Creating Skills

Skills are instruction folders that teach Claude specialized workflows. The core principle is progressive disclosure: minimize tokens while maintaining expertise.

Skill Structure

skill-name/
├── SKILL.md           # Required. Frontmatter + instructions
├── references/        # Optional. Deep documentation (loaded on-demand)
├── scripts/           # Optional. Executable code
└── assets/            # Optional. Templates, icons

SKILL.md Template

---
name: kebab-case-name
description: Use when [trigger conditions]. [What it does]. [When to apply it].
---

# Skill Title

## Overview
One paragraph: what this skill does and the core principle.

## Instructions
Step-by-step guidance or reference content.

## References
- Link to [references/detail.md](references/detail.md) for deep content

Writing the Description

The description field is the most critical part. It's always loaded in Claude's system prompt and determines auto-invocation.

Rules:

• Start with "Use when..." followed by trigger conditions
• Include natural keywords users would say
• Keep under 1024 characters
• Be specific about what AND when

Good:

description: Use when creating new skills, editing existing skills, or reviewing skill quality. Use when user mentions building, writing, or improving a SKILL.md file.

Bad:

description: Helps with skills

Content Guidelines

Be Concise

The context window is shared. Claude is already smart — only add domain-specific knowledge it doesn't have. Keep SKILL.md body under 500 lines.

Match Degrees of Freedom to Task

Freedom	When	Format
High	Multiple valid approaches, creative work	Text guidelines
Medium	Preferred pattern exists, some variation OK	Pseudocode with params
Low	One correct approach, variations cause failure	Exact code/commands

Use Consistent Terminology

Pick one term per concept. Don't alternate between "skill", "command", and "instruction" — use "skill" throughout.

Progressive Disclosure (3 Levels)

1. Frontmatter — Always loaded. Just enough for Claude to know when to use it
2. SKILL.md body — Loaded when skill is invoked. Full instructions
3. Referenced files — Loaded on-demand. Deep documentation, examples, APIs

Frontmatter Reference

Field	Default	Purpose
name	Directory name	Display name, becomes /slash-command. Kebab-case, max 64 chars
description	First paragraph	When to use. Claude uses this for auto-invocation
argument-hint	—	Autocomplete hint: [issue-number], [filename]
disable-model-invocation	false	true = manual-only (deploy, commit, dangerous ops)
user-invocable	true	false = hidden from / menu (background knowledge)
allowed-tools	—	Tool restrictions: Read, Grep, Bash(gh *)
model	—	Override model for this skill
context	—	fork = run in isolated subagent
agent	general-purpose	Subagent type when context: fork

Skill Types

Reference skills — Add knowledge Claude applies to current work (conventions, patterns, domain knowledge). Run inline with conversation context.

Task skills — Step-by-step instructions for specific actions. Often use disable-model-invocation: true and context: fork.

String Substitutions

Variable	Description
$ARGUMENTS	All arguments passed to the skill
$ARGUMENTS[N] or $N	Specific argument by index (0-based)
${CLAUDE_SESSION_ID}	Current session ID

Distribution

Level	Path	Scope
Personal	~/.claude/skills/<name>/SKILL.md	All your projects
Project	.claude/skills/<name>/SKILL.md	This project only
Plugin	<plugin>/skills/<name>/SKILL.md	Where plugin enabled

Priority: Enterprise > Personal > Project > Plugin.

Testing Checklist

Run these three test types before considering a skill complete:

1. Triggering Test

• Ask Claude something that should activate the skill (without /invoking it)
• Verify the description keywords match natural language triggers
• Test phrases that should NOT trigger it

2. Functional Test

• Invoke with /skill-name directly
• Verify Claude follows instructions correctly
• Test with and without arguments
• Check edge cases

3. Performance Test

• Is SKILL.md under 500 lines?
• Does Claude complete tasks in reasonable steps?
• Are referenced files loaded only when needed?

Common Mistakes

Mistake	Fix
Description too vague	Add "Use when..." + specific trigger phrases
SKILL.md too long (>500 lines)	Move detail to references/
Over-explaining basics Claude knows	Remove — Claude knows markdown, YAML, git, etc.
Inconsistent terminology	Pick one term per concept
No testing before deployment	Run all 3 test types
Missing disable-model-invocation on dangerous ops	Add it for deploy, commit, send-message skills

References

• Anthropic Official Guide — Condensed content from the 32-page PDF
• Patterns & Troubleshooting — 5 reusable patterns + debugging