Create Skill

Overview

This skill guides you through creating high-quality Claude Code skills following Anthropic's best practices and prompt engineering principles.

Quick Start

When creating a new skill:

1. Ask for skill requirements - Understand what the skill should do
2. Design the skill structure - Plan files and progressive disclosure
3. Write the SKILL.md - Create the main instruction file
4. Add reference files - Create supporting documentation if needed
5. Test and iterate - Verify the skill works as expected

Skill Creation Process

Step 1: Gather Requirements

Ask the user:

• What should the skill do?
• When should it be triggered (auto-discovery keywords)?
• What tools does it need access to?
• Should it be user-invocable (slash command)?

Step 2: Create Directory Structure

~/.claude/skills/{skill-name}/
├── SKILL.md              # Required: Main instructions
├── reference.md          # Optional: Detailed documentation
├── examples.md           # Optional: Usage examples
└── scripts/              # Optional: Utility scripts
    └── helper.py

Location Guide:

Step 3: Write SKILL.md

Use this template:

---
name: skill-name
description: [Specific description with keywords for auto-discovery. Describe WHAT it does and WHEN to use it.]
user-invocable: true
allowed-tools: [Optional: Tool restrictions]
---

# Skill Title

## Quick Start

[Essential instructions - what Claude needs to know immediately]

## Workflow

[Step-by-step process]

## Examples

[Concrete input/output examples]

## Advanced Features

For detailed reference, see [reference.md](reference.md).

Best Practices Checklist

Description (Critical for Auto-Discovery)

<good_description> 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. </good_description>

<bad_description> description: Helps with documents </bad_description>

Rules:

• Include specific keywords users might mention
• State WHAT it does AND WHEN to use it
• Write in third person ("Processes files" not "I process files")
• Be specific, not vague

Content Guidelines

<progressive_disclosure>

• Put essential info in SKILL.md
• Link to detailed docs in separate files
• Claude loads additional files only when needed </progressive_disclosure>

<consistent_terminology>

• Choose one term and stick with it
• "API endpoint" everywhere, not "URL/route/path"
• "Extract" everywhere, not "pull/get/retrieve" </consistent_terminology>

Prompt Engineering Principles

<explicit_instructions> Be direct and explicit. Tell Claude exactly what to do.

LESS EFFECTIVE: "Handle errors appropriately" MORE EFFECTIVE: "Catch exceptions, log the error message, and return a structured error response with status code and message." </explicit_instructions>

<provide_context> Explain WHY, not just WHAT.

LESS EFFECTIVE: "Always validate input" MORE EFFECTIVE: "Validate input because this data comes from external users and could contain malicious content or unexpected formats." </provide_context>

<use_examples> Show concrete input/output pairs:

Example: Input: User request to format date Output:

const formatted = new Intl.DateTimeFormat('en-US').format(date);

</use_examples>

<xml_tags> Use XML tags for structured sections - Claude is fine-tuned to respond to them:

<task>Description</task>
<constraints>Limits</constraints>
<output_format>Expected format</output_format>

</xml_tags>

Degrees of Freedom

Tool Restrictions

allowed-tools: Read, Grep, Glob, Bash(python:*)

Common patterns:

• Read-only: Read, Grep, Glob
• With Python: Read, Bash(python:*)
• Full access: omit allowed-tools

Frontmatter Reference

---
name: skill-name                    # Required: Unique identifier
description: Description here       # Required: For auto-discovery
user-invocable: true               # Optional: Show in slash menu (default: true)
disable-model-invocation: false    # Optional: Block Skill tool (default: false)
allowed-tools: Read, Grep          # Optional: Restrict available tools
context: fork                      # Optional: Run in isolated context
agent: Explore                     # Optional: Use specific agent type
hooks:                             # Optional: Lifecycle hooks
  PreToolUse:
    - matcher: "Bash"
      hooks:
        - type: command
          command: "./check.sh"
---

Testing the Skill

After creating the skill:

1. Test auto-discovery - Ask Claude something matching the description
2. Test slash command - Use /skill-name directly
3. Test with different models - Haiku needs more guidance, Opus needs less
4. Iterate based on results - Refine instructions if needed

Common Patterns

Validation Workflow

1. Run validation: `script validate input`
2. Review results
3. Fix any issues
4. Re-validate until passing

Multi-Step Process

## Workflow
1. [First step]
2. [Second step]
3. [Verification step]
4. [Final step]

Script Integration

## Scripts

- `scripts/analyze.py` - Analyzes input
- `scripts/process.py` - Processes data
- `scripts/validate.py` - Validates output

Run scripts, don't load their contents into context.

Additional Resources

For detailed examples and advanced patterns, see:

• examples.md - Complete skill examples
• reference.md - Full frontmatter reference