Skill Integration

Skill Integration Skill

Standardized patterns for how agents discover, reference, and use skills effectively in Claude Code 2.0+.

When This Activates

• Working with agent prompts or skill references
• Implementing new agents or skills
• Understanding skill architecture
• Optimizing context usage
• Keywords: "skill", "progressive disclosure", "skill discovery", "agent integration"

Overview

The skill-integration skill provides standardized patterns for:

• Skill discovery: How agents find relevant skills based on task keywords
• Progressive disclosure: Loading skill content on-demand to prevent context bloat
• Skill composition: Combining multiple skills for complex tasks
• Skill reference format: Consistent way agents reference skills in prompts

Progressive Disclosure Architecture

What It Is

Progressive disclosure is a design pattern where:

1. Metadata stays in context - Skill names, descriptions, keywords (~50 tokens)
2. Full content loads on-demand - Detailed guidance only when needed (~5,000-15,000 tokens)
3. Context stays efficient - Support 50-100+ skills without bloat

Why It Matters

Without progressive disclosure:

• 20 skills × 500 tokens each = 10,000 tokens in context
• Context bloated before agent even starts work
• Can't scale beyond 20-30 skills

With progressive disclosure:

• 100 skills × 50 tokens each = 5,000 tokens in context
• Full skill content only loads when relevant
• Scales to 100+ skills without performance issues

How It Works

┌─────────────────────────────────────────────────────────┐
│                   Agent Context                          │
│                                                          │
│  Agent Prompt: ~500 tokens                              │
│  Skill Metadata: 20 skills × 50 tokens = 1,000 tokens  │
│  Task Description: ~200 tokens                          │
│                                                          │
│  Total: ~1,700 tokens (efficient!)                      │
└─────────────────────────────────────────────────────────┘
                          │
                          │ Agent encounters keyword
                          │ matching skill
                          ↓
┌─────────────────────────────────────────────────────────┐
│              Skill Content Loads On-Demand               │
│                                                          │
│  Skill Full Content: ~5,000 tokens                      │
│  Loaded only when needed                                │
│                                                          │
│  Total context: 1,700 + 5,000 = 6,700 tokens           │
│  Still efficient!                                        │
└─────────────────────────────────────────────────────────┘

Skill Discovery Mechanism

Keyword-Based Activation

Skills auto-activate when task keywords match skill keywords:

Example: testing-guide skill

---
name: testing-guide
keywords: test, testing, pytest, tdd, coverage, fixture
auto_activate: true
---

Task triggers skill:

• "Write tests for user authentication" → matches "test", "testing"
• "Add pytest fixtures for database" → matches "pytest", "fixture"
• "Improve test coverage to 90%" → matches "testing", "coverage"

Manual Skill Reference

Agents can explicitly reference skills in their prompts:

## Relevant Skills

You have access to these specialized skills:

- **testing-guide**: Pytest patterns, TDD workflow, coverage strategies
- **python-standards**: Code style, type hints, docstring conventions
- **security-patterns**: Input validation, authentication, OWASP compliance

Benefits:

• Agent knows which skills are available for its domain
• Progressive disclosure still applies (metadata in context, content on-demand)
• Helps agent make better decisions about when to consult specialized knowledge

Skill Composition

Combining Multiple Skills

Complex tasks often require multiple skills:

Example: Implementing authenticated API endpoint

Task: "Implement JWT authentication for user API endpoint"

Skills activated:
1. **api-design** - REST API patterns, endpoint structure
2. **security-patterns** - JWT validation, authentication best practices
3. **python-standards** - Code style, type hints
4. **testing-guide** - Security testing patterns
5. **documentation-guide** - API documentation standards

Progressive disclosure:
- All 5 skill metadata in context (~250 tokens)
- Full content loads only as needed (~20,000 tokens total)
- Agent accesses relevant sections progressively

Skill Layering

Skills can reference other skills:

## Relevant Skills

- **testing-guide**: Testing patterns (references python-standards for test code style)
- **security-patterns**: Security best practices (references api-design for secure endpoints)
- **documentation-guide**: Documentation standards (references python-standards for docstrings)

Benefits:

• Natural skill hierarchy
• Agent discovers related skills automatically
• No need to list every transitive dependency

Standardized Agent Skill References

Template Format

Every agent should include a "Relevant Skills" section:

## Relevant Skills

You have access to these specialized skills when [agent task]:

- **[skill-name]**: [Brief description of what guidance this provides]
- **[skill-name]**: [Brief description of what guidance this provides]
- **[skill-name]**: [Brief description of what guidance this provides]

**Note**: Skills load automatically based on task keywords. Consult skills for detailed guidance on specific patterns.

Best Practices

✅ Do's:

• List 3-7 most relevant skills for agent's domain
• Use consistent skill names (match SKILL.md name: field)
• Keep descriptions concise (one line)
• Add note about progressive disclosure
• Trust skill discovery mechanism

❌ Don'ts:

• List all 21 skills (redundant, bloats context)
• Duplicate skill content in agent prompt
• Provide detailed skill guidance inline
• Override skill content with conflicting guidance
• Assume skills are "just documentation"

Example: implementer Agent

## Relevant Skills

You have access to these specialized skills when implementing features:

- **python-standards**: Code style, type hints, docstring conventions
- **api-design**: REST API patterns, error handling
- **database-design**: Query optimization, schema patterns
- **testing-guide**: Writing tests alongside implementation
- **security-patterns**: Input validation, secure coding practices
- **observability**: Logging, metrics, tracing
- **error-handling-patterns**: Standardized error handling and recovery

**Note**: Skills load automatically based on task keywords. Consult skills for detailed guidance on specific patterns.

Token impact:

• Before: 500+ tokens of inline guidance
• After: 150 tokens referencing skills
• Savings: 350 tokens (70% reduction)

Token Reduction Benefits

Per-Agent Savings

Typical agent with verbose "Relevant Skills" section:

Before (verbose inline guidance):

## Relevant Skills

### Testing Patterns
- Use pytest for all tests
- Follow Arrange-Act-Assert pattern
- Use fixtures for setup
- Aim for 80%+ coverage
- [... 300 more words ...]

### Code Style
- Use black for formatting
- Add type hints to all functions
- Write Google-style docstrings
- [... 200 more words ...]

### Security
- Validate all inputs
- Use parameterized queries
- [... 150 more words ...]

Token count: ~500 tokens

After (skill references):

## Relevant Skills

You have access to these specialized skills when implementing features:

- **testing-guide**: Pytest patterns, TDD workflow, coverage strategies
- **python-standards**: Code style, type hints, docstring conventions
- **security-patterns**: Input validation, secure coding practices

**Note**: Skills load automatically based on task keywords. Consult skills for detailed guidance.

Token count: ~150 tokens

Savings: 350 tokens per agent (70% reduction)

Across All Agents

• 20 agents × 350 tokens saved = 7,000 tokens
• Plus: Skills themselves deduplicate shared guidance
• Result: 20-30% overall token reduction in agent prompts

Scalability

With inline guidance (doesn't scale):

• 20 agents × 500 tokens = 10,000 tokens
• Can't add more specialized guidance without bloating prompts
• Context budget limits agent capability

With skill references (scales infinitely):

• 20 agents × 150 tokens = 3,000 tokens
• Can add 100+ skills without impacting agent prompt size
• Progressive disclosure ensures context efficiency

Real-World Examples

Example 1: researcher Agent

Before:

## Relevant Skills

### Research Patterns
When researching, follow these best practices:
- Start with official documentation
- Check multiple sources for accuracy
- Document sources with URLs
- Identify common patterns across sources
- Note breaking changes and deprecations
- Verify information is current (check dates)
- Look for code examples and real-world usage
- [... 400 more words ...]

Token count: ~600 tokens

After:

## Relevant Skills

You have access to these specialized skills when researching:

- **research-patterns**: Web research methodology, source evaluation
- **documentation-guide**: Documentation standards for research findings

**Note**: Skills load automatically based on task keywords.

Token count: ~100 tokens

Savings: 500 tokens (83% reduction)

Example 2: planner Agent

Before:

## Relevant Skills

### Architecture Patterns
Follow these architectural patterns:
- [... 300 words ...]

### API Design
When designing APIs:
- [... 250 words ...]

### Database Design
For database schemas:
- [... 200 words ...]

### Testing Strategy
Plan testing approach:
- [... 200 words ...]

Token count: ~700 tokens

After:

## Relevant Skills

You have access to these specialized skills when planning:

- **architecture-patterns**: Design patterns, SOLID principles
- **api-design**: REST API patterns, versioning strategies
- **database-design**: Schema design, query optimization
- **testing-guide**: Test strategy, coverage planning

**Note**: Skills load automatically based on task keywords.

Token count: ~130 tokens

Savings: 570 tokens (81% reduction)

Detailed Documentation

For comprehensive skill integration guidance:

• Skill Discovery: See docs/skill-discovery.md for keyword matching and activation
• Skill Composition: See docs/skill-composition.md for combining skills
• Progressive Disclosure: See docs/progressive-disclosure.md for architecture details

Examples

• Agent Template: See examples/agent-skill-reference-template.md
• Composition Example: See examples/skill-composition-example.md
• Architecture Diagram: See examples/progressive-disclosure-diagram.md

Integration with autonomous-dev

All 20 agents in the autonomous-dev plugin follow this skill integration pattern:

• Each agent lists 3-7 relevant skills
• No inline skill content duplication
• Progressive disclosure prevents context bloat
• Scales to 100+ skills without performance issues

Result: 20-30% token reduction in agent prompts while maintaining full access to specialized knowledge.

Version: 1.0.0 Type: Knowledge skill (no scripts) See Also: agent-output-formats, documentation-guide, python-standards