AI Skill Manager

AI Skill Manager

Comprehensive guide for managing Claude Code skills throughout their complete lifecycle, from planning and creation to maintenance and deprecation.

When to Use This Skill

• Creating new skills - Plan structure, write descriptions, test discovery
• Updating existing skills - Detect outdated content, apply 2026 best practices
• Optimizing token usage - Implement progressive disclosure, measure efficiency
• Managing quality - Audit skills, enforce standards, track metrics
• Testing discovery - Verify trigger phrases, optimize keywords
• Deprecating skills - Plan transitions, archive safely, maintain backwards compatibility

Quick Reference

Creating a New Skill

# 1. Create structure
mkdir -p .claude/skills/my-skill/scripts

# 2. Write SKILL.md with frontmatter
cat > .claude/skills/my-skill/SKILL.md << 'EOF'
---
name: my-skill
description: [Action verbs] + [what it does] + [when to use] + [trigger keywords]
---

# My Skill

## When to Use
- Use case 1
- Use case 2

## Quick Reference
[Complete minimal example]

## Core Patterns
[Main implementation]
EOF

# 3. Test discovery
# Ask Claude: "What Skills are available?"
# Test trigger: Use keywords from description

Auditing an Existing Skill

# Run skill audit script
python .claude/skills/ai-skill-manager/scripts/skill-audit.py .claude/skills/my-skill/

# Check output for:
# - Line count (should be < 500)
# - Token usage (should be < 5k)
# - Description quality
# - Missing sections

Optimizing Token Usage

# Count tokens
python .claude/skills/ai-skill-manager/scripts/token-counter.py .claude/skills/my-skill/SKILL.md

# If > 500 lines, split into:
# - SKILL.md (overview + quick ref)
# - REFERENCE.md (detailed API docs)
# - EXAMPLES.md (comprehensive examples)
# - TROUBLESHOOTING.md (common issues)

Core Workflow

1. Planning Phase

Before writing any content:

Naming Patterns:

• ✅ database-migration-manager, flutter-query-testing, web-performance-metrics
• ❌ db-stuff, helpers, utils, tools

2. Writing Effective Descriptions

Description Formula:

[Action verbs] + [what it does] + [when to use] + [trigger keywords users would say]

Examples:

✅ GOOD:

description: "Extract text and tables from PDF files, fill forms, merge documents. Use when working with PDF files or when the user mentions PDFs, forms, or document extraction."

❌ BAD:

description: "Helps with documents"  # Too vague, no triggers

Trigger Keyword Strategy:

• Action verbs: create, analyze, optimize, validate, generate, test
• Domain nouns: database, migration, query, performance, RLS
• File extensions: .ts, .tsx, .sql, .md, .dart
• Error patterns: "RLS error", "query failure", "N+1 query"

3. Content Structure (SKILL.md)

---
name: skill-name
description: [Formula above]
version: "1.0.0"
last_updated: "YYYY-MM-DD"
---

# Skill Name

Brief overview (1-2 sentences)

## When to Use This Skill

- **Use case 1** - Description
- **Use case 2** - Description

## Quick Reference

```bash
# Essential commands with complete working examples
./command.sh

Core Workflow

Step 1: [Action]

Detailed instructions

Step 2: [Action]

More instructions

Common Patterns

Pattern 1: [Name]

[Code example]

Pattern 2: [Name]

[Code example]

Troubleshooting

Related Resources

• For complete API details, see REFERENCE.md
• For examples, see EXAMPLES.md
• Related skill: other-skill

### 4. Progressive Disclosure (for skills > 500 lines)

**File Structure:**

my-skill/ ├── SKILL.md # Overview + navigation (< 500 lines) ├── REFERENCE.md # Detailed API docs (loaded when needed) ├── EXAMPLES.md # Comprehensive examples (loaded when needed) ├── TROUBLESHOOTING.md # Common issues (loaded when needed) └── scripts/ # Utilities (execute, not loaded) └── helper.py

**In SKILL.md, reference supporting files:**
```markdown
## Core Patterns

[Essential patterns here - most common 80% use cases]

## Additional Resources

For complete API reference, see [REFERENCE.md](REFERENCE.md)
For detailed examples, see [EXAMPLES.md](EXAMPLES.md)
For troubleshooting, see [TROUBLESHOOTING.md](TROUBLESHOOTING.md)

Token Savings: 95-98% reduction by loading docs only when needed

5. Testing Discovery

Verification Steps:

1. Check availability:

   Ask Claude: "What Skills are available?"
   

2. Test trigger phrases:

   Use keywords from description naturally in a request
   Example: "Create a migration for adding email column"
   Should trigger: database-migration-manager
   

3. Verify invocation:

   • Claude should show skill confirmation
   • Instructions should load correctly
   • Examples should work as shown

4. Test script execution:

   ./scripts/test-discovery.sh my-skill
   

6. Quality Audit Checklist

Run before committing any skill:

• Description includes specific trigger keywords
• SKILL.md under 500 lines (or uses progressive disclosure)
• Code examples are tested and working
• Quick reference shows complete minimal example
• Troubleshooting section included
• Related skills/files documented
• YAML frontmatter is valid
• Scripts have execute permissions (chmod +x)
• Token usage < 5k tokens for full load
• Discovery tested with trigger phrases

Advanced Patterns

Skills with Utility Scripts

Zero-context execution - scripts run without loading contents:

## Validation

To validate your migration:
```bash
python scripts/validate-migration.py path/to/migration.sql

The script checks:

• Idempotency (IF NOT EXISTS/IF EXISTS)
• Naming conventions
• RLS policy patterns

**Script Best Practices:**
- Return structured JSON output
- Handle errors gracefully
- Provide actionable error messages
- Exit with proper status codes (0 = success, 1 = error)

### Skills with Tool Restrictions

```yaml
---
name: secure-operations
description: [...]
allowed-tools: Read, Bash(python:*), Grep  # Restrict to safe operations
---

Skills with Model Overrides

---
name: simple-formatting
description: [...]
model: claude-haiku-3-5-20250110  # Fast, cheap model for simple tasks
---

Skills with Hooks

---
name: safe-migrations
description: [...]
hooks:
  PreToolUse:
    - matcher: "Write"
      hooks:
        - type: command
          command: "./scripts/validate-before-write.sh"
          once: true
---

Maintenance Workflow

When to Update Skills

Indicators:

• Framework version upgrades (Next.js, Flutter, Riverpod, etc.)
• API changes or deprecations
• New best practices emerge
• Users report skill isn't working
• Token usage too high (> 5k tokens)
• Performance issues identified

Update Triggers:

Version Control

Use semantic versioning:

---
name: my-skill
version: "1.2.0"  # MAJOR.MINOR.PATCH
last_updated: "2026-01-12"
---

• MAJOR: Breaking changes to skill interface
• MINOR: New features, backward-compatible
• PATCH: Bug fixes, documentation updates

Maintain CHANGELOG.md:

# Changelog

## [1.2.0] - 2026-01-12
### Added
- New pattern for X
- Support for Y

### Changed
- Improved Z

### Fixed
- Bug in W

Deprecation Process

1. Mark as deprecated:

description: "[DEPRECATED] Use 'new-skill' instead. [Old description...]"

2. Update content to redirect:

# DEPRECATED: Old Skill Name

This skill is deprecated. Please use `/new-skill` instead.

See [new-skill](../new-skill/SKILL.md) for the replacement.

## Migration Guide
[How to transition from old to new]

3. Archive after transition (e.g., 1 sprint):

mkdir -p .claude/skills/_archived
mv .claude/skills/old-skill .claude/skills/_archived/

Optimization Strategies

Token Efficiency Techniques

1. Progressive Disclosure (95-98% token reduction)

   • SKILL.md: Overview + quick ref
   • Supporting files: Loaded only when needed

2. Use Tables Over Prose

   # ❌ Verbose
   When you need auth, use withAuth. For user context, use withAuthParams.
   
   # ✅ Scannable
   | Wrapper | Use When |
   |---------|----------|
   | withAuth | Need authentication |
   | withAuthParams | Need user context |
   

3. Script-Based Execution (zero-context)

   • Only script output consumes tokens
   • Script contents don't load into context

4. Template Reuse

   Use templates in `resources/templates/`:
   - `migration-template.sql` - Copy and modify
   

Content Strategy

Few-Shot Learning:

## Examples

### Example 1: Common Use Case
**Input**: [User request]
**Expected**: [Desired output]
**Code**:
```code
[Complete example]

Example 2: Edge Case

[Same structure]

**Anti-Pattern Documentation:**
```markdown
## Common Mistakes

### ❌ Don't Do This
[Bad example with explanation]

### ✅ Do This Instead
[Good example with explanation]

Quality Metrics

Measuring Skill Effectiveness

Key Metrics:

• Discovery rate: How often Claude finds skill when appropriate (target: > 90%)
• Success rate: How often skill instructions work correctly (target: > 95%)
• Token efficiency: Cost per successful use (target: < 5k tokens)
• User satisfaction: Team feedback (track informally)

Monitoring:

# Count skill invocations
grep "skill-name" ~/.claude/projects/*/transcript.jsonl | wc -l

# Check token usage
python scripts/token-counter.py .claude/skills/skill-name/SKILL.md

Troubleshooting

Related Resources

Detailed Guides (Progressive Disclosure)

For comprehensive workflows and examples, see:

• CREATION_GUIDE.md - Detailed skill creation workflow with examples
• MAINTENANCE_GUIDE.md - Version control, updates, deprecation
• OPTIMIZATION_GUIDE.md - Token efficiency, performance tuning
• BEST_PRACTICES.md - Industry standards, Claude Code patterns, examples library

Utility Scripts

Run from project root:

# Audit existing skill
python .claude/skills/ai-skill-manager/scripts/skill-audit.py .claude/skills/my-skill/

# Count tokens
python .claude/skills/ai-skill-manager/scripts/token-counter.py .claude/skills/my-skill/SKILL.md

# Test discovery
./.claude/skills/ai-skill-manager/scripts/test-discovery.sh my-skill

Official Documentation

• Claude Code Skills Documentation
• Skill Authoring Best Practices
• Claude Code Hooks Reference

Related Skills

• user-stories-manager - Track feature development
• wip-lifecycle-manager - Manage work-in-progress documentation
• code-quality-tools - Automated quality fixes

Success Criteria

A well-crafted skill should:

✅ Be discovered automatically when relevant (> 90% discovery rate) ✅ Provide working examples that execute successfully ✅ Stay under 500 lines or use progressive disclosure ✅ Consume < 5k tokens when fully loaded ✅ Include troubleshooting for common issues ✅ Reference related skills and documentation ✅ Be versioned and maintained over time ✅ Follow 2026 industry best practices

Last Updated: 2026-01-12 Version: 1.0.0 Maintainer: Ballee Engineering Team