Creating Skills

Guide for creating Claude Code skills following Anthropic's official best practices.

Quick Start

# 1. Create skill directory
mkdir -p ~/.claude/skills/<skill-name>/references

# 2. Create SKILL.md with frontmatter
cat > ~/.claude/skills/<skill-name>/SKILL.md << 'EOF'
---
name: <skill-name>
description: <what it does>. Use when <trigger phrases>. <key capabilities>.
---

# <Skill Title>

<One-line description>

## Quick Start
<Minimal steps to use the skill>

## Core Workflow
<Step-by-step instructions>

## Important Rules
<Critical constraints and requirements>
EOF

# 3. Add helper scripts if value-add
# 4. Add reference files for detailed docs

SKILL.md Structure

Frontmatter (REQUIRED)

---
name: skill-name          # lowercase, hyphens, no spaces
description: <desc>       # CRITICAL for discovery (max 1024 chars)
---

Advanced Frontmatter (v3.0)

Inspired by Claude Code's internal skills architecture (loadSkillsDir.ts).

---
name: skill-name
description: <What it does>. <Key capabilities>.

# v3.0 fields — separate trigger logic from description
when_to_use: >
  Use when user wants to create PR, open pull request,
  or merge feature. Also triggers on 'review', 'submit'.
  
# Security: restrict which tools this skill can use
allowed-tools:
  - run_command
  - view_file
  - grep_search
  # Omit dangerous tools like write_to_file for read-only skills

# Auto-activate when editing files matching these paths
paths:
  - src/auth/**
  - lib/security/**

# Parameterized skills: $1, $2 substitution
arguments: [target_file, test_scope]
argument-hint: "<file> [scope]"

# Control execution depth
effort: thorough  # Options: minimal, standard, thorough

# Specify a model for this skill (rare)
# model: claude-sonnet-4-20250514

# Whether users can invoke directly (vs agent-only)
user-invocable: true

# Pure text skill (no LLM invocation)
# disable-model-invocation: false

# Isolated execution context
# context: fork
---

When to use

when_to_use

description

• description

  = WHAT the skill does (appears in listings)

• when_to_use

  = WHEN to trigger it (matching logic, trigger phrases)
• Keeping them separate prevents the description from becoming a keyword dump

Description Formula

<What it does>. Use when <trigger phrases>. <Key capabilities>.

Example:

Creates GitHub Pull Requests with automated validation. Use when user wants
to create PR, open pull request, or merge feature. Validates tasks, runs
tests, generates Conventional Commits format.

Trigger phrases to include:

• Action verbs: "create", "handle", "manage", "process"
• User intent: "wants to", "needs to", "asks for"
• Keywords users would say: "PR", "pull request", "review comments"

Body Sections (ORDER MATTERS)

1. Title -

   # Skill Name

2. One-liner - Single sentence summary
3. Quick Start - Minimal steps (copy-paste ready)
4. Core Workflow - Numbered steps with code blocks
5. Helper Scripts (if any) - Table with purpose
6. Important Rules - Critical constraints (bold ALWAYS/NEVER)

Naming Conventions

Format Options

processing-pdfs

github-pr-creation

github-pr-*

Rules

• Lowercase only
• Hyphens between words (no underscores)
• No spaces
• Descriptive but concise (2-4 words)

Token Budget

If approaching 500 lines:

1. Move detailed examples to

   references/examples.md

2. Move troubleshooting to

   references/troubleshooting.md

3. Keep SKILL.md focused on workflow

Helper Scripts Guidelines

When to Create Scripts

DO create scripts for:

• Complex logic (severity classification, commit analysis)
• Multi-step processing with JSON output
• Reusable utilities across invocations

DON'T create scripts for:

• Single command wrappers (

  gh api ...

  )
• Simple operations Claude can do inline
• One-line bash commands

Script Requirements

#!/usr/bin/env python3
"""Script description."""

import json
import sys

def main():
    # Read from stdin or args
    # Process data
    # Output JSON to stdout
    print(json.dumps(result))

if __name__ == "__main__":
    main()

• Output JSON for structured data
• Use stdin/stdout for piping
• Include clear error messages
• Keep focused on single responsibility

Directory Structure

~/.claude/skills/<skill-name>/
├── SKILL.md              # Main skill file (< 500 lines)
├── scripts/              # Optional helper scripts
│   └── helper.py         # Only if value-add
└── references/           # Optional detailed docs
    ├── examples.md       # Extended examples
    └── guide.md          # Deep-dive documentation

Core Principles

1. Claude is Already Smart

"Default assumption: Claude is already very smart. Only add context Claude doesn't already have."

Challenge each line:

• Does Claude really need this explanation?
• Can I assume Claude knows this?
• Does this paragraph justify its token cost?

2. Progressive Disclosure

SKILL.md (primary)
    ↓ references/ (when needed)
        ↓ external links (rarely)

• Start minimal, expand as needed
• Don't front-load all information
• Let Claude discover details when relevant

3. User Confirmation for Critical Actions

**ALWAYS** confirm before:
- Modifying files
- Running destructive commands
- Creating external resources (PRs, issues)

4. Structured Output

Prefer JSON for script output:

# Good: Structured, parseable
python3 script.py | jq '.result'

# Bad: Unstructured text
python3 script.py | grep "Result:"

Quality Checklist

Before finalizing a skill:

• Frontmatter: name + description present
• Description: includes WHAT + WHEN triggers + capabilities
• Naming: lowercase, hyphens, descriptive
• Quick Start: copy-paste ready, < 30 lines
• Line count: SKILL.md < 500 lines
• Scripts: only value-add, no wrappers
• Rules: critical constraints marked with bold
• Test: skill triggers on expected phrases

Anti-Patterns to Avoid

Examples

Good Description

Handles PR review comments with severity classification. Use when user
wants to resolve PR comments, handle review feedback, or fix review
comments. Fetches via GitHub CLI, classifies by severity, proposes fixes.

Good Quick Start

# 1. Get PR number
PR=$(gh pr view --json number -q '.number')

# 2. Fetch and classify comments
gh api repos/{owner}/{repo}/pulls/$PR/comments | \
  python3 ~/.claude/skills/github-pr-review/scripts/classify_comments.py

# 3. Process each comment: show → propose → confirm → apply
# 4. Commit: git commit -m "fix(scope): address review - [desc]"
# 5. Reply and push

Good Important Rules

## Important Rules

- **ALWAYS** confirm before modifying files
- **ALWAYS** verify ALL issues in multi-issue comments
- **NEVER** skip user confirmation for file changes
- Group related changes → single commit

References

• references/official_best_practices.md

  - Full Anthropic documentation

• references/skill_examples.md

  - Complete skill examples