creating-new-skills（创造新技能）

Purpose

Create new Agent Skills for AI Agents by following a structured workflow based on best practices and comprehensive documentation.

Instructions

This skill helps you create new Agent Skills for Claude Code. Before starting, read the comprehensive documentation files in the docs/ directory for complete context.

Prerequisites

Required Reading - Read these files in order before creating a skill:

1. docs/claude_code_agent_skills.md - Complete guide to creating and managing skills
2. docs/claude_code_agent_skills_overview.md - Architecture and how skills work
3. docs/blog_equipping_agents_with_skills.md - Design principles and best practices

Understanding Skills

What is a Skill?

• A directory containing a SKILL.md file with YAML frontmatter
• Instructions that Claude loads on-demand when relevant
• Optional supporting files (scripts, documentation, templates)
• Like an onboarding guide for a new team member

Three Sources of Skills:

1. Personal Skills: ~/.claude/skills/ - Available across all your projects
2. Project Skills: .claude/skills/ - Shared with team via git
3. Plugin Skills: Bundled with Claude Code plugins, automatically available when installed

Progressive Disclosure (3 Levels):

1. Metadata (always loaded): name and description in YAML frontmatter
2. Instructions (loaded when triggered): Main body of SKILL.md
3. Resources (loaded as needed): Additional files, scripts, templates

Key Principle: Only relevant content enters the context window at any time.

Skill Creation Workflow

Step 1: Define the Skill's Purpose

Ask the user these questions:

1. What task or domain should this skill cover?
2. When should Claude use this skill? (triggers)
3. What expertise or workflows need to be captured?
4. Does it need scripts, templates, or other resources?

Document the answers for reference.

Step 2: Create the Skill Directory Structure

Create skills in the project's .claude/skills/ directory for team sharing:

mkdir -p .claude/skills/<skill-name>

Naming conventions:

• Use gerund form (verb + -ing): processing-pdfs, analyzing-spreadsheets, managing-databases
• Must contain only lowercase letters, numbers, and hyphens
• Maximum 64 characters
• Cannot contain XML tags or reserved words ("anthropic", "claude")
• Avoid generic names like helper, utils, tools

Note: Project skills (.claude/skills/) are automatically shared with your team via git. For personal skills only you use, create in ~/.claude/skills/ instead.

Step 3: Design the SKILL.md Structure

Every skill must have:

---
name: Your Skill Name
description: Brief description of what this Skill does and when to use it
---

# Your Skill Name

## Instructions
[Clear, step-by-step guidance for Claude]

## Examples
[Concrete examples of using this Skill]

Frontmatter Requirements:

• name: Required

- Max 64 characters - Lowercase letters, numbers, and hyphens only - Cannot contain XML tags or reserved words ("anthropic", "claude") - Use gerund form: processing-pdfs, analyzing-data

• description: Required, max 1024 characters

- MUST be written in third person (critical for discovery) - Include BOTH what it does AND when to use it - Mention key trigger words/phrases - Be specific, not vague - Cannot contain XML tags

Optional Frontmatter (Claude Code only):

• allowed-tools: Restrict which tools Claude can use (e.g., Read, Grep, Glob)

Step 4: Write the Instructions Section

Structure the instructions as:

1. Prerequisites - Required dependencies, tools, environment setup
2. Workflow - Step-by-step process (numbered steps)
3. Supporting Details - Additional context, script usage, error handling

Best Practices:

• Use clear, actionable language
• Number sequential steps
• Use bullet points for options/lists
• Include code blocks with bash commands
• Reference supporting files with relative links: [reference.md](reference.md)
• Keep focused on one capability
• Keep SKILL.md body under 500 lines for optimal performance
• Avoid deeply nested references - keep references one level deep from SKILL.md
• For files over 100 lines, include a table of contents at the top

Example workflow format:

### Workflow

1. **First step description**:

command to run

- Additional context
- Options or variations

1. **Second step description**:
  - Detailed instructions
  - What to look for
  - Expected outcomes
2. **Third step**...

Step 5: Write the Examples Section

Provide 2-4 concrete examples showing:

• Different use cases
• Various input formats
• Step-by-step execution
• Expected outcomes

Example format:

### Example 1: Descriptive Title

User request:

User's exact request text

You would:
1. First action
2. Second action with command:

actual command

1. Next steps...
2. Final result

Step 6: Add Supporting Files (Optional)

If the skill needs additional context:

1. Create files alongside SKILL.md
2. Reference them from instructions: [forms.md](forms.md)
3. Use progressive disclosure - split by topic/scenario

Common supporting file types:

• Additional instructions (e.g., advanced_usage.md)
• Reference documentation (e.g., api_reference.md)
• Scripts in scripts/ directory
• Templates in templates/ directory
• Configuration examples

Script guidelines:

• Make executable: chmod +x scripts/*.py
• Add PEP 723 inline dependencies for Python scripts
• Include usage instructions in SKILL.md
• Return clear output for Claude to parse

Step 7: Test the Skill

1. Verify file structure:

   ls -la .claude/skills/<skill-name>/

1. Check YAML frontmatter is valid: head -10.claude/skills/<skill-name>/SKILL.md
2. Test with relevant queries:

- Ask questions matching the skill's description - Verify Claude loads and uses the skill - Check that instructions are clear and actionable

1. Iterate based on testing:

- Refine description if skill doesn't trigger - Clarify instructions if Claude struggles - Add examples for common edge cases

Step 8: Commit to Version Control

Since project skills are automatically shared with your team, commit them to git:

git add .claude/skills/<skill-name>
git commit -m "Add <skill-name> skill"
git push

Note: Team members will get the skill automatically when they pull the latest changes.

Best Practices Summary

Description writing:

• ✅ "Transcribes audio/video files to text using Fireworks API. Use when user asks to transcribe, convert speech to text, or needs transcripts."
• ❌ "Helps with audio"
• Always write in third person - descriptions are injected into system prompt
• ✅ "Processes Excel files and generates reports"
• ❌ "I can help you process Excel files"

Instruction organization:

• Keep SKILL.md body under 500 lines
• Split complex content into linked files
• Use progressive disclosure for optional/advanced content
• Keep references one level deep from SKILL.md

Skill scope:

• One skill = one capability or workflow
• Don't combine unrelated tasks
• Make focused, composable skills

File references:

• Use relative paths: [file.md](file.md) not absolute paths
• Always use forward slashes (Unix style): scripts/helper.py ✓
• NOT Windows style: scripts\helper.py ✗
• Reference scripts with full path from skill root
• Make it clear when Claude should read vs execute files

Degrees of Freedom:

• High freedom (text instructions): When multiple approaches are valid
• Medium freedom (pseudocode/templates): When a preferred pattern exists
• Low freedom (specific scripts): When operations are fragile or consistency is critical

MCP Tool References:

• Use fully qualified names: ServerName:tool_name
• Example: BigQuery:bigquery_schema, GitHub:create_issue

Common Patterns from Existing Skills

Pattern 1: Transcription skill

• Prerequisites section with environment setup
• Clear numbered workflow
• Multiple examples showing different formats
• Supporting file for corrections/mappings

Pattern 2: Morning debrief skill

• Two-step process (transcribe, extend)
• Reference to detailed prompt in separate file
• File organization step
• Clear output structure specification

Pattern 3: Meta-skill (this one)

• Extensive prereading documentation
• Step-by-step creation workflow
• Multiple examples with variations
• Best practices and common patterns

Examples

Example 1: Creating a Simple Code Review Skill

User request:

Create a skill that reviews Python code for best practices

You would:

1. Read the documentation files in docs/
2. Ask clarifying questions:

- What specific best practices? (PEP 8, security, performance?) - Should it check only or suggest fixes? - Any specific frameworks or libraries?

1. Create the skill directory: mkdir -p.claude/skills/python-code-review
2. Write SKILL.md with: --- name: reviewing-python-code description: Reviews Python code for PEP 8 compliance, security issues, and performance. Use when reviewing Python code, checking code quality, or analyzing Python files. allowed-tools: Read, Grep, Glob ---
3. Add Instructions section with:

- Prerequisites (none needed, uses built-in tools) - Workflow: 1. Read the Python file(s) 2. Check PEP 8 compliance 3. Identify security issues 4. Suggest performance improvements 5. Provide summary with specific line references

1. Add 3 examples:

- Example 1: Single file review - Example 2: Multi-file project review - Example 3: Focused security review

1. Test with sample Python files

Example 2: Creating a Data Analysis Skill with Scripts

User request:

Build a skill for analyzing CSV data with statistics and visualizations

You would:

1. Read documentation files
2. Define scope with user:

- What statistics? (mean, median, correlations?) - What visualizations? (charts, plots?) - Output format? (markdown report, images?)

1. Create structure: mkdir -p.claude/skills/csv-analysis/scripts mkdir -p.claude/skills/csv-analysis/templates
2. Write SKILL.md referencing:

- scripts/analyze.py - Statistical analysis script - scripts/visualize.py - Chart generation script - templates/report_template.md - Output template

1. Create Python scripts with inline dependencies: # /// script # requires-python = ">=3.10" # dependencies = ["pandas", "matplotlib", "seaborn"] # ///
2. Write clear instructions for:

- When to run which script - How to interpret output - How to customize analysis

1. Add examples showing:

- Basic statistics - Visualization generation - Custom report creation

1. Test with sample CSV files

Example 3: Creating a Multi-File Documentation Skill

User request:

Create a skill for writing technical documentation with our company's style guide

You would:

1. Read documentation files
2. Gather requirements:

- Get company style guide document - What types of docs? (API, user guides, architecture?) - Any templates or examples?

1. Create comprehensive structure: mkdir -p.claude/skills/tech-docs/{templates,examples,guidelines}
2. Organize content:

- SKILL.md - Overview and workflow - guidelines/style_guide.md - Company style rules - guidelines/api_docs.md - API documentation specifics - guidelines/user_guides.md - User guide standards - templates/api_template.md - API doc template - templates/guide_template.md - User guide template - examples/ - Sample documentation

1. Write SKILL.md that:

- References guidelines by doc type - Uses progressive disclosure (only load needed guidelines) - Provides workflow for each doc type

1. Add examples for:

- API endpoint documentation - User guide creation - Architecture decision records

1. Test with various documentation requests

Example 4: Extending an Existing Skill

User request:

Add spell correction to our transcribe skill

You would:

1. Read current skill: cat.claude/skills/transcribe/SKILL.md
2. Identify where to add the feature:

- After transcription step - Before final output

1. Create supporting file: touch.claude/skills/transcribe/spell_corrections.md
2. Write correction mappings in new file: # Spell Corrections - "cloud code" → "claude code" - "API" → "API" (ensure caps)...
3. Update SKILL.md workflow:

- Add step: "Apply spell corrections from spell_corrections.md" - Reference the corrections file

1. Update examples to show correction step
2. Test with audio that has common errors

Summary

Creating skills is about packaging expertise into discoverable, composable capabilities. Follow these principles:

1. Read the docs first - Understand progressive disclosure and skill architecture
2. Write clear descriptions - Include what AND when
3. Keep instructions focused - Use supporting files for additional context
4. Test thoroughly - Verify Claude discovers and uses the skill correctly
5. Iterate with feedback - Refine based on actual usage

Skills transform general-purpose Claude into a specialist for your domain. Start small, test early, and expand as needed.