UI Skills

When invoked, apply these opinionated constraints for building better interfaces.

MANDATORY: Kimi Delegation for UI Implementation

All UI implementation work MUST be delegated to Kimi K2.5 via MCP.

Kimi excels at frontend development. Claude reviews, Kimi builds:

// Delegate UI implementation to Kimi
mcp__kimi__spawn_agent({
  prompt: `Implement [component/feature].
Apply ui-skills constraints:
- Tailwind CSS defaults, cn() utility
- Accessible primitives (Base UI/Radix/React Aria)
- No h-screen (use h-dvh), respect safe-area-inset
- Animator only transform/opacity, max 200ms feedback
- text-balance for headings, tabular-nums for data
Existing patterns: [reference files]
Output: ${targetPath}`,
  thinking: true
})

Workflow:

1. Define constraints → Claude (this skill)
2. Implement UI → Kimi (Agent Swarm)
3. Review quality → Claude (expert panel review)

Anti-pattern: Implementing UI yourself instead of delegating to Kimi.

How to use

• /ui-skills Apply these constraints to any UI work in this conversation.

• /ui-skills <file> Review the file against all constraints below and output:

  • violations (quote the exact line/snippet)
  • why it matters (1 short sentence)
  • a concrete fix (code-level suggestion)

Stack

• MUST use Tailwind CSS defaults unless custom values already exist or are explicitly requested
• MUST use motion/react (formerly framer-motion) when JavaScript animation is required
• SHOULD use tw-animate-css for entrance and micro-animations in Tailwind CSS
• MUST use cn utility (clsx + tailwind-merge) for class logic

Components

• MUST use accessible component primitives for anything with keyboard or focus behavior (Base UI, React Aria, Radix)
• MUST use the project’s existing component primitives first
• NEVER mix primitive systems within the same interaction surface
• SHOULD prefer `Base UI` for new primitives if compatible with the stack
• MUST add an aria-label to icon-only buttons
• NEVER rebuild keyboard or focus behavior by hand unless explicitly requested

Interaction

• MUST use an AlertDialog for destructive or irreversible actions
• SHOULD use structural skeletons for loading states
• NEVER use h-screen, use h-dvh
• MUST respect safe-area-inset for fixed elements
• MUST show errors next to where the action happens
• NEVER block paste in input or textarea elements

Animation

• NEVER add animation unless it is explicitly requested
• MUST animate only compositor props (transform, opacity)
• NEVER animate layout properties (width, height, top, left, margin, padding)
• SHOULD avoid animating paint properties (background, color) except for small, local UI (text, icons)
• SHOULD use ease-out on entrance
• NEVER exceed 200ms for interaction feedback
• MUST pause looping animations when off-screen
• SHOULD respect prefers-reduced-motion
• NEVER introduce custom easing curves unless explicitly requested
• SHOULD avoid animating large images or full-screen surfaces

Typography

• MUST use text-balance for headings and text-pretty for body/paragraphs
• MUST use tabular-nums for data
• SHOULD use truncate or line-clamp for dense UI
• NEVER modify letter-spacing (tracking-*) unless explicitly requested

Layout

• MUST use a fixed z-index scale (no arbitrary z-*)
• SHOULD use size-* for square elements instead of w-* + h-*

Performance

• NEVER animate large blur() or backdrop-filter surfaces
• NEVER apply will-change outside an active animation
• NEVER use useEffect for anything that can be expressed as render logic

Design

• NEVER use gradients unless explicitly requested
• NEVER use purple or multicolor gradients
• NEVER use glow effects as primary affordances
• SHOULD use Tailwind CSS default shadow scale unless explicitly requested
• MUST give empty states one clear next action
• SHOULD limit accent color usage to one per view
• SHOULD use existing theme or Tailwind CSS color tokens before introducing new ones

Expert Panel Review (MANDATORY)

Before returning ANY design output to the user, it MUST pass expert panel review.

See full details: references/expert-panel-review.md

Quick Reference

1. Simulate 10 world-class advertorial experts:

   • Ogilvy (advertising), Rams (industrial design), Scher (typography)
   • Wiebe (conversion copy), Laja (CRO), Walter (UX)
   • Cialdini (persuasion), Ive (product design), Wroblewski (mobile)
   • Millman (brand strategy)

2. Each expert scores 0-100 with specific improvement feedback

3. Threshold: 90+ average required

4. If below 90: implement feedback, iterate, re-review

5. Only return design to user when 90+ achieved

Example Output

Expert Panel Review: Hero Section

| Expert | Score | Critical Improvement |
|--------|-------|---------------------|
| Ogilvy | 88 | Lead with benefit, not feature |
| Rams | 94 | Clean, focused |
| Scher | 86 | H2 needs more weight contrast |
| Wiebe | 81 | "Get Started" → "Start Free Trial" |
| Laja | 77 | No social proof above fold |
| Walter | 90 | Good emotional resonance |
| Cialdini | 83 | Add urgency element |
| Ive | 92 | Refined execution |
| Wroblewski | 88 | Touch targets good |
| Millman | 85 | Voice slightly inconsistent |

**Average: 86.4** ❌ Below threshold

Implementing: Laja (social proof), Wiebe (CTA), Cialdini (urgency)...

Anti-Patterns

• ❌ Skipping review for "quick fixes"
• ❌ Accepting 85+ as "close enough"
• ❌ Generic feedback ("make it better")
• ❌ Returning design without 90+ score