From Hero to System

Turning Human Knowledge
into AI Skills

A practical guide for design architects

February 2026 · Prepared by Bob for Ako

The "Hero Knowledge" Problem

Every team has unwritten expertise trapped in one person's head.

🧠 Expert knows
"use red-600 for errors"
🏖️ Expert goes
on vacation
🤖 AI uses
wrong red

Heroes don't scale. AI can — if you structure the knowledge.

The Core Insight

"The teams winning at AI aren't the ones with
the best models. They're the ones who've
packaged their knowledge best."

What Are Claude Skills?

Modular instruction packages that teach Claude
how to handle specific tasks or domains.

📦 Package
expertise once
🔍 Claude
auto-discovers
✨ Consistent
output every time

Think of them as "onboarding docs for AI"

Anatomy of a Skill


my-brand-skill/
├── SKILL.md           # Core instructions (required)
├── REFERENCE.md       # Supplemental info (optional)
├── resources/
│   ├── helper.py      # Executable scripts
│   └── examples.json  # Reference data
  

Minimum: one folder + one SKILL.md file

SKILL.md In Practice

---
name: brand-guidelines
description: >
  Apply company brand guidelines for marketing
  copy, presentations, and customer-facing content.
dependencies:
  - python3
---

# Brand Guidelines Skill

## Voice & Tone
- Professional but approachable
- Active voice preferred

## Color Usage
- Primary: #1A73E8 (trust blue)
- Error: #DE3E25 (destructive only)
- Never use raw hex — use semantic tokens

## ✅ Good
"We've updated your dashboard with new insights."

## ❌ Bad
"The UI has been refactored with enhanced data viz."
  

Smart Loading: 3 Tiers

TierWhat Claude SeesWhen
Metadata Name + description Always — every conversation
Full Instructions Entire SKILL.md body Only when relevant
Resources Scripts, data, refs Only when executing

10 skills loaded → only 1 fully expanded → token-efficient

Skills ≠ Output Styles

🎨 Styles

  • Modify how Claude responds
  • Tone, formatting, structure
  • Always active once selected

⚡ Skills

  • Modify what Claude can do
  • Domain expertise, workflows
  • On-demand based on context

The Convergence

Everyone is solving the same problem — different shapes, same DNA.

Claude Skills Custom GPTs Cursor Rules AGENTS.md
(GitHub Copilot)
Figma MCP
FormatSKILL.md + folderWeb UI config.mdc filesPlain markdownAuto-generated
ActivationAuto by contextUser selectsGlob / AI-decidedAlways loadedMCP protocol
Progressive✅ 3-tier⚠️ Glob
Executable✅ Scripts
Composable

Composable = can mix & combine multiple skills/rules · AGENTS.md = GitHub Copilot's instruction file format

Choosing Your Format

📄 Consistent document output?Claude Skills (scripts for PDF, Excel)

💬 Customer-facing assistant?Custom GPTs (easy sharing, web UI)

🖥️ Coding conventions?Cursor Rules (auto-attach by file type)

🎨 Design system compliance?Figma MCP + token docs

🏢 Team process capture?AGENTS.md (simplest approach)

They're not mutually exclusive — combine as needed

Structuring Knowledge for AI

AI has zero tacit knowledge. Everything must be explicit.

Knowledge TypeExampleEncoding
Rules"Error text = red-600"Explicit constraints
Patterns"Cards have 16px padding"Templates + examples
Preferences"Prefer shorter headlines"Ranked examples ✅/❌
Judgment"Layout feels too dense"Decision trees
Process"PRs need design review"Step-by-step workflows

The Power of Anti-Patterns

AI learns almost as much from what NOT to do.


## Common Mistakes (DO NOT)
- ❌ Using primitive color tokens directly (e.g., `red-6`)
- ❌ Hardcoding pixel values instead of spacing tokens
- ❌ Creating new variants when an existing one works
- ❌ Skipping loading/error states in components
  

The "new hire" test: If a smart new hire could follow
your instructions and produce 90%+ acceptable output,
your instructions are good enough for AI.

Tokens Without Intent

🚫 What AI sees


{
  "red-6": {
    "value": "#DE3E25",
    "type": "color"
  }
}
      

No idea that red-6 is a primitive,
not for direct use

✅ What AI needs


{
  "color-feedback-error": {
    "value": "#DE3E25",
    "description": "Error states:
      alerts, destructive buttons",
    "usage": ["Alert.error",
      "Button.destructive"],
    "pairedWith": [
      "color-feedback-errorText"]
  }
}
      

Design System → AI Bridge

Layer 4 · Usage Rules & Guardrails
"Buttons must use semantic tokens"
Layer 3 · Component Specifications
Props, variants, composition rules
Layer 2 · Semantic Tokens
color-feedback-error, spacing-card-padding
Layer 1 · Primitive Tokens
red-6, space-4, font-size-14

AI-readability starts at Layer 2 · Guardrails live at Layer 4

Component Spec Example

# Component: Button

## Variants
- `primary`: Main actions (1 per section)
- `secondary`: Supporting actions
- `destructive`: Delete, irreversible actions

## Token Usage
| Property     | Token                         |
|------------- |-------------------------------|
| Background   | `color-interactive-primary`   |
| Text         | `color-text-onColor`          |
| Padding      | `spacing-component-btn-pad`   |
| Border radius| `radius-component-button`     |
| Min height   | `sizing-touch-target-min`     |

## Rules
- NEVER use raw hex values
- ALWAYS pair with appropriate text color token
- Maximum 1 primary button per section
- Destructive buttons require confirmation
  

Figma MCP: The Bridge

Connects your Figma designs directly to your IDE's AI agent

🎨 Figma
Components
🔌 MCP
Server
🤖 AI Agent
(Cursor / Claude)
💻 Code using
YOUR components

📦 Sends: Components, styles, variables, Code Connect

🎯 Enables: AI uses your components, not generic ones

Killer feature: Auto-generates design system rules from your codebase

7 Principles of AI Knowledge

1. Be explicit, not implicit

2. Structure beats prose (tables & lists > paragraphs)

3. Include examples — good AND bad

4. State what NOT to do

5. Use progressive disclosure

6. Test with edge cases

7. Version and iterate

"Make it look professional" means nothing.
"16px body, 48px headers, neutral-700" means everything.

Knowledge Capture Process

👀 Observe
the expert
🔍 Extract
patterns
📝 Encode
as rules
🧪 Test with
AI

5 extraction questions:

  1. "Walk me through the last time you did [task]."
  2. "What's the most common mistake juniors make?"
  3. "One rule only — what is it?"
  4. "When do you break your own rules?"
  5. "Show me a great example and a terrible one."

Real-World Success Stories

🚀 Trigger.dev — Cursor rules for task writing
Explicit SDK version + full examples → dramatically improved consistency

🎨 Figma MCP — Design system context for AI agents
AI reuses existing components instead of generating generic code

🏗️ Builder.io — Context-aware design-to-code
AI learns from your actual repo and uses your components

🤖 OpenClaw / AGENTS.md — Agent self-knowledge
Plain markdown in project root — simple but effective

🎯 Romina Kavcic — AI-readable design tokens
Added descriptions + usage arrays → Claude picked the right tokens automatically

Your 5-Phase Roadmap

📋 Phase 1: Audit (Day 1)

  • Top 20 tokens
  • Top 10 components
  • 5 "unwritten rules"

📝 Phase 2: Describe (Days 2-3)

  • Add description fields
  • Add usage & pairedWith
  • Write component specs

🛡️ Phase 3: Guardrail (Days 4-5)

  • Usage rules
  • Anti-pattern examples
  • Decision trees

📦 Phase 4: Package (Week 2)

  • Choose format(s)
  • Progressive disclosure
  • Set up Figma MCP

🔄 Phase 5: Test & Iterate — ongoing, review monthly

The Minimum Viable Skill

If you only do ONE thing, create this file:


# Design System Quick Reference for AI

## Tokens (Use These, Not Raw Values)
- Primary action: `color-interactive-primary` (#1A73E8)
- Error: `color-feedback-error` (#DE3E25)
- Body text: `color-text-primary` (#1F2937)
- Spacing: sm (8px) · md (16px) · lg (24px)
- Border radius: `radius-default` (8px)

## Component Rules
- Use `

90% of the value · 10% of the effort

Start Today: 15-Minute Win

Add description to your top 10 most-used tokens:

  1. Primary button color
  2. Error color
  3. Text primary color
  4. Background color
  5. Border color
  6. Button padding
  7. Input padding
  8. Border radius (default)
  9. Focus ring color
  10. Disabled state opacity

That's it. You've just made your design system AI-readable.

The Universal Pattern

All approaches converge on four pillars:

📋
Structured
Metadata
name, triggers
📝
Markdown
Instructions
rules, examples
📦
Reference
Resources
data, scripts
🔍
Progressive
Disclosure
load on demand

🔗 Emerging standard: agentskills.io

Key Takeaways

1️⃣  Heroes don't scale — package the knowledge

2️⃣  Be explicit — AI has zero tacit understanding

3️⃣  Anti-patterns are as valuable as patterns

4️⃣  Semantic tokens with descriptions = AI-readable design

5️⃣  Figma MCP is the highest-leverage integration

6️⃣  Start with one file — iterate from there

Further Reading

📘 Anthropic: How to Create Skills for Claude

📗 Figma: Design Systems and AI — Why MCP Servers Are the Unlock

📙 Design Tokens That AI Can Actually Read — Romina Kavcic

📕 Trigger.dev: How to Write Great Cursor Rules

📓 Agent Skills Specification — Emerging cross-platform standard


Thank you! 🙏

Research & presentation by Bob · February 2026