What is UX/UI Agent Skills?
UX/UI Agent Skills, also called XIAS, is a kit by plugin87 for AI coding agents. It combines design-system doctrine, DTCG design tokens, 26 runnable skills and 2,144 design system specs with 52 objective gates, 34 of which open a real browser to check contrast, keyboard use, target size and overflow.
UX/UI Agent Skills, short name XIAS, is a pure instruction and knowledge layer for coding agents, with no build tools or runtime. It carries design doctrine, tokens, component specs and a library of design systems that the agent reads before it designs.
The second half is enforcement. The kit ships 52 objective gates behind one command. The README says 34 of them open a real browser and measure rendered HTML, and the other 18 read files such as token JSON and component specs.
What changes in your agent's output
- Colors, spacing and type come from a three-tier token set (primitive, semantic, component) in DTCG format instead of hardcoded values.
- Components are specified with anatomy, variants, states and accessibility, and and the doctrine includes a set of eight states.
- Destructive actions must wear a danger color. A gate fails a Delete button filled with a primary blue.
- Contrast is checked on the real render in light and dark, per state, and touch targets must reach 24 by 24 pixels.
- Common AI tells are flagged, such as indigo to purple gradients, emoji used as icons and one radius and shadow everywhere.
Who it is for
Best for
- Teams that want measurable design and accessibility checks on agent-written UI, rather than a style guide the agent may forget.
- Projects with a design system or a target look, since the kit maps to Material 3, Apple HIG, Fluent, Carbon, shadcn/ui and others.
- Developers who want a second opinion step: the README describes a critic that renders the work and argues for rejection.
Not for
- Quick one-off tweaks. The full kit needs Node 20, Python 3 and Chrome, and 34 gates cannot measure without a browser.
- Anyone who expects a score to equal good taste. The README separates correctness, which gates measure, from quality, which is judged.
- Contributors wanting to send pull requests, because the README says pull requests are not merged.
Use it with UiChemy
UiChemy Figma to WordPress converts a design into an editable page, and this kit can check the rules behind that design first. Pair it with web-accessibility-skill after publishing. Click Get Started on uichemy.com.
How to install UX/UI Agent Skills: Design System Rules and 52 Gates
/plugin marketplace add plugin87/ux-ui-agent-skillsNeeds Node 20 or newer and Python 3. Pull requests are not merged, per the README.Claude Code plugin
Run the marketplace command above, then
/plugin install ux-ui-agent-skills@ux-ui-agent-skills. This gives 26 skills and the design-critic agent with no files copied into your repo.Other agents with npx
Run
npx [email protected] init --agent codexto write AGENTS.md and the kit. The README says Codex, Cursor, Copilot and Aider read this file, and a gemini option exists for Gemini CLI.Preview first
Run
npx [email protected] demoto open the rendered examples without installing, or add the dry-run flag to init to list every file it would write.Install a browser for the gates
In a project run
npm installandnpx playwright install chrome. Without Chrome the 34 browser gates report that they could not measure.
Check that it loaded
Run node scripts/accuracy_report.mjs in a clone. It should print 52/52, or report which gates are required and failing.
Key features
- Design token generationProduces DTCG JSON for color, typography, spacing, shadows, borders, breakpoints and motion in a primitive, semantic and component structure.
- 52 objective gatesOne command reports a real N/N. 34 gates open a real browser and 18 read files, and a gate that cannot look reports that it could not rather than a pass.
- 26 runnable skillsFourteen load when the work calls for them, such as design-tokens, a11y-audit and design-review. Twelve start only when you invoke them, such as critique and gate.
- 2,144 design systemsA library of design system specs for layout variance and visual direction. The install fetches the one the agent picks and checks its sha256 against an index.
- Framework adaptersAn adapter protocol targets React with Tailwind, Next.js, SwiftUI, Vue, Svelte, Angular, Flutter, Jetpack Compose and plain CSS.
- Accessibility auditingWCAG 2.2 AA and AAA findings ranked P0, P1 and P2, with state-aware contrast, axe roles, focus traps, RTL and reduced motion checks.
- Design critic in a separate contextThe critique skill runs as a forked subagent that never sees the conversation, renders the page at two widths in both themes and returns a verdict.
- Design review scoringDesigns are scored across six dimensions with Nielsen's ten heuristics and a structured findings table.
- Motion and UX writingTokenized durations and easing with a reduced-motion strategy, plus error and empty-state copy formulas.
- Edit-time hooksOn the plugin route, hooks block hardcoded values and emoji on write, and refuse to end a session that edited UI and measured none of it. The README says other agents get the rules but not the hooks.
Frequently asked questions
What is UX/UI Agent Skills?
It is an open source kit by plugin87, also called XIAS, that gives coding agents design-system rules, tokens, 26 skills and a library of design systems, then runs 52 gates against what the agent built. The README describes it as a pure instruction and knowledge layer with no runtime.
How do I install UX/UI Agent Skills?
On Claude Code, add the plugin marketplace and install the plugin with two commands. For other agents, run the npx init command with an agent flag such as codex or gemini. A third route is Homebrew, and a clone and copy also works.
Which AI coding agents does it work with?
The README names Claude Code, Codex, Cursor, Copilot, Jules, Aider, VS Code, Gemini CLI and Antigravity. Support differs: Claude Code gets skills and hooks, the AGENTS.md route gets the doctrine and runnable gates, and Gemini gets 23 ported skills.
Do I need Chrome or Node?
Yes for the full set. It needs Node 20 or newer and Python 3, and 34 of the 52 gates open a real browser. Without one, those gates print SKIPPED when run alone and count as failing in the full report.
Does a passing gate mean the design is good?
No. The README says the gates measure correctness and the critic judges quality, and that a passing gate is never evidence of taste. Its seeded critic fixtures pass all 52 gates and still carry fourteen design defects.
Is there an MCP server?
Yes. The README describes an MCP server with six tools: list_gates, run_gate, review_ui, get_doctrine, get_design_system and get_tokens. It lets any MCP client read the doctrine and run the gates, with exit codes that distinguish found, none and could not look.
Is it safe to run init in my project?
The README points to a SECURITY file that lists every file init writes and what each hook does, and says a dry-run flag prints the file list and changes nothing. It also lists the one network request the kit makes.
How does it compare with other design skills?
It is larger and enforcement-focused, with measured gates that check rendered pages instead of trusting the agent. For a lighter rules-only skill see ui-ux-pro-max, and for anti-generic taste rules see design-taste-frontend.


