skill-optimizer（技能优化器）

Skill Optimizer

Purpose

Optimize existing Claude Code skills to minimize token consumption by leveraging the three-level content loading architecture and progressive disclosure patterns.

When to Use

Use this skill when:

• Analyzing existing skills for optimization opportunities
• Skills are too large (approaching or exceeding 500 lines)
• Need to reduce context consumption
• Restructuring skills for progressive disclosure
• Converting monolithic skills to multi-file architecture
• Optimizing YAML frontmatter for better discovery
• Improving skill performance and load times
• Auditing skills for token efficiency
• Creating new skills with optimization in mind

Content Loading Architecture

Three-Level Loading System

Level 1: Metadata (Always Loaded - ~100 tokens/skill)

• YAML frontmatter in SKILL.md
• Loaded at startup into system prompt
• Enables skill discovery without context overhead
• Optimization Target: Description field (max 1024 chars)

Level 2: Instructions (Loaded When Triggered - <5,000 tokens)

• Main SKILL.md content
• Loads dynamically when skill is relevant
• Contains workflows, best practices, guidance
• Optimization Target: Keep under 500 lines

Level 3: Resources (Loaded As Needed - Variable)

• Additional markdown files (REFERENCE.md, EXAMPLES.md, etc.)
• Code scripts in scripts/ directory
• Templates, schemas, documentation
• Optimization Target: No penalty until accessed

Key Principle

"Files don't consume context until accessed" - Bundle comprehensive documentation in reference files without context penalty.

Quick Optimization Workflow

1. Analyze Current State

Check skill size:

wc -l .claude/skills/skill-name/SKILL.md

Identify optimization opportunities:

• SKILL.md > 500 lines?
• Detailed API docs in main file?
• Extensive examples in main file?
• Long reference tables or schemas?
• Code snippets that could be scripts?
• Multiple distinct topics/sections?

2. Apply 500-Line Rule

If SKILL.md > 500 lines:

✅ Keep in main file:

• Purpose and when to use
• Quick start / common workflows
• Critical best practices
• Brief examples (5-10 lines)
• Cross-references to detailed files

❌ Move to reference files:

• Comprehensive API documentation
• Extensive code examples (>20 lines)
• Detailed troubleshooting guides
• Pattern libraries
• Schema definitions
• Long reference tables

3. Structure Reference Files

Create organized reference hierarchy:

skill-name/
├── SKILL.md              # <500 lines, workflows & quick ref
├── REFERENCE.md          # Comprehensive documentation
├── EXAMPLES.md           # Detailed code examples
├── PATTERNS.md           # Pattern library
├── TROUBLESHOOTING.md    # Debug guide
└── scripts/              # Executable utilities
    └── helper.sh

Add table of contents to files >100 lines

4. Optimize YAML Frontmatter

Description optimization (max 1024 chars):

✅ Include:

• All trigger keywords and phrases
• Use cases and scenarios
• File types and technologies
• Common action verbs
• Related concepts

❌ Avoid:

• Generic descriptions
• Missing key terms
• Overly verbose explanations

Example:

---
name: database-development
description: Database development guidance for PostgreSQL including schema design, migrations, RLS policies, indexing strategies, privacy-preserving patterns, query optimization, and Prisma ORM. Use when working with tables, columns, indexes, migrations, RLS, Row-Level Security, database schema, SQL queries, Prisma schema, database optimization, privacy architecture, or PostgreSQL best practices.
---

5. Implement Progressive Disclosure

Pattern:

## Topic Overview

Brief explanation (2-3 sentences).

**Key Points:**
- Important consideration 1
- Important consideration 2

**For detailed information**: [REFERENCE.md](REFERENCE.md#topic-details)

**Quick Example:**
\`\`\`typescript
// Minimal working example (5-10 lines)
\`\`\`

**For more examples**: [EXAMPLES.md](EXAMPLES.md#topic-examples)

Optimization Patterns Summary

Pattern 1: Extract API Documentation

Strategy: Move detailed API docs to REFERENCE.md

Before: 80+ lines of API documentation in SKILL.md After: 10-line summary with link to complete docs Savings: ~70 lines (~1,400 tokens)

For complete pattern details: REFERENCE.md

Pattern 2: Extract Pattern Libraries

Strategy: Move code patterns to PATTERNS.md

Before: 200+ lines of pattern code in SKILL.md After: 18-line summary with quick example Savings: ~182 lines (~3,640 tokens)

For complete pattern details: REFERENCE.md

Pattern 3: Extract Troubleshooting

Strategy: Move debug guides to TROUBLESHOOTING.md

Before: 300+ lines of troubleshooting in SKILL.md After: 18-line summary with quick diagnostics Savings: ~282 lines (~5,640 tokens)

For complete pattern details: REFERENCE.md

Pattern 4: Convert Code to Scripts

Strategy: Move executable code to scripts/ directory

Before: 55+ lines of bash script in SKILL.md After: 7-line reference to executable script Savings: ~48 lines (~960 tokens) + script code never enters context

For complete pattern details: REFERENCE.md

Common Anti-Patterns

Avoid these common mistakes when creating or optimizing skills:

❌ Anti-Pattern 1: Monolithic Skills

• Single SKILL.md with 1000+ lines
• Solution: Split into main + reference files

❌ Anti-Pattern 2: Incomplete References

• Reference files exist but not linked
• Solution: Link all reference files from SKILL.md

❌ Anti-Pattern 3: Nested References

• References pointing to other references (>1 level)
• Solution: Keep hierarchy flat (max 1 level)

❌ Anti-Pattern 4: Sparse Frontmatter

• Minimal YAML description missing keywords
• Solution: Rich description with all triggers

❌ Anti-Pattern 5: Code as Documentation

• 100+ line scripts embedded in markdown
• Solution: Move to scripts/ directory

For detailed anti-patterns and solutions: REFERENCE.md

Migration Workflow

Quick Migration Steps

Phase 1: Discovery

# Check skill size
wc -l .claude/skills/skill-name/SKILL.md

# Identify sections to extract
grep "^##" .claude/skills/skill-name/SKILL.md

Phase 2: Planning

• Design file hierarchy (SKILL.md + reference files)
• Plan content distribution
• Identify scripts to extract

Phase 3: Implementation

# Create reference files
touch REFERENCE.md EXAMPLES.md
mkdir -p scripts

# Extract content systematically
# 1. Copy section to reference file
# 2. Replace in SKILL.md with summary + link
# 3. Verify link works
# 4. Remove detailed content from SKILL.md

Phase 4: Optimization

• Trim remaining content
• Optimize frontmatter with trigger keywords
• Add navigation aids (table of contents)

Phase 5: Validation

# Verify under 500 lines
wc -l .claude/skills/skill-name/SKILL.md

# Test links work
grep -o '\[.*\](.*\.md#.*)' SKILL.md

For complete migration workflow: REFERENCE.md

Advanced Techniques

Quick reference to advanced optimization strategies:

Technique 1: Conditional Content Loading

• Structure content so advanced sections load only when needed
• Benefits: Most users don't load advanced content

Technique 2: Layered Examples

• Progressive complexity: minimal → production → enterprise
• Benefits: Beginners see simple examples, advanced users access complex ones

Technique 3: Executable Documentation

• Scripts that both execute and document
• Benefits: Zero token cost, self-documenting output

Technique 4: Tabular Compression

• Use tables to compress structured data
• Benefits: 60%+ space reduction for configurations/options

Technique 5: Smart Chunking

• Group related small sections instead of individual extraction
• Benefits: Better narrative flow, fewer cross-references

For detailed advanced techniques: REFERENCE.md

Optimization Checklist

When optimizing a skill, verify:

Content Structure

• SKILL.md is under 500 lines
• Main file contains quick reference only
• Detailed docs moved to reference files
• Reference files have table of contents (if >100 lines)
• Cross-references use relative links
• No deeply nested references (max 1 level)

YAML Frontmatter

• Description includes all trigger keywords
• Description is under 1024 characters
• Description covers use cases and scenarios
• Description mentions file types/technologies
• Name follows kebab-case convention

Progressive Disclosure

• Overview → Details pattern used
• Quick examples in main file (5-10 lines)
• Extensive examples in EXAMPLES.md
• Brief summaries with references to details
• Common workflows highlighted in main file

File Organization

• Reference files named clearly
• Scripts in scripts/ directory
• Scripts are executable (chmod +x)
• No redundant content across files
• Each file has single, clear purpose

Token Efficiency

• Eliminated verbose explanations
• Removed duplicate information
• Used bullet points vs paragraphs
• Moved large code blocks to reference files
• Converted reusable code to scripts

Measurement & Validation

Token Estimation

Quick estimate:

lines=$(wc -l < SKILL.md)
tokens=$((lines * 20))  # Conservative: 20 tokens/line
echo "Estimated tokens: ~$tokens"

Target: Keep SKILL.md under 10,000 tokens (~500 lines)

Before/After Comparison

# Calculate savings
BEFORE=850  # lines before optimization
AFTER=420   # lines after optimization
SAVINGS=$((BEFORE - AFTER))
TOKEN_SAVINGS=$((SAVINGS * 20))

echo "Reduced by $SAVINGS lines"
echo "Estimated token savings: ~$TOKEN_SAVINGS tokens"

Quality Checks

• All original information preserved
• Links work correctly
• Main file comprehensive for common cases
• Reference files well-organized
• Navigation intuitive

For detailed measurement methods: REFERENCE.md

Best Practices Summary

✅ DO:

• Keep SKILL.md minimum but use other reference files to keep the full and detailed knowledge base
• Use progressive disclosure (overview → details)
• Include all trigger keywords in description
• Create clear cross-references to detailed docs
• Add table of contents to reference files >100 lines
• Convert reusable code to scripts
• Test optimization with real usage

❌ DON'T:

• Exceed 500 lines in SKILL.md
• Nest references more than 1 level deep
• Include API docs in main file
• Embed long code examples in main file
• Create reference files without linking them
• Use sparse YAML descriptions
• Optimize without preserving critical info

Quick Reference

File size limits:

• SKILL.md: <500 lines (strict)
• Reference files: No limit (loaded on-demand)
• YAML description: 1024 chars max

Optimization priority:

1. Apply 500-line rule to SKILL.md
2. Extract API docs to REFERENCE.md
3. Move examples to EXAMPLES.md or PATTERNS.md
4. Convert scripts to scripts/ directory
5. Enrich YAML frontmatter description

Common extractions:

• API documentation → REFERENCE.md
• Code examples → EXAMPLES.md
• Troubleshooting → TROUBLESHOOTING.md
• Pattern library → PATTERNS.md
• Scripts → scripts/ directory

Typical token savings:

• API extraction: ~1,400 tokens
• Pattern library: ~3,640 tokens
• Troubleshooting: ~5,640 tokens
• Scripts: ~960 tokens + no code in context

For comprehensive optimization patterns and detailed workflows: REFERENCE.md

Real-World Example

Before optimization:

skill-example/
└── SKILL.md  (850 lines, ~17,000 tokens)

After optimization:

skill-example/
├── SKILL.md              (420 lines, ~8,400 tokens)
├── REFERENCE.md          (350 lines, loaded on-demand)
├── EXAMPLES.md           (180 lines, loaded on-demand)
└── scripts/
    ├── validate.sh       (code never enters context)
    └── setup.sh          (code never enters context)

Result: 50% token reduction on initial load, comprehensive docs still available on-demand

Next Steps:

1. Audit existing skills: wc -l.claude/skills/*/SKILL.md
2. Identify candidates for optimization (>500 lines)
3. Apply optimization workflow
4. Measure token savings
5. Validate with real usage

For detailed guidance: See REFERENCE.md for complete optimization patterns, anti-patterns, migration workflows, and advanced techniques.