creating-skills（创造技能）

Creating Skills

This skill helps you create new Claude Agent Skills following Anthropic's official specifications and best practices.

This skill includes comprehensive resources:

• BEST_PRACTICES.md: Detailed authoring guidelines
• REFERENCE.md: Technical specifications and detailed examples
• README.md: Overview and quick start guide
• templates/: Ready-to-use skill templates
• examples/: Sample skills for reference
• scripts/validate_skill.py: Skill validation tool

When to Use This Skill

Invoke this skill when the user:

• Asks to create a new Claude skill
• Wants to generate a skill template
• Needs help structuring a custom capability
• Requests skill scaffolding or boilerplate
• Wants to validate an existing skill

Skill Creation Workflow

Step 1: Gather Requirements

Ask the user for:

1. Skill name: What should the skill be called?

- Lowercase letters, numbers, hyphens only - Maximum 64 characters - Use gerund form (e.g., "processing-data", "analyzing-logs") - Avoid vague names like "helper" or "utils"

1. Skill description: What does the skill do and when should it be used?

- Maximum 1024 characters - Write in third person - Include specific triggers - Format: "[What it does]. Use when [conditions]."

1. Skill complexity: Simple or complex?

- Simple: Single SKILL.md file (under 300 lines) - Complex: SKILL.md + REFERENCE.md + FORMS.md + scripts

1. Core functionality: What are the main tasks?

Step 2: Choose and Copy Template

Simple Template (for focused, single-file skills):

cp -r templates/simple-skill-template/ ../your-skill-name/

Complex Template (for multi-file skills with extensive docs):

cp -r templates/complex-skill-template/ ../your-skill-name/

See REFERENCE.md section "Template Selection Guide" for detailed criteria.

Step 3: Customize the Template

Edit SKILL.md frontmatter:

---
name: your-skill-name
description: What it does. Use when triggers.
---

Fill in required sections:

1. When to Use This Skill
2. Core functionality with steps
3. Common patterns
4. Error handling
5. Examples
6. Validation checklist

For Complex Skills:

• Edit REFERENCE.md for technical details
• Edit FORMS.md for output templates
• Create scripts for automation

See REFERENCE.md section "Customization Guide" for detailed instructions.

Step 4: Validate the Skill

Run the validation script:

python scripts/validate_skill.py ../your-skill-name/SKILL.md --strict

The validator checks:

• YAML frontmatter syntax and fields
• Name format (lowercase-with-hyphens, ≤64 chars, gerund form)
• Description (≤1024 chars, includes triggers)
• No reserved words ("anthropic", "claude")
• File structure and references
• Best practices compliance

See REFERENCE.md section "Validation Tool" for detailed usage.

Step 5: Review Best Practices

Before finalizing:

cat BEST_PRACTICES.md

Key principles:

• Conciseness: Only include non-standard information
• Progressive Disclosure: Keep SKILL.md under 500 lines
• Freedom Levels: Match specificity to task fragility
• Consistent Terminology: Use same terms throughout
• Validation Steps: Include checklists for complex workflows

See BEST_PRACTICES.md for complete guidelines.

Frontmatter Requirements

Every SKILL.md must start with:

---
name: skill-name
description: Clear description. Use when triggers.
---

Quick Rules:

• name: lowercase-with-hyphens, ≤64 chars, gerund form
• description: ≤1024 chars, includes "Use when..."
• No XML tags or reserved words

See REFERENCE.md section "Frontmatter Specifications" for complete rules and examples.

Progressive Disclosure

Skills load in three levels:

Level 1 - Metadata (~100 tokens, always loaded):

• YAML frontmatter for skill discovery

Level 2 - Instructions (<5,000 tokens, triggered):

• Main SKILL.md content

Level 3 - Resources (on-demand):

• REFERENCE.md, FORMS.md, scripts
• Load only when referenced

Example:

## Basic Processing
[Instructions for common case]

## Advanced Techniques
See REFERENCE.md section "Advanced Methods" for details.

Example Skills

Simple Skill: Code Reviewer (examples/simple-skill-example/)

• Single SKILL.md file
• Clear workflow with checklists
• ~350 lines

Complex Skill: Data Analyzer (examples/complex-skill-example/)

• SKILL.md + REFERENCE.md + FORMS.md + scripts
• Statistical methods in REFERENCE.md
• Report templates in FORMS.md

View examples:

cat examples/simple-skill-example/SKILL.md
cat examples/complex-skill-example/SKILL.md

Quick Start

1. Choose template based on complexity:

• Simple: Single focused task, <300 lines
• Complex: Multiple features, needs extensive docs

2. Copy template:

cp -r templates/simple-skill-template/ ../my-skill/

3. Edit frontmatter and fill sections:

• Replace all [placeholders]
• Add specific examples
• Create validation checklist

4. Validate:

python scripts/validate_skill.py ../my-skill/SKILL.md --strict

5. Review BEST_PRACTICES.md and test

Validation Checklist

Before using a new skill:

Frontmatter:

• Valid name (gerund form, ≤64 chars)
• Description with triggers (≤1024 chars)
• No prohibited content

Content:

• "When to Use This Skill" section
• Core functionality with steps
• Examples included
• Validation checklist
• Concise (no unnecessary info)

Structure:

• Progressive disclosure applied
• SKILL.md under 500 lines (or split)
• Referenced files exist

Testing:

• Passed validator
• Functionally tested
• Works as expected

Common Patterns

Pattern 1: Simple Single-File Skill

• Use simple-skill-template
• Focus on one capability
• Include 2-3 examples
• Add validation checklist

Pattern 2: Complex Multi-File Skill

• Use complex-skill-template
• SKILL.md: High-level workflow with references
• REFERENCE.md: Technical details, algorithms
• FORMS.md: Output templates
• scripts/: Automation utilities

Pattern 3: Skill with Automation

• Include scripts/ directory
• Document script usage in SKILL.md
• Scripts execute without loading to context
• Use only pre-installed packages

Runtime Constraints

Remember when designing skills:

• ❌ No network access or external API calls
• ❌ No runtime package installation
• ✅ Only pre-installed packages
• ✅ Scripts execute via bash without context loading
• ✅ Progressive disclosure minimizes context usage

Error Handling

Common Issue: Validation Fails

• Check YAML syntax
• Verify name format (lowercase-with-hyphens)
• Ensure description includes "Use when..."
• Remove any reserved words

Common Issue: Skill Too Long

• Split into SKILL.md + REFERENCE.md
• Move technical details to REFERENCE.md
• Move templates to FORMS.md
• Keep SKILL.md under 500 lines

Common Issue: References Not Found

• Ensure referenced files exist
• Use relative paths
• Check file names match exactly

Additional Resources

Internal:

• REFERENCE.md: Technical specs, detailed examples, troubleshooting
• BEST_PRACTICES.md: Complete authoring guidelines
• README.md: Quick start and overview
• templates/: Ready-to-use templates
• examples/: Working sample skills

External:

• Claude Skills Documentation
• Skills Cookbook

Getting Help

Review templates:

cat templates/simple-skill-template/SKILL.md
cat templates/complex-skill-template/SKILL.md

Study examples:

cat examples/simple-skill-example/SKILL.md
cat examples/complex-skill-example/SKILL.md

Read best practices:

cat BEST_PRACTICES.md

Check technical specs:

cat REFERENCE.md

Validate your skill:

python scripts/validate_skill.py ../your-skill/SKILL.md --strict

For detailed technical specifications, troubleshooting, and comprehensive examples, see REFERENCE.md.