design-code — Claude Code skill: UI code and installation
This skill generates component and screen code for a requested framework. Use it when you need token-based UI with accessibility checks.
npx skills add plugin87/ux-ui-agent-skills --skill design-code -a claude-codeWhat design-code does
The agent first identifies the framework and styling method, then reads the shared protocol and a matching adapter. It also uses component specs, ARIA patterns, and JSON token files. When the requested stack has no adapter, the protocol checklist guides the agent in creating one.
The output should contain complete files rather than placeholders. The instructions call for tokens instead of hardcoded values, applicable component states, keyboard interaction, dark mode, mobile-first layouts, and reduced-motion support. Lint, accessibility, and theme scripts are included for verification.
The skill reads a kit from the tokens, frameworks, components, and accessibility directories. If its check returns KIT_MISSING, the agent must stop and point to npx ux-ui-agent-skills init or plugin installation. Without those files, it cannot follow the skill's code-generation rules.
What it helps with
- build a component for a chosen framework
- create a screen using project tokens
- add ARIA roles and keyboard interaction
- check component states and dark mode
- find hardcoded colors and dimensions in UI code
Example prompts for the agent
- Build a button component for our stack and check its accessibility.
- Create a settings screen using project tokens and dark mode.
- Check this component for hardcoded values and missing UI states.
How to install design-code
1Skills CLI (Vercel)
npx skills add plugin87/ux-ui-agent-skills --skill design-code -a claude-code2Manually: put the files in place
~/.claude/skills/design-code/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/design-code/
1Skills CLI (Vercel)
npx skills add plugin87/ux-ui-agent-skills --skill design-code -a codex2Manually: put the files in place
.agents/skills/design-code/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add plugin87/ux-ui-agent-skills --skill design-code -a cursor2Manually: put the files in place
.cursor/skills/design-code/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install plugin87/ux-ui-agent-skills design-code2Manually: put the files in place
.github/skills/design-code/SKILL.md1Skills CLI (Vercel)
npx skills add plugin87/ux-ui-agent-skills --skill design-code -a gemini-cli2Platform CLI
gemini skills install https://github.com/plugin87/ux-ui-agent-skills --path .claude/skills/design-codeManually: put the files in place
.devin/skills/design-code/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add plugin87/ux-ui-agent-skills --skill design-code -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about design-code
How do I install design-code in Claude Code?
Run npx skills add plugin87/ux-ui-agent-skills --skill design-code -a claude-code. The skill file is placed at ~/.claude/skills/design-code/SKILL.md.
What does design-code do?
It generates component and screen code for a requested framework and styling method, drawing on adapters, component specs, and tokens. Its instructions also cover accessibility, UI states, and theme checks.
What does design-code need to run?
It reads kit files from tokens, frameworks, components, and accessibility. If the kit check returns KIT_MISSING, the agent should stop and suggest npx ux-ui-agent-skills init or plugin installation.
Source text
SKILL.md · MIT
---
name: design-code
description: Generate production-ready, accessible, token-driven component code for ANY framework (React+Tailwind, Next.js, SwiftUI, Vue, Svelte, Angular, Solid, Web Components/Lit, React Native, Flutter, Jetpack Compose, vanilla CSS, or CSS-in-JS). Use when the user wants working UI code for a component or screen in a specific stack.
---
# Skill: Design Code
> **Step 0: is the kit here?** This skill reads files from the kit. Check once:
> `ls ${CLAUDE_SKILL_DIR}/../../../tokens >/dev/null 2>&1 && echo KIT_OK || echo KIT_MISSING`
> On `KIT_MISSING` only the skill folders were installed, which is what
> `npx skills add` does. Say so plainly, point the user at
> `npx ux-ui-agent-skills init` or the plugin install, and stop. Do not guess the
> contents of a file you could not open.
Render components into a target framework via the adapter system.
## Steps
1. Identify the **target framework** + styling method from the request (or ask).
2. Read `${CLAUDE_SKILL_DIR}/../../../frameworks/adapter-protocol.md` (the universal contract: token resolution, component contract, styling, dark mode, motion + output rules).
3. Load the concrete adapter:
- Full references: `${CLAUDE_SKILL_DIR}/../../../frameworks/react-tailwind.md`, `${CLAUDE_SKILL_DIR}/../../../frameworks/nextjs.md`, `${CLAUDE_SKILL_DIR}/../../../frameworks/swiftui.md`.
- Concise adapters: `${CLAUDE_SKILL_DIR}/../../../frameworks/adapters/{vue,svelte,angular,solid,web-components-lit,react-native,flutter,jetpack-compose,vanilla-css,css-in-js}.md`.
- **No file for the target?** Generate an adapter on the fly using the protocol's "author a new adapter" checklist.
4. Pull the component spec from `${CLAUDE_SKILL_DIR}/../../../components/*` and the ARIA pattern from `${CLAUDE_SKILL_DIR}/../../../accessibility/aria-patterns.md`.
5. Resolve all values to tokens (`${CLAUDE_SKILL_DIR}/../../../tokens/*.json`); apply aesthetic direction if requested (`apply-aesthetic` skill).
## Output rules (mandatory)
Use tokens (never hardcode) · include a11y on every interactive element · handle all applicable of the 8 states · support dark mode at the semantic layer · mobile-first · honor reduced motion · **deliver complete files, no placeholders** (`${CLAUDE_SKILL_DIR}/../../../workflows/redesign-audit.md` → Output Completeness).
## Verification (mandatory before declaring done)
Code is the highest-stakes output. Self-check every time:
1. **No hardcoded values**: every color/size/radius/shadow/duration/**font** traces to a token (CSS var / theme key / asset). Run `${CLAUDE_SKILL_DIR}/../../../scripts/lint_hardcodes.py` over the output; zero raw hex/px/ms AND zero raw Tailwind palette utilities (`bg-gray-500`, `text-blue-600`). Use semantic utilities/tokens (`bg-surface`, `text-primary`). Run `${CLAUDE_SKILL_DIR}/../../../scripts/validate_theme_refs.py` so every `var(--…)` resolves to a defined theme token (no floating tokens). One allowed exception: 3rd-party theme-config (MUI/Mantine) mapping our tokens INTO their API.
2. **All applicable states present**: Default, Hover, Focus(-visible ring), Active, Disabled, Loading(+`aria-busy`), Error, Selected, or justified N/A.
3. **Accessibility wired**: correct role/ARIA per `${CLAUDE_SKILL_DIR}/../../../accessibility/aria-patterns.md`, keyboard model, focus management, ≥24px target; verify contrast (`${CLAUDE_SKILL_DIR}/../../../scripts/contrast.py`) for any new color pair.
4. **Dark mode + reduced motion + responsive**: semantic tokens swap; motion has a `prefers-reduced-motion` fallback; layout is mobile-first.
5. **Completeness**: full files, no `// ...`; if asked for N, deliver N. If any check fails, fix before returning (run `a11y-audit` if unsure).
6. **Single-theme consistency**: consume the project's ONE shared token theme (the root CSS-var layer); never define a per-page palette or new colors. Across multiple pages/screens, the same semantic tokens must drive every surface so the whole product stays visually identical and themeable from one place (`${CLAUDE_SKILL_DIR}/../../../.claude/rules/tokens-and-color.md` → Single-Theme Consistency).
7. **One shared primitive layer**: build ONE reusable component per atom (`Button`, `Input`, `Modal`, `Badge` via `cva`/equivalent); never repeat utility-class clusters inline across files or hand-roll a div-as-modal per screen. Overlays reuse the single `Modal` primitive: focus trap, `role="dialog"`, `aria-modal="true"`, `aria-labelledby`, Escape, **return focus on close**, backdrop (WCAG 2.4.3 + 2.1.2). See `${CLAUDE_SKILL_DIR}/../../../examples/golden/Button.tsx` + `${CLAUDE_SKILL_DIR}/../../../examples/golden/Modal.tsx`.
8. **Font loading**: NEVER `@import` web fonts in CSS (render-blocking). Use a framework loader (`next/font`) or `<link rel="preconnect">` + `<link rel="preload">` with `font-display: swap`; self-host when possible. Font family comes from a token (`--font-sans`), never a literal.
9. **Semantic token BY INTENT + consistency**: the token's *meaning* must match the action: destructive (Delete/Remove) → `action.destructive` (danger), **never** `action.primary`; secondary = neutral outline/transparent with dark text (**never a colored fill** → no dark-text-on-blue). The SAME action uses the SAME variant **everywhere** (a trigger button and its confirm button must match, not red in one place and blue in another). Gates don't catch this. You must.
10. **No emoji as icons**: use a real icon set (default **lucide**) as inline SVG with `currentColor`; never an emoji, including in JS that swaps a label (swap the `<svg>`, not a text/emoji string). Mentally run `${CLAUDE_SKILL_DIR}/../../../scripts/lint_taste.py` (it flags emoji-as-icon).
11. **Destructive confirmation UX**: irreversible actions (delete account/data, anything stated "cannot be undone") need real friction per WCAG 3.3.4 / 3.3.6: a confirmation dialog whose **confirm button restates the action** ("Delete account", not "Delete"/"OK"/"Yes") and, for high-stakes, a **type-to-confirm** step. See `${CLAUDE_SKILL_DIR}/../../../examples/sample-app/preview.html` and `${CLAUDE_SKILL_DIR}/../../../content/voice-tone.md`.
12. **Taste pre-flight**: run the 12-point aesthetic check in `${CLAUDE_SKILL_DIR}/../../../taste/design-taste.md` and, for any rendered HTML, `DS_REQUIRE_BROWSER=1 node ${CLAUDE_SKILL_DIR}/../../../scripts/taste_audit.mjs <file>` (render-based: flags timid type-scale contrast, uniform repetition, over-wide measure, palette sprawl). Taste is heuristic: treat findings as a strong signal and pair with a real screenshot review; it is NOT auto-provable like correctness.
13. **Run the gates, then report (mandatory).** Before declaring done, RUN `node ${CLAUDE_SKILL_DIR}/../../../scripts/gate.mjs verify_states.mjs <file>` + `--dark` (every element, default/hover/focus) and `node ${CLAUDE_SKILL_DIR}/../../../scripts/accuracy_report.mjs`; report their actual output. Never type a contrast number or "100%" you didn't just measure (CLAUDE.md → Verification Protocol).
© plugin87 · MIT · plugin87/ux-ui-agent-skills