XIAS(UX/UI Agent Skills):用令牌、规格与可运行门禁约束编程代理出 UI
开源 MIT 项目 UX/UI Agent Skills(XIAS)提供无构建依赖的设计系统教义层:138 套品牌级系统、组件规格、DTCG 令牌与数十条客观门禁(含真实浏览器渲染检查),可挂到 Claude Code、Cursor、Codex 与 MCP。
开源项目 plugin87/ux-ui-agent-skills(又名 XIAS,MIT)面向「模型会画漂亮界面,却难守设计系统」:它不是又一个运行时 UI 库,而是可丢进任意项目的指令与知识层——无构建步骤、无依赖、无运行时——给编程代理完整的设计系统教义、令牌、规格与门禁。
仓库自述能力包括:约 138 套品牌级设计系统;数十条客观门禁(其中大量会打开真实浏览器检查),覆盖真实渲染下的 WCAG 对比度、axe-core 角色、键盘可达、触控目标尺寸、窄屏溢出等;按意图强制令牌(例如「危险色」不能随便填成品牌蓝);可运行 skills、组件规格、DTCG 令牌文件与多框架适配器;兼容 Claude Code、Codex、Cursor、Copilot、Aider 与 MCP,目标栈可到 React、Next、SwiftUI、Tailwind 等。README 用「收据」演示:未受约束模型写出的渐变看板、emoji 与蓝色「删除账号」会被门禁客观拒绝,而按规则建的参考页可通过全部门禁。安装示例包括 Claude Code 插件市场,以及把 MCP 门禁服务挂进客户端。
对产品设计智能化读者,价值在把「审美争论」换成可重复的失败信号:门禁不能观察时不得报通过。落地时仍需人工定品牌令牌与组件边界;把门禁接入 CI 前确认无头浏览器环境稳定,避免把 SKIPPED 当成绿。
UX/UI Agent Skills (XIAS) for coding agents
Turn Claude into a Senior Design Architect with structured instructions, design tokens, and 138 corporate-brand level design systems to achieve consistent, accessible, and token-based design outputs.
It's called UX/UI Agent Skills (XIAS).
It's a pure instruction and knowledge layer you drop into any project. No build step, no dependencies, no runtime. It hands your coding agent a full design-system doctrine plus the tokens, specs, and gates to actually enforce it.
→ 138 brand-grade design systems to build against, any framework, any stack → 52 objective gates that pass or fail the output, 34 of them opening a real browser to check it → Real-render WCAG contrast in every state, axe-core roles, keyboard, target size, no overflow at 280px → Token-by-intent enforcement: a destructive button filled with your brand blue gets rejected because that is not a danger color → 26 runnable skills, 52 component specs, 14 DTCG token files, 16 framework adapters → Works with Claude Code, Codex, Cursor, Copilot, Aider, and MCP → React 19, Next.js 15, SwiftUI 6, Tailwind v4, or whatever you already use
Here's the part that proves it works:
The repo ships the receipt. It takes the exact dashboard a model writes when nothing stops it, the gradient, the four equal cards, the emoji, the blue Delete Account, and runs it through the gates. Ten of them reject it, none on taste. A white heading on a gradient scoring 1.00:1 contrast. A 15px icon button under the 24px minimum. 63 hardcoded values instead of tokens. The reference page built to the rules passes all 52.
You can click through 49 live pages of the output yourself, both themes, no install.
[add live star count] GitHub stars. MIT License. 100% open source.
Link in the comments. What's the ugliest AI-generated UI you've had to fix?