polanyi-design

"We can know more than we can tell."
— Michael Polanyi, The Tacit Dimension (1966)

A frontend design cognitive engine that surfaces the visual judgment

senior designers know but rarely articulate.

Not a role-play. Not a template generator.

An operational thinking system grounded in tacit knowledge theory.

Install · See the Difference · How It Works · Measured Impact · Credits

Same prompt. Different design intelligence.

Without Skill	With polanyi-design
Blue gradients, centered everything, Inter	Deep green, Playfair+Inter pairing, 60/40 editorial split

Both pages generated from the identical prompt: "设计一个软件服务公司的 landing page"

Live baseline · Live enhanced · Side-by-side comparison

📦 Install

Method 1: Global (works across all projects)

mkdir -p ~/.claude/skills/polanyi-design

curl -o ~/.claude/skills/polanyi-design/SKILL.md \
  https://raw.githubusercontent.com/SunflowersLwtech/polanyi-design/main/SKILL.md

Method 2: Project-local (this repo only)

mkdir -p .claude/skills/polanyi-design

curl -o .claude/skills/polanyi-design/SKILL.md \
  https://raw.githubusercontent.com/SunflowersLwtech/polanyi-design/main/SKILL.md

Method 3: Manual

1. Download SKILL.md
2. Place it at ~/.claude/skills/polanyi-design/SKILL.md (global) or .claude/skills/polanyi-design/SKILL.md (project)

Usage

Once installed, invoke directly:

/polanyi-design 设计一个登录页面

Or let it auto-activate when you ask design-related questions — the skill triggers on keywords like "design", "layout", "UI", "looks like a template", "feels off", etc.

Verify Installation

cat ~/.claude/skills/polanyi-design/SKILL.md | head -5
# Should show: ---
#              name: polanyi-design

👁 See the Difference

Same prompt, same model. One with the skill loaded, one without.

Prompt: "设计一个软件服务公司的 landing page"

Both pages were generated in fresh Claude Code sessions with identical prompts.
The only difference: one session had polanyi-design loaded.

View Baseline (no skill) · View Enhanced (with skill)

Key Design Decisions Only the Skill Produces

Fashion Brand Demo

Also included: a luxury fashion website (NOIR ATELIER) with the same A/B methodology:

View Baseline · View Enhanced

🧠 How It Works

Three systems fire in sequence:

System 1: Knowledge Filter

┌─────────────────────────────────┐
│  SKIP (Explicit)                │  Don't restate Tailwind docs.
├─────────────────────────────────┤
│  EXTRACT (Implicit) ★          │  Senior designer reasoning.
│  This is where you spend tokens │  The unspoken "why."
├─────────────────────────────────┤
│  PUSH (Tacit)                   │  "This feels off" → structural
│  Embodied · Holistic · Context  │  diagnosis + concrete fix.
└─────────────────────────────────┘

System 2: Five Design Lenses + Cross-Cutting

Cross-Cutting: Accessibility (focus-visible, contrast AA, reduced-motion) · Motion System (duration scale, easing, choreography) · Data Visualization Colors (categorical, sequential, divergent, colorblind-safe) · Quality Verification (axe-core, 200% zoom, VoiceOver)

Aesthetic Judgment Encoding

Six patterns from frontend-design research:

1. Negation > Assertion — Ban defaults (Inter, shadow-md, rounded-lg) rather than prescribing "good taste"
2. Polarization > Compromise — Commit to "brutally minimal" or "maximalist editorial." The middle is where templates live.
3. Intentionality > Intensity — "What makes this UNFORGETTABLE?" beats "What follows best practices?"
4. Anti-Convergence — Prevent AI from reproducing the same safe choices every time
5. Taxonomy > Rules — Aesthetic directions (editorial, brutalist, organic) instead of fixed rules
6. Multi-Layer Refinement — Direction → Domain → Tactic → Negation, all coherent

System 3: Output Protocol

[1-sentence design thesis — declared before any code]
[Concrete implementation — px, hex, font names]
[Tacit rationale — why, not just what]
[What NOT to do — prevents the most likely wrong choice]
[Escape hatch — when this advice shouldn't apply]

📊 Measured Impact

Evaluated via A/B testing across 5 design prompts (landing page diagnosis, dashboard layout,
token architecture, component API review, responsive strategy), each scored on 5 dimensions
(0-10 scale) by an independent Claude evaluator agent. The evaluator received both responses
without knowing which had the skill loaded. 3 rounds of iteration, self-evaluated — take
scores as directional, not peer-reviewed:

R1  ████████████████████████░░░░░░░░░░░░░░░░  +23%  (AC regression found)
R2  ████████████████████████████████░░░░░░░░░  +31%  (AC fixed, +3.0 swing)
R3  ███████████████████████████████░░░░░░░░░░  +30%  (production ready)

Theoretical Foundation

Built on Michael Polanyi's tacit knowledge framework (1958-1966), operationalized through:

Credits

This is an original work, not a fork. We gratefully acknowledge:

• enzyme2013/polanyi-skill — First public Polanyi skill, built with nuwa-skill
• Google Design MD (Stitch) — Structured design system spec format
• Michael Polanyi — Personal Knowledge (1958), The Tacit Dimension (1966)
• The Polanyi Society — polanyisociety.org

See references/CREDITS.md for full attribution.

License

MIT License — See LICENSE for full text.