCAG & ARIA references + inclusive design
│   ├── wcag-checklist.md      # WCAG 2.2 checklist (POUR, P0/P1/P2)
│   ├── aria-patterns.md       # WAI-ARIA patterns for 19 components
│   ├── cognitive.md · vision.md · i18n-rtl.md   # cognitive · low-vision/CVD/forced-colors · RTL
│   └── wcag-aaa.md            # AAA upgrade delta
│
├── workflows/                 # Design process + ops/pipeline guides
│   ├── design-review.md · design-to-code.md · prototyping.md · redesign-audit.md
│   └── governance.md · token-build.md · figma-integration.md · design-qa.md · performance.md
│
└── frameworks/                # Implementation patterns — ANY framework
    ├── adapter-protocol.md    # Universal translation contract
    ├── react-tailwind.md · nextjs.md · swiftui.md   # full references
    └── adapters/              # vue · svelte · angular · solid · web-components-lit · qwik · astro
                               # mui · mantine · chakra · bootstrap
                               # react-native · flutter · jetpack-compose · vanilla-css · css-in-js

[!NOTE] The design-taste layer (taste/) and the 138-system library (design-systems/library/) set visual direction; the system (tokens, components, accessibility) keeps it correct. Taste serves the Aesthetics tier and never overrides accessibility. Skills under .claude/skills/ run with agent permissions — review before use.

Token Architecture

The design token system follows a 3-tier hierarchy using the DTCG standard:

┌─────────────────────┐     ┌─────────────────────┐     ┌─────────────────────┐
│  COMPONENT TOKENS   │ ──► │  SEMANTIC TOKENS    │ ──► │  PRIMITIVE TOKENS   │
│  button-bg-primary  │     │  action.primary     │     │  blue.600 = #2563EB │
│  (use in code)      │     │  (use in design)    │     │  (raw palette)      │
└─────────────────────┘     └─────────────────────┘     └─────────────────────┘

Dark mode works by swapping semantic tokens — primitives stay the same.

Supported Frameworks — any

The Framework Adapter Protocol defines a universal token→framework contract, so the agent can target a stack even with no dedicated file (it generates an adapter on demand).

Full references

Concise adapters — Vue 3 · Svelte 5 · Angular · SolidJS · Web Components (Lit) · React Native · Flutter · Jetpack Compose · vanilla CSS · CSS-in-JS (emotion/vanilla-extract/Panda)

Design-System Interop — any

Adopt, build on, or migrate between external design systems via a role-based crosswalk (interop-protocol + crosswalk). Curated tables: Material Design 3 · Apple HIG · Fluent 2 · Carbon · shadcn/ui · Radix (others derived on demand). Plus a library of 138 brand-grade design systems (apple, linear, stripe, vercel, notion, spotify, tesla…) under design-systems/library/.

Accessibility Standards

All outputs follow WCAG 2.2 Level AA as a minimum:

• Color contrast: 4.5:1 (text), 3:1 (UI components)
• Keyboard navigable with visible focus indicators
• Screen reader compatible with proper ARIA roles and live regions
• Touch targets: 24×24px minimum (WCAG 2.5.8)
• WCAG 2.2 criteria: Focus Not Obscured, Target Size, Accessible Authentication

Design Review Output

When reviewing designs, the agent scores across 6 weighted dimensions:

Findings are categorized: Critical (must fix) → Major (fix this sprint) → Minor (when convenient) → Enhancement (backlog).

Customization

This is a starter kit — make it yours:

• Brand colors — edit tokens/colors.json primitives, then update semantic references
• Typography — swap font families in tokens/typography.json and framework files
• Components — add new components following the existing spec format in components/
• Frameworks — add new framework files in frameworks/ (e.g., vue.md, flutter.md)
• Workflows — adapt review rubrics and checklists in workflows/ to your team's process

Requirements

• Claude Code CLI or any Claude-powered IDE
• A Claude model with sufficient context (Sonnet, Opus, or Haiku)

Changelog

v2.5.1

The release that checks the checkers. Every gate in this kit pointed at examples that already pass, which proves a gate says yes to good work and never that it can still say no. tests/ supplies the other half, and building it turned up four places where a gate reported green without having looked at anything.

tests/ — negative fixtures, one per gate

• npm run test:gates runs the full suite (needs Playwright and real Chrome); npm run test:unit runs the browser-free half. Runner is Node's built-in node --test — no new dependency.
• Every gate now has input built to break exactly the thing it measures: an emoji in UI, raw hex, a dangling alias, a sub-AA token pair, a floating var(), a button that only fails contrast on hover, a control with no accessible name, a 16x16 target with a crowded neighbour, a blue Delete, a 280px overflow, a toggle Enter and Space cannot operate, silently clipped text, content revealed only by an entrance animation, a dialog Tab walks out of, an aria-sort header that sorts nothing, a physical margin that will not mirror under RTL, the hardcoded indigo gradient, and a 1.5x type scale. Each must exit 1. tests/fixtures/good/clean-panel.html must survive all sixteen render gates in light and dark.
• A pass is not accepted on trust: rejects() requires exit status exactly 1 and refuses three near-misses that would otherwise read as a detection — a usage message (nine gates print one and exit 0 when handed nothing, so a typo'd path would look clean), a SKIPPED line, and an exit 1 whose output does not match the signal under test.
• The suite itself was verified by mutation, not assertion: check_no_emoji made to stop reporting its hits, validate_contrast made to stop failing a bad pair, and verify_states made to always exit 0 — each turned the suite red.
• tests/README.md records what is still not covered rather than implying it is: five scripts hardwire the repo root and ignore argv, verify_keyboard.mjs filters on tabbability before auditing so a <div role="button"> with no tabindex is invisible to it, build_tokens.mjs has no failure mode at all, and taste remains unmeasurable.

Four false greens, found and fixed

• A missing browser no longer reads as a pass. Fourteen of the fifteen render gates printed SKIPPED and exited 0 when Playwright could not be imported — on a clean CI runner or a fresh install, accuracy_report.mjs would have reported 100% with fourteen gates having measured nothing. DS_REQUIRE_BROWSER=1 turns that into exit 1, and accuracy_report.mjs, evals/run.mjs and CI's render job all set it. Without the flag the gates stay skippable for local convenience.
• validate_contrast.py passed a token file containing no tokens. A required pair whose tokens could not be resolved printed ? and was skipped, so an empty or differently-shaped design-tokens.json produced "OK: all required contrast pairs pass". Its own docstring always promised that a missing token fails; now it does.
• check_no_emoji.py and lint_hardcodes.py returned OK for a path that does not exist. Scanning zero files read as clean, so a renamed directory in CI would have passed silently. Both now error on a missing path, and on a path that exists but holds nothing scannable.
• Gate counts in prose had drifted and nothing checked them. ci.yml said "12 objective gates" where the array had 14, and the README said "thirteen" and "34-check gate". tests/meta/registry.test.mjs now counts the real arrays and holds every claim in CI, the README and accuracy_report.mjs to them.

v2.5.0

The release where the kit stopped taking its own word for anything. Enforcement went from 25 to 37 objective checks, the always-on brief was cut in half, and the two things a gate genuinely cannot do — judge taste, and prove the kit works from a cold start — got real machinery instead of a disclaimer.

Heads-up if you vendored CLAUDE.md

• CLAUDE.md is now ~276 lines, not 576. Depth moved to .claude/rules/ (7 files) and loads only when the work calls for it. Headings are unchanged, so pointers into them still resolve, but anyone who vendored the old single-file brief should re-copy: npx ux-ui-agent-skills add claude rules. Always-on and non-negotiable: the emoji ban, the gate protocol, token-by-intent, one-theme, the 8-state table, output completeness. validate_instruction_surface.py fails the build if one of them is ever demoted, if a rule file is orphaned, or if the brief regrows past its budget.
• init installs a new area (rules), and the package now ships templates/, .claude/rules/, .claude/commands/.

Enforcement: 25 -> 37 checks, still all-or-nothing

• Five render-based gates for rules the kit preached and nothing checked: verify_target_size.mjs (WCAG 2.5.8 with the spec's real spacing / inline / label-hit-area exceptions), verify_reduced_motion.mjs (policy present, motion stopped, and no content lost — catches content only an entrance animation reveals), verify_keyboard.mjs (WCAG 2.1.1, ARIA-aware: roving tabindex and aria-activedescendant widgets judged by orphan-widget and dead-arrow signals, not by Tab), lint_intent.mjs (token by intent, measured on the render: a destructive action wearing action.primary fails), verify_overflow.mjs (silently clipped text and overlapping controls, with screen-reader-only text correctly exempt). slop_tells.mjs became a hard gate.
• Each was proven to FAIL on a deliberate violation before it was trusted, and together they found real bugs the previous 25 checks passed: a dead prefers-reduced-motion rule lost to CSS specificity, a harness with no motion policy at all, and three composite widgets that declared a roving-tabindex model with zero arrow-key handlers.
• Responsive is proven against font metrics, not one machine. verify_responsive.mjs --scale renders the same narrow widths under a larger root font: the proxy for another platform's wider fallback font (Linux Chromium measured a label 2px wider than macOS Chrome and broke a 280px layout that looked clean locally) and for a user with larger text. Every example now holds at 280px at 1.25x, and that run is a gate. Root causes worth knowing: an <input> keeps an intrinsic ~20-character width that sizes its grid column; a grid item keeps min-width:auto and can widen its own track; a white-space:nowrap tooltip has no upper bound; a <table> without table-layout:fixed is sized by its cells' min-content.
• New harness edge-cases.html renders what production actually contains: unbroken 60-character strings, empty and single-item collections, missing values, ten-digit counts, forty rows.

Judgement, where measurement ends

• /critique and the design-critic subagent. Its stance is adversarial by design: the work is mediocre until the render proves otherwise, a passing gate is never evidence of taste, and every finding must cite evidence — a file and line, a measured number, or something specific in a specific screenshot. It refuses to review from source alone, screenshots at 1280 and 390 in light and dark, clicks every control, and returns a verdict plus the three reasons a senior designer would send the work back.
• The critique was run on the kit's own examples, and it returned reject. What it found is fixed here, not filed: a data-table header that declared aria-sort, drew a chevron and sorted nothing; a datepicker day drawn as selected that never moved; an Export button styled like a live action with no handler; and an "Email notifications" control that was a role="switch" in name only, identical in both states, ignoring the kit's own Toggle spec. The reference app now leads with one hero metric instead of four equal cards, closes with a real footer instead of empty canvas, carries an elevation scale rather than one flat shadow, and every harness gained display type at --text-4xl (the doctrine's own 2.5x bar) and a closing note. taste_audit and slop_tells are now clean on every example, in both themes.
• Gate 37, verify_interactive.mjs, exists because of that critique. A control that declares aria-sort / aria-pressed / aria-expanded / aria-checked / aria-selected, or wears a state-bearing role, must change something on a real click - any attribute, any DOM mutation, or focus landing somewhere other than itself. data-demo-state opts a deliberate state rendering out, and has to be written by hand so the exemption is on the record. It caught the original bug, two more like it, and then an eval fixture written hours after the critique that named the failure mode.
• taste_audit measures a real character now. Its line-length check assumed 1ch was half an em and therefore flagged max-width: 65ch - the width the kit itself recommends. It measures the element's own font instead.

Evals: does the kit survive a cold start?

• evals/ — four cold-start briefs and node evals/run.mjs <brief-id>, which points fourteen objective gates at what an agent actually produced and prints the brief's requirements for human judgement. --self-test scores the reference app with the same gates and is itself a gate, so the harness cannot rot unnoticed. Building it exposed two real defects the full accuracy gate had missed: the reference app overflowed at 280px, and verify_overflow was flagging screen-reader-only text as silently clipped.

Starting a real product with the kit

• templates/product-design/ is the recommended Claude Code design-project layout, ready to use: a lean always-on brief, .claude/{rules,skills,commands,settings.json}, a WCAG-verified design-tokens.json (light + dark), src/components/, public/images/, reference/. Scaffold with npx ux-ui-agent-skills new <dir> or /scaffold-project, and validate_template.py keeps it sound (layout complete, aliases resolve, required contrast pairs pass in both themes).
• Scripts take a path, so a product repo can gate its own single-file theme: validate_tokens.py [file|dir] (with explicit paths an unresolved alias FAILS), validate_contrast.py [file], build_tokens.mjs --in <file|dir>. Verified end to end inside a freshly scaffolded repo.
• Skills declare how they are invoked — invocation: user|model across all 17 SKILL.md, regrouped in this README: six user-invoked skills that orchestrate a job, eleven model-invoked ones that are the discipline the agent applies while it works.

v2.4.0

• Project layout aligned to the recommended Claude Code design-project structure (Phase A1 of docs/restructure-plan.md). Additive only — no knowledge folders moved, no path references changed, accuracy_report stays 25/25 = 100%.
• CONTEXT.md — ubiquitous language. A shared domain glossary (3-tier tokens, the 8 states, POUR, gate, token-by-intent, anti-slop, RENDER-AND-LOOK) so the agent names the problem precisely and spends fewer tokens doing it.
• Custom slash commands under .claude/commands/ — /gate (run the one-command gate, report the real N/N), /ship (pre-release gate + README/changelog checklist), /scaffold-project (generate a new design-product skeleton in the reference layout).
• .claude/settings.json — shared permissions (the scripts allowlist), checked into git so the team gates without per-call prompts.
• .mcp.json — project-scoped Figma MCP connection, secret-free (${FIGMA_API_KEY} env expansion).
• CLAUDE.local.md (gitignored) for personal preferences, and reference/ for real screens the agent studies before an image-to-code / redesign / design-review pass.

v2.3.1

• Zero emoji, everywhere. Purged every emoji / decorative pictograph from the entire repo — the agent instruction surface (CLAUDE.md, skills, component / workflow / content / accessibility specs, design-system library), the gate scripts' output, cover.html, docs/, and the README. Generated output was picking up emoji because the files the model reads were full of them; the instruction surface is now clean, so the model has nothing to imitate.
• No-emoji rule is now global + absolute — promoted from a narrow "no emoji as UI icons" note to a top-of-file ABSOLUTE rule covering UI, code, JSON, copy, comments, and commit messages. Replacements are lucide icons or plain words.
• The gate now guards the instruction surface too. scripts/check_no_emoji.py previously scanned only examples/ + taste/; it now also scans CLAUDE.md, the skills, and the spec directories, so emoji cannot drift back in. accuracy_report stays 25/25 = 100%.

v2.3.0

• 22 component-states harnesses — full set under examples/component-states/, each gated in light + dark: Button, Input, Modal, Tabs, Select/Combobox, Checkbox-Radio-Switch, Toast, Feedback, Navigation, Overlays, Misc, Card, Data Table, Drawer, Date Picker, File Upload, Search/Form-Field, Charts, Tree/Carousel/Image carousel/Divider, Command Palette, App Shell (header + sidebar landmarks), Context Menu. accuracy_report now 25/25 = 100%.
• Charts / data-viz — Bar, Line+area, Donut, Sparkline, Scatter (role="img" + legend), animated on entry (pathLength line-draw, scaleY bars, staggered points) with prefers-reduced-motion parity.
• lucide icon sprite — examples/component-states/icons.js injects one <symbol> sprite; icons are referenced by name <svg class="ico"><use href="#i-bell"/></svg> — no per-use path, no network, offline + gate-safe. Every harness icon converted (hand-drawn approximations that rendered as broken glyphs are gone).
• Responsive gate — scripts/verify_responsive.mjs fails on any horizontal overflow at 280 / 320 / 414 px. Fixed fixed-width / unreset-list-padding / non-wrapping-flex / minmax traps across the set.
• New theme tokens — motion (--duration-*, --ease-*, --transition-micro), color-blind-aware chart palette (--color-chart-1..6), and a dark-aware --color-surface-brand.
• Gate refinements — verify_states holds graphical / icon-only controls to 3:1 (WCAG 1.4.11, not 4.5) and exempts disabled controls; verify_focustrap no longer false-passes a display:none position:fixed dialog (caught a drawer that never closed).
• Verified patterns — thin custom checkbox/radio (real <input> under a pointer-events:none box, check + dash as two <path> in one <svg>), smooth grid-rows accordion, auto-fit (not auto-fill) card grids, equal-height panels, mobile sidebar that pushes content down instead of overlapping.
• design-component skill — added "RENDER AND LOOK — gates don't prove pixels", responsive, motion, layout, icons-by-sprite, and graphical-control rules.

v2.2.5

• Input + Modal states harnesses — examples/component-states/input.html (default/hover/focus/disabled/read-only/error/loading/filled, each with an associated <label>) and modal.html (focus trap + Escape + return-focus). Both gated by verify_states + axe + measure_render (+ verify_focustrap for the modal), light + dark. Added a verified --color-text-error token (light + dark). accuracy_report now 18/18 = 100%.

v2.2.4

• Component accuracy — verify every state. A component is "correct" only when every variant × state renders right, not just the resting default. New examples/component-states/button.html renders all variants × states (default/hover/focus/disabled/loading/selected) and is gated by verify_states + axe + measure_render (light + dark). design-component skill now mandates a states harness + running the gates. Wired into accuracy_report (now 16/16 = 100%).

v2.2.3

• Focus-trap gate — scripts/verify_focustrap.mjs opens a modal and verifies with a real keyboard that Tab stays trapped, role/aria-modal/name are present, and Escape closes + returns focus (WCAG 2.1.2 / 2.4.3). Proven to catch a leaking modal.
• RTL gate — scripts/verify_rtl.mjs renders LTR vs dir="rtl" and flags layout that overflows only when mirrored (the tell of physical left/right instead of logical properties).
• Token build (real artifact) — scripts/build_tokens.mjs (npm run build:tokens) resolves all DTCG aliases (incl. cross-file + dark) and emits a CSS-variable theme (:root + [data-theme="dark"]). 85 light + 17 dark color vars, fully resolved.
• All three wired into accuracy_report (now 15/15 = 100%) and CI.

v2.2.2

• axe-core a11y gate — scripts/axe_audit.mjs (npm run test:axe) runs axe-core (WCAG 2.0/2.1/2.2 A + AA) against rendered HTML, catching ARIA/role/label/landmark/name issues the contrast + state gates can't. Already caught a real missing-<label> bug in an example. Wired into accuracy_report (now 12 checks, 100%) and CI.
• CI render job now also runs the state-aware gate + axe across examples.

v2.2.1

• State-aware WCAG gate — scripts/verify_states.mjs (npm run test:states) measures real computed contrast of every interactive element in default / hover / focus (light + dark), catching state bugs the resting-state gate missed (e.g. a secondary button picking up the primary fill on hover). Wired into accuracy_report (now 11 checks, 100%).
• Verification Protocol — new top-of-CLAUDE.md rule: never report a quality number you didn't measure; verify all states; run the gate before declaring done; build with the gates. design-code step 13 and a11y-audit now run the render gates instead of eyeballing.
• Fixed real hover-state AA failures in examples/apple-demo; added examples/brandkit-demo (generated OKLCH foundation, light+dark, every state AA).

v2.2.0

• 2 new runnable skills (15 → 17) — image-to-code (reference image/screenshot → infer the design system → token-driven, verified code) and brandkit (brief → complete primitive→semantic→component DTCG token foundation + theme.css, light + dark, WCAG-verified). Both native, fit the kit's gates.
• Render-based taste audit — scripts/taste_audit.mjs (npm run taste) measures structural slop tells from real computed styles: timid type-scale contrast, uniform repetition, over-wide measure, palette sprawl. Wired into the design-code taste pre-flight. Heuristic by design — a strong signal, not proof (taste is subjective).
• De-emoji'd the taste doctrine itself; router rows + skills list updated.

v2.1.0

• Accuracy report — npm run verify (scripts/accuracy_report.mjs) runs every objective correctness gate as one all-or-nothing, reproducible check: token validity + alias resolution, WCAG contrast (token pairs), component-spec completeness, no hardcoded values (golden + sample), theme-ref resolution, no-emoji, and real headless-Chrome WCAG measurement (sample-app, light + dark). Prints N/N = 100% or the exact failures.
• Block-level lint exemption — scripts/lint_hardcodes.py supports ds-allow-hardcode:start / :end for justified illustration blocks (e.g. CSS product art), keeping the rest of the file strictly token-only.

v2.0.0 — Enforcement layer (breaking)

Breaking: dark-mode token values changed (link, primary action) and border.strong now meets 3:1 — re-verify any snapshots/visual tests. The kit moves from advisory guidance to enforced gates.

Theme consistency + real WCAG, enforced (driven by real-world audit feedback):

• Single-theme consistency — every page renders from ONE shared token theme; CLAUDE.md rule + examples/golden/ (theme.css with full color/type/spacing/breakpoint tokens, Button.tsx, Modal.tsx). design-code/redesign require consuming the one theme — no per-page palettes.
• Hardcode linter catches real drift — scripts/lint_hardcodes.py now flags raw hex/px/ms, raw Tailwind palette utilities (bg-gray-500, text-blue-600), and literal font-family (the 527-hardcode problem).
• No floating tokens — scripts/validate_theme_refs.py proves every var(--…) a component uses is defined in the theme (precision/consistency gate).
• Real WCAG gate — scripts/validate_contrast.py checks required text/action/border pairs in light + dark; fixed genuine dark-mode contrast bugs (link, primary action) and made border.strong meet 3:1 for essential control borders.
• One Modal primitive — golden Modal.tsx + hardened spec: focus trap, role="dialog", aria-modal, return-focus on close (fixes the 0/14-focus-trap class of bug, WCAG 2.4.3 + 2.1.2).
• One CI enforces all of it — .github/workflows/ci.yml runs tokens + contrast + spec + hardcode + theme-ref + npm test on every push/PR. Drift, contrast regressions, off-theme colors, and floating tokens cannot merge.
• 5 new runnable skills (10 → 15) — governance, token-build, figma-integration, design-qa, performance; verification steps added to design-code/prototype/ux-writing; router rows for all newer knowledge.
• validate_component_spec.py + lint_taste.py; fixed validate_tokens.py cross-file aliases; atoms Button/Input document all 8 states.

v1.2.1

• Docs — added docs/WORKFLOW.md: end-to-end how-it-works guide (Request Router, on-demand layer loading, real usage scenarios, automated release pipeline) + linked from the README
• CI — first release shipped fully automatically by the release.yml workflow (GitHub Release notes from this changelog + npm publish --provenance on tag push)

v1.2.0

• 8 new component specs — data-display.md (Calendar, Carousel, Tree) + data-viz.md (Bar, Line/Area, Pie/Donut, Sparkline, Scatter) → 50 components; plus components/icon-system.md
• Data-viz tokens — tokens/data-viz.json: color-blind-aware (Okabe–Ito) categorical/sequential/diverging palettes + axis/grid/tooltip → 14 token files
• ⟨⟩ 6 new framework adapters — Qwik, Astro, MUI, Mantine, Chakra, Bootstrap → 16 adapters
• Extended interop crosswalks — Ant Design 5, Shopify Polaris, GitHub Primer, Atlassian, Bootstrap (color-role tables + per-system notes)
• Accessibility depth — cognitive.md (load, plain language, dyslexia, reduced-data), i18n-rtl.md (logical properties, RTL mirroring, text expansion), vision.md (color blindness, low vision, forced-colors), wcag-aaa.md (AAA upgrade delta); +4 ARIA patterns (Carousel, Grid, Toolbar, Feed)
• Ops & pipeline workflows — governance.md (SemVer, contribution, deprecation), token-build.md (Style Dictionary / DTCG → multi-platform), figma-integration.md (token↔Variable sync, Figma MCP), design-qa.md (visual regression + a11y CI), performance.md (Core Web Vitals)

v1.1.0

• npm package — install into any project with npx ux-ui-agent-skills init (zero-dependency CLI: init / add / list, --force/--dry)
• Runnable skills — 10 invocable /skills under .claude/skills/ + real helper scripts (validate_tokens.py, contrast.py, design_systems.py, scaffold_component.py)
• Intelligence layer — Request Router in CLAUDE.md; Framework Adapter Protocol (target any framework) with 10 concise adapters; Design-System Interop Protocol + crosswalk (map to/from any design system)
• Native design-taste — taste/ (anti-slop doctrine, aesthetic archetypes, motion choreography) + a library of 138 brand-grade design systems
• 6 new token categories — gradients, opacity, blur, sizing, states, theming (multi-brand + density)
• 16 new component specs — Tabs, Breadcrumb, Pagination, Stepper, Menu, Toast, Banner, Skeleton, Progress, Empty State, Combobox, Select, Slider, Date Picker, File Upload, Popover, Command Palette, Divider
• Removed the externally-bundled taste skills in favor of native, first-party content

v1.0.1

• Motion tokens — tokens/motion.json: duration scale, easing curves, transition presets, keyframes, reduced-motion strategy
• UX writing guide — content/voice-tone.md: voice & tone system, error/empty-state formulas, microcopy patterns, inclusive language, pre-ship checklist
• Added cover.html repo cover image

v1.0.0

• Initial release — agent persona, 6 token files, 26 components (Atomic Design), WCAG 2.2 + ARIA references, 3 workflow guides, 3 framework guides

License

Released under the MIT License.