design-exploration（设计探索）

DEPRECATED. Use /evolve instead. /evolve replaces this skill entirely. It explores design directions with a genetic algorithm, persistent cross-project memory, multi-provider delegation, and taste-frontend.md injection. The investigation → catalogue → refinement → selection cycle maps directly to evolve's detect → generate → select → advance → lock workflow. Original skill content preserved below for reference.

Design Exploration

Investigate existing designs, generate a visual catalogue of 6-12 diverse proposals, facilitate iterative refinement, and output a selected direction for implementation.

MANDATORY: Kimi Delegation for Visual Work

All visual proposal generation MUST be delegated to Kimi K2.5 via MCP.

Kimi excels at frontend development and visual coding. For design exploration:

1. Research direction → Gemini (web grounding, trend research)
2. Generate proposals → Kimi (visual implementation via Agent Swarm)
3. Review/iterate → Claude (orchestration, quality gates)

// Delegate proposal generation to Kimi
mcp__kimi__spawn_agent({
  prompt: `Generate design proposal for [component].
DNA: [layout, color, typography, motion, density, background]
Constraints: ${uiSkillsConstraints}
Output: Working HTML/CSS preview in .design-catalogue/proposals/[name]/`,
  thinking: true  // Enable for complex proposals
})

// For parallel proposal generation (4.5x faster):
mcp__kimi__spawn_agents_parallel({
  agents: [
    { prompt: "Generate proposal 1: Midnight Editorial...", thinking: true },
    { prompt: "Generate proposal 2: Swiss Brutalist...", thinking: true },
    { prompt: "Generate proposal 3: Warm Workshop...", thinking: true },
    // ... up to 100 parallel agents
  ]
})

Anti-pattern: Generating proposals yourself instead of delegating to Kimi. Pattern: Research (Gemini) → Generate (Kimi) → Review (Claude)

When to Use

• Before /aesthetic or /polish when direction is unclear
• When user wants to see multiple design options visually
• When redesigning or refreshing existing UI
• When maturity < 6 and no clear aesthetic direction exists

Core Principle: Visual Over Verbal

The catalogue is a working visual, not markdown descriptions.

Users see actual styled components, real typography, live color palettes - not just words describing what things *would* look like.

Phase 0: Backend Detection

Check available rendering backends:

// Check if Pencil MCP is available
try {
  mcp__pencil__get_editor_state({ include_schema: false })
  // Pencil available → use pencil-renderer
  BACKEND = "pencil"
} catch {
  // Fallback to HTML renderer
  BACKEND = "html"
}

Backend selection:

• Pencil (preferred): Native.pen file rendering, design system integration
• HTML (fallback): Static HTML/CSS proposals served locally

Workflow

1. Investigation

Understand what exists:

# Screenshot current state via Chrome MCP
mcp__claude-in-chrome__tabs_context_mcp
mcp__claude-in-chrome__navigate url="[current app URL]"
mcp__claude-in-chrome__computer action="screenshot"

Analyze the current design:

• Typography: What fonts, sizes, weights?
• Colors: What palette, what semantic meaning?
• Layout: Centered? Asymmetric? Grid-based?
• Motion: Any animations? What timing?
• Components: What patterns exist?

Infer current DNA:

DNA Axes:
- Layout: [centered|asymmetric|grid-breaking|full-bleed|bento|editorial]
- Color: [dark|light|monochrome|gradient|high-contrast|brand-tinted]
- Typography: [display-heavy|text-forward|minimal|expressive|editorial]
- Motion: [orchestrated|subtle|aggressive|none|scroll-triggered]
- Density: [spacious|compact|mixed|full-bleed]
- Background: [solid|gradient|textured|patterned|layered]

Identify:

• Strengths to preserve
• Weaknesses/opportunities
• Constraints (brand colors, accessibility, tech stack)

Research via Gemini:

gemini -p "I'm building [describe component/page]. Research:
1. Distinctive design approaches (avoid AI-slop aesthetics)
2. Real-world examples of excellent [component type]
3. Current typography/color/layout trends for this context
4. Unexpected alternatives to obvious solutions"

2. Build Visual Catalogue

Generate 6-12 proposals using DNA variation system.

If BACKEND = "pencil":

Use pencil-renderer primitive:

// For each proposal
// 1. Get style guide matching DNA mood
mcp__pencil__get_style_guide_tags()
mcp__pencil__get_style_guide({ tags: [mapped_tags] })

// 2. Render via pencil-renderer workflow
// See: pencil-renderer/SKILL.md

All proposals render to a single.pen document as separate frames.

If BACKEND = "html":

Create catalogue directory:

.design-catalogue/
├── index.html              # Main viewer
├── styles/catalogue.css    # Viewer chrome styles
├── proposals/
│   ├── 01-[name]/
│   │   ├── preview.html
│   │   └── styles.css
│   ├── 02-[name]/
│   │   └── ...
│   └── ... (6-12 total)
└── assets/

DNA Variation Rule: No two proposals share >2 axes.

Required diversity:

• At least 2 bold/dramatic directions
• At least 2 refined/subtle directions
• At least 1 unexpected/wild card
• At least 1 that preserves current strengths

Each proposal preview includes:

• Hero section in that style
• Sample card component
• Button states (default, hover, active)
• Typography scale (h1-h6, body, caption)
• Color palette swatches with hex codes
• DNA code badge

Reference templates:

• references/viewer-template.html - Main catalogue viewer
• references/proposal-template.html - Individual proposal page
• references/catalogue-template.md - Metadata structure

Anti-convergence check: Reference aesthetic-system/references/banned-patterns.md

• No Inter, Space Grotesk, Roboto as primary fonts
• No purple gradients on white
• No centered-only layouts for all proposals

3. Present & Serve

If BACKEND = "pencil":

// Screenshot each proposal frame
for (const frame of proposalFrames) {
  mcp__pencil__get_screenshot({ nodeId: frame.id })
}

Present in chat with screenshots.

If BACKEND = "html":

cd .design-catalogue && python -m http.server 8888 &
echo "Catalogue available at http://localhost:8888"

Open in Chrome MCP:

mcp__claude-in-chrome__navigate url="http://localhost:8888"
mcp__claude-in-chrome__computer action="screenshot"

Present to user:

Design Catalogue Ready

I've built [N] visual proposals exploring different directions for [project].

Quick overview:
1. Midnight Editorial - [soul statement]
2. Swiss Brutalist - [soul statement]
3. Warm Workshop - [soul statement]
...

[If Pencil] View the .pen file directly in your editor
[If HTML] Browse the catalogue: http://localhost:8888

Tell me which 2-3 resonate.

4. Collaborative Refinement

First selection:

AskUserQuestion:
"Which 2-3 directions interest you most?"
Options: [Proposal names with brief descriptions]

Refinement dialogue:

• "What specifically appeals about [selection]?"
• "Anything you'd change or combine?"
• "Should I generate hybrid proposals?"

If refinement requested:

• Generate 2-3 new proposals based on feedback
• Add to catalogue as new entries
• Update viewer to highlight refined options

5. Direction Selection

Present finalists with expanded detail:

• Full color palette (all semantic tokens)
• Complete typography scale with specimens
• Component transformation examples (before → after)
• Implementation priority list

Final selection:

AskUserQuestion:
"Which direction should we implement?"
Options: [Finalist names]
+ "Make more changes first"

5.5. Expert Panel Review (MANDATORY)

Before presenting any design to user, run expert panel review.

See: ui-skills/references/expert-panel-review.md

1. Simulate 10 world-class advertorial experts (Ogilvy, Rams, Scher, Wiebe, Laja, Walter, Cialdini, Ive, Wroblewski, Millman)
2. Each expert scores 0-100 and provides specific improvement feedback
3. Calculate average score
4. If average < 90: Implement highest-impact feedback, iterate, re-review
5. If average ≥ 90: Proceed to handoff

Expert Panel Review: [Design Name]

| Expert | Score | Critical Improvement |
|--------|-------|---------------------|
| Ogilvy | 88 | Headline needs stronger benefit |
| Rams | 92 | Clean execution |
| ...
**Average: 89.2** ❌ → Iterating...

Do not return design until 90+ average achieved.

6. Output & Handoff

Return selected direction:

## Selected Direction: [Name]

**DNA:** [layout, color, typography, motion, density, background]

**Typography:**
- Headings: [Font family, weights]
- Body: [Font family, weights]
- Code/Mono: [Font family]

**Color Palette:**
- Background: [hex]
- Foreground: [hex]
- Primary: [hex]
- Secondary: [hex]
- Accent: [hex]
- Muted: [hex]

**Implementation Priorities:**
1. [First change - highest impact]
2. [Second change]
3. [Third change]

**Preserve:**
- [What to keep from current design]

**Transform:**
- [What changes dramatically]

**Anti-patterns to avoid:**
- [Specific things NOT to do]

Handoff routing:

AskUserQuestion:
"Ready to generate implementation code?"
Options:
- "Yes, generate React components" → invoke handoff
- "No, I'll implement manually" → return spec only

Cleanup (HTML backend only):

# Stop local server
pkill -f "python -m http.server 8888"
# Optionally remove catalogue
# rm -rf .design-catalogue

Quick Reference

Integration

Invoked by:

• /design command (directly)
• /aesthetic at Phase 0 (recommended)
• /polish when maturity < 6 (suggested)

Outputs to:

• pencil-to-code — when Pencil backend, user wants code
• design-theme — when HTML backend, user wants implementation
• /aesthetic — guides all subsequent phases
• /polish — constrains DNA for refinement loop

Related Skills

• pencil-renderer — Pencil MCP rendering primitive
• pencil-to-code —.pen → React/Tailwind export
• aesthetic-system — DNA codes, anti-convergence rules
• design-tokens — Token system patterns
• design-theme — Theme implementation

References

• references/viewer-template.html - Catalogue viewer HTML
• references/proposal-template.html - Individual proposal HTML
• references/catalogue-template.md - Metadata structure
• aesthetic-system/references/dna-codes.md - DNA axis definitions
• aesthetic-system/references/banned-patterns.md - Anti-convergence rules