UX/UI Agent Skills
aka XIAS
Design-system doctrine for coding agents, and 52 gates that check what they build instead of taking their word for it.
Instructions, design tokens, 26 runnable skills and 2,144 design systems (2,087 of them measured from the live site) that make a coding agent design to a senior standard, in any framework and against any design system. Then the gates measure what it built.
Current release v3.0.0 · Changelog · No build tools, dependencies, or runtime: a pure instruction and knowledge layer for AI agents.
The output, rendered
Not a mockup. These are screenshots of the files in examples/, taken by
node scripts/screenshot_docs.mjs from the same HTML the 52 gates measure, so
what you see below is what the gate run passed, in both themes.
Click through them yourself: plugin87.github.io/ux-ui-agent-skills. There are 56 live pages (twenty whole-product screens, five websites from a product launch to a Thai public service page, every component harness, both reference screens and an iPhone Duo simulator), with a theme toggle. No install, no clone.
Meridian Terminal is the density test: ten chart types on one screen:
candlestick, volume, moving average, depth, order book, tape, sparklines,
stacked area, waterfall, bubble and a correlation matrix. Every one of them is
inline SVG drawn from the same tokens, with no chart library.
Open it live.
Build one yourself with /data-dashboard.
Atlas is one HTML file in examples/showcase/, built only from the kit's
tokens and rules: the hero figure leads at 3x the body size, the three
breakdowns are deliberately three different shapes rather than three identical
cards, the range tabs move the numbers, and the table headers really sort.
Open it live.
One theme, two modes, no per-page palette. The destructive action wears the
danger variant in both. Loading keeps full strength and swaps in a spinner
instead of borrowing the disabled dimming. Every one of those is a rule in
CLAUDE.md that a gate or a critic enforces.
What the gates actually catch
Left: the same dashboard written the way a model writes it when nothing stops it: the indigo-to-purple gradient, four equal cards with no focal point, emoji as icons, one radius and one shadow everywhere, grey-on-white body text, and a blue Delete Account. Right: the reference app in this repo.
The difference is not a matter of opinion, and that is the point. The page on the
left is tests/fixtures/bad/slop-screen.html; here is what the gates say about it:
Ten gates reject it, none of them on a matter of taste. The right-hand page passes all 52.
The last row is there because building this comparison broke a gate open.
lint_intent originally read that blue Delete Account as fine: it resolved
"primary" and "danger" from the page's own CSS variables, and a page with no
tokens resolved neither, so it skipped the page and reported zero intent-bearing
controls. An untokenised page is precisely where intent gets picked by
convenience, so the gate no longer looks away: a destructive label filled with a
saturated colour outside the danger hue range is wrong-intent with or without a
theme. Re-checked against every example in both themes afterwards: no false
positives.
tests/meta/browser-gates.test.mjs pins every claim in this table, so it cannot
rot.
It was run blind, and it came back as rework
Two subagents were handed a brief and a project scaffolded by ux-ui-skills new,
which installs the kit and ships no example screens, and nothing else. No
hints, no warning that anything would be scored, no access to the conversation
that built the kit. They met CLAUDE.md and .claude/rules/ the way a new
user's agent does.
Both scored 14/14 on an independent run of evals/run.mjs, scored here
rather than self-reported. The rules transfer across a cold start.
And the two runs cost the kit seven defects that four in-session runs never
hit, because a familiar run keeps reaching for a finished demo theme instead of
the template a real user gets: a secondary button at 1.13:1 dark-on-dark
because the component tier never followed the dark map; a reduced-motion policy
in an external stylesheet read as "no policy" (Chromium treats a file:// linked
sheet as cross-origin, so the gate was blind exactly where every real project
lives); a missing scrim token; a theme that emitted colours and no spacing; an
intent gate that did not know "cancel subscription" is destructive.
Then the outputs went to /critique, which renders the work and argues for
rejection. Verdict on two pages that pass all fourteen gates: rework, eight
findings, five Major, and not one of them measurable: an empty state that reads
as a page that failed to load, a theme toggle that communicates nothing about its
own state, a toast claiming "Draft project created" over a list that never
changed, and a loading state wearing the disabled dimming so it reads as "you
cannot do this".
The blind outputs were never edited. They are in evals/out/ as the record
of what a cold-start agent produced; patching them would be editing the
experiment. The findings were spent on the kit instead. One of them became a
gate, verify_interactive.mjs, which fails any control that declares a state
contract and changes nothing when clicked.
The whole log, including what is still unproven, is in evals/RESULTS.md.
What It Does
Quick Start
Five ways in, and which one is yours: docs/INSTALL.md walks through the Claude Code plugin, the
AGENTS.mdsurface for Codex and Cursor, the MCP server, Homebrew andnpx: what each one installs, and what you give up by choosing it.
Option A: Install as a Claude Code plugin (recommended)
Two lines in Claude Code, and every skill, command, and agent is available in any project you open, with no files copied into your repo:
/plugin marketplace add plugin87/ux-ui-agent-skills
/plugin install ux-ui-agent-skills@ux-ui-agent-skills
You get 26 skills and the design-critic agent.
Fourteen the model reaches for on its own when the work calls for them:
design-tokens, design-component, design-code, design-review,
a11y-audit, apply-aesthetic, data-dashboard, design-qa,
figma-integration, performance, token-build, ux-writing,
design-doctrine, and design-screen, the one that claims "design a page /
screen / app", which until 2026-10-05 no skill did, so the most common request
there is matched nothing strongly and the kit got used at a fraction of its
depth.
Twelve you start yourself, because each one takes an action or sets a
direction that should be your call: /brandkit, /governance, /image-to-code,
/migrate-design-system, /prototype, /redesign, /gate, /critique,
/grill-me, /ship, /scaffold-project, /observe. They carry
disable-model-invocation: true, which is the field that actually stops the
model invoking them. Until 2026-10-05 six of them said invocation: user, a
key Claude Code does not read, so they were auto-invocable the whole time. The
last five were commands and are skills now, so they have a CLAUDE_SKILL_DIR of
their own and their paths resolve under a plugin install.
design-doctrine carries the house rules that CLAUDE.md carries in the repo,
because a plugin root CLAUDE.md is not loaded as project context, as Claude
Code's own plugin validate says so.
Then just work. Ask for the thing you want and the right skill loads itself:
"Design a notification component with all states and accessibility"
"Build the billing settings screen, one shared theme, light and dark"
"/grill-me" interrogate the brief before anything is built
"/gate" run all 52 checks and report the real N/N
"/critique" hand the result to a critic that argues for rejection
"/observe" did the agent follow the rules? sessions, blocks, findings left open
If the skills do not show up straight away, start a new session. To check what is loaded, update, or remove it:
claude plugin details ux-ui-agent-skills # inventory + token cost per skill
claude plugin update ux-ui-agent-skills
claude plugin uninstall ux-ui-agent-skills
claude plugin marketplace remove ux-ui-agent-skills
uninstall also deletes the plugin's data folder, which holds the gate
receipts and the /observe history.
34 of the gates open a real browser. They use the Chrome installed on the
machine, or the one npx playwright install chrome downloads. Without either
they cannot measure: run through scripts/gate.mjs, as the skills do, they fail
with REQUIRED, FAILING instead of passing on nothing.
See it first, install nothing
npx ux-ui-agent-skills demo # copies the rendered examples and opens them
Every page it opens is a page the gates measure. Delete the folder afterwards; nothing was installed. Or skip the copy entirely and use the live demo.
Option B: Use it from any MCP client
claude mcp add ux-ui-gates -- npx -y --package=ux-ui-agent-skills ux-ui-mcp
Six tools: list_gates, run_gate, review_ui, get_doctrine,
get_design_system, get_tokens. Read the doctrine and the tokens before
generating, run the gates on what you produced after.
AGENTS.md already lets any agent read the doctrine. This lets any MCP
client run the 52 gates, which reading cannot do. Plenty of things can
describe good UI, very little can tell you afterwards that the contrast you
shipped is 3.9:1 on hover.
It returns what the gates printed, unedited, with their exit codes: 0 looked and
found nothing, 1 found something, 2 could not look, never a pass. No SDK
dependency; the protocol is implemented directly so "dependencies": {} stays
true. Details in docs/MCP.md.
Option C: Install with Homebrew
brew install plugin87/tap/ux-ui-agent-skills
ux-ui-skills init # CLAUDE.md + .claude/ skills, rules, hooks
ux-ui-skills init --agent codex # AGENTS.md, for Codex, Cursor, Copilot, Aider
Same npm tarball the npx route runs, so this is a convenience rather than a
second product that can drift. The formula is generated from the registry and
its test do block asserts the installed CLI reports the version the formula
names, and that --agent codex writes AGENTS.md without CLAUDE.md or
.claude/. Details in docs/HOMEBREW.md.
Option D: Install with npx
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
npx ux-ui-agent-skills library get stripe # fetch one design-system spec
Flags: --force (overwrite existing files) · --dry (preview, change nothing).
The 2,144 design-system specs are 27.6MB, and a project uses one or two, so
init does not copy them. It installs an index of every spec with its sha256,
and the agent fetches the one it picks with node scripts/library.mjs get <name>. A fetched file whose sha256 does not match the index is refused and
nothing is written. library pull fetches all of them for offline work. The
plugin (Option A) and a clone carry every spec already. To browse all 2,144, open
design-systems/catalog.md: GitHub's folder view lists
only the first 1,000.
Working on the kit itself, or want it vendored? Clone and copy instead.
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, and 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"
Proving It, and Admitting What Cannot Be Proven
The kit ships 52 objective gates behind one command:
node scripts/accuracy_report.mjs # 52/52 or it fails, no partial credit
34 of them open a real browser, so they need one installed. npx ux-ui-agent-skills init does not install Playwright. Whether /plugin install
does depends on the Claude Code version: 2.1.301 installed the kit's npm
dependencies, Playwright included, into the plugin folder. Neither installs the
browser itself. In a clone, run this once before expecting a full score:
npm install # playwright
npx playwright install chrome # real Chrome: six gates require the channel
Without it those 34 report REQUIRED, FAILING under accuracy_report.mjs, which
is the honest answer. Run individually they print SKIPPED and exit 0, so
run a single render gate through scripts/gate.mjs (or prefix it with
DS_REQUIRE_BROWSER=1) if you are reading its exit code, rather than reading
silence as green. Every skill already does.
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 from 280 to 951px and folded on iPhone Duo, 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.
What that number covers, stated exactly. 34 of the 52 checks open a real
browser, so what they measure is rendered HTML: the 23 component harnesses,
the twenty industry screens, the reference app, the live demo, the starter
template, and the React source compiled and server-rendered. The other 18 read
files: token JSON and alias resolution, contrast math on the token source,
component specs, hardcoded values, theme references, emoji, hooks, the
instruction surface held in step across CLAUDE.md and AGENTS.md, and
destructive-intent declarations in framework source.
The split is not typed here: tests/meta/counts.test.mjs derives it from which
gates import Playwright and fails if this paragraph drifts from the report. It
read 35/16 until 2026-10-07, when the derivation was first run. Those were two numbers
nobody had measured, in the section about not stating numbers you did not
measure.
Framework source (.tsx, .vue, .swift) gets the file-reading checks: no
emoji, no hardcoded values, every var(--…) resolving to the theme, and, since a
blue Delete shipped in this repo's own Settings.tsx while the HTML twin of that
screen was correct, every destructive control declaring its intent
(lint_intent_source.mjs). That last one proves a declaration, never a
colour: variant="destructive" wired to a blue token passes it.
Only a render catches that, so the React components are now compiled with esbuild,
server-rendered, and put through the same browser gates as the HTML
(render_framework_source.mjs, which also fails if any class the components emit
resolves in no linked stylesheet). .vue and .swift are still file-checks only,
and are not proven by this number until they are rendered and measured.
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.
It runs as a forked subagent, so the critic never sees the conversation that produced the work. That is not a performance decision. A critic holding the maker's reasoning is reviewing the argument instead of the artifact, and it agrees far too easily.
The critic itself is scored. tests/fixtures/critic/ holds four pages carrying
fourteen seeded design defects (no focal point, an empty state with no way
forward, a toast that claims a project was created while the list is unchanged),
and every one of those pages passes all 52 gates. That is the property that
makes the number mean something: nothing in the harness can reach these, so
catching them takes judgement. node evals/score_critic.mjs scores a critique
against the key, prints the matching excerpt for every hit and the rule for every
miss, and says in its own output that the number is a floor until you have read
them.
/observe answers a different question: not whether the work is right, but
whether the agent did what the rules say while making it. The hooks record one
event per session start, per UI edit (with what each fast gate caught) and per
end of turn (which pages had a passing gate, and whether the Stop hook passed,
blocked, or let the agent go with pages still open). The report joins those to
the gate receipts and gives each session a verdict from three counted rules:
every edited page measured, no finding left on a file's latest version, never
finished while blocked. --last and --session <id> print a timeline;
--html writes the same thing as a page. Everything stays on the machine,
files older than 30 days are deleted, and DS_OBSERVE=0 turns it off.
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.
Documentation
Not using Claude Code?
The kit works with Codex, Cursor, Copilot, Jules, Aider, VS Code and about
twenty more, through the open AGENTS.md convention, and with Gemini CLI
and Antigravity, which read GEMINI.md and .agents/skills/.
npx ux-ui-agent-skills init --agent codex # AGENTS.md + the whole kit
npx ux-ui-agent-skills init --agent gemini # GEMINI.md + 23 ported skills: Gemini CLI, Antigravity
npx ux-ui-agent-skills init --agent both # CLAUDE.md and AGENTS.md
npx ux-ui-agent-skills init # Claude Code (default)
gemini extensions install https://github.com/plugin87/ux-ui-agent-skills # Gemini CLI, every project
What transfers, and what does not, stated exactly:
A skill is ported, not copied: Gemini CLI and Antigravity do not expand
${CLAUDE_SKILL_DIR}, so each path is rewritten for where the skill lands, and
the three that need Claude Code (critique, observe, scaffold-project) stay
there. Setup and what was measured: docs/INSTALL.md.
The hooks are the real difference. On Claude Code, three rules are enforced
without the model choosing to: no hardcoded values and no emoji on every write,
and a refusal to end a session that edited UI and measured none of it. On the
AGENTS.md surface every rule is a request and the gates run when you run them.
scripts/validate_agents_surface.py holds the two files in step, so a rule added
to one cannot quietly skip the other. Full setup, and the two defects a real
Codex run found in the first draft: docs/AGENTS-SETUP.md.
Contributing
Pull requests are not merged here. Every line is written by the maintainer, deliberately. Asking for work against a bar and then declining it would waste your time, so the policy is stated up front instead of at the end of a review.
What is wanted, and credited by name in the CHANGELOG: a gate gap, a reproduction, or a proposal. Forking is welcome; the MIT licence says so.
Two commands are the whole bar for anything that does ship:
node scripts/accuracy_report.mjs (52/52, no partial credit) and
npm run test:gates (every gate must still reject its broken fixture). Paste
the real output rather than describing it.
The most valuable issue this repo can receive is a gate gap: a case where a gate said yes to work it should have caught. There is a template for exactly that, because a gate that passes broken work is worse than a missing one: it turns a real defect into a green tick.
CONTRIBUTING.md covers setup, the non-negotiable rules, how to add a gate that can still say no, and the SemVer table. CODE_OF_CONDUCT.md applies to every space in the project.
License
Released under the MIT License: free to use, modify, and distribute, including commercially, as long as the copyright notice and the permission notice travel with the copy.
Copyright (c) 2026 Thientan Soparat (@plugin87).
Built by Thientan Soparat (@plugin87).
If this kit helps you, star it on GitHub so others can find it.
