Skill Creator

Overview

Anthropic's official guide for creating high-quality Claude skills. A systematic process from requirement definition through packaging and iteration.

What Skills Provide

• Specialized workflows: Step-by-step domain guidance
• Tool integrations: Connect external services via MCP or scripts
• Domain expertise: Encode expert knowledge as reusable instructions
• Bundled resources: Scripts, references, templates, and assets

Core Design Principles

Conciseness

The context window is a shared resource. Skills should contain only the minimum information needed.

Appropriate Degrees of Freedom

Overly rigid skills can't adapt to real scenarios. Leave reasonable room for Claude's judgment.

Progressive Disclosure

Three-level loading system for maximum token efficiency:

Skill File Structure

your-skill-name/
├── SKILL.md                  # Required — main skill file
├── scripts/                  # Optional — executable code
│   ├── process_data.py
│   └── validate.sh
├── references/               # Optional — documentation
│   ├── api-guide.md
│   └── examples/
└── assets/                   # Optional — templates, icons
    └── report-template.md

Six-Step Creation Process

Step 1: Understand the Skill

Define use cases with concrete examples. Write 2-3 real requests users would make.

Step 2: Plan Reusable Content

Determine what's universal guidance (SKILL.md) vs. specialized reference (references/).

Step 3: Initialize the Skill

Create folder structure and write YAML frontmatter:

---
name: your-skill-name          # kebab-case, required
description: >                 # Required: what + when
  Describe skill functionality and trigger conditions.
  Include specific phrases users might say.
---

Step 4: Write SKILL.md

• Put critical instructions at the top
• Use numbered steps and clear structure
• Mark key info with ## Important / ## Critical headers
• Reference bundled resources with clear paths

Step 5: Package the Skill

• Folder names in kebab-case
• SKILL.md spelling is case-sensitive
• No README.md inside the skill folder
• Keep SKILL.md under 5,000 words

Step 6: Iterate Based on Usage

YAML Frontmatter Fields

Invocation Control Matrix

Skill Locations & Priority

Priority: Enterprise > Personal > Project. Plugin skills use plugin-name:skill-name namespace.

Description budget: 2% of context window, fallback 16,000 characters. Override with SLASH_COMMAND_TOOL_CHAR_BUDGET env var.

String Substitutions

Dynamic variables available in skill body:

Dynamic Context Injection

Use !`command` syntax to execute shell commands before skill content is sent to Claude, injecting real-time data into context:

---
name: pr-summary
context: fork
agent: Explore
allowed-tools: Bash(gh *)
---
## Pull Request Context
- PR diff: !`gh pr diff`
- PR comments: !`gh pr view --comments`

This is a powerful pattern — it lets skills access live data (Git status, API responses, environment info) rather than just static text.

Subagent Execution Pattern

Set context: fork to run the skill in an isolated subagent, where the skill body becomes the subagent's prompt:

---
name: code-review
context: fork
agent: general-purpose
model: sonnet
allowed-tools: Read, Grep, Glob
---
You are a senior code reviewer. Review the recent changes...

Use cases:

• Isolated context to avoid polluting the main conversation
• Restricted tool access for safety
• Different model selection for specific tasks

Best Practices

• Code is more reliable than language: Use scripts for critical validation, not just prose
• Iterate on one task first: Get a single hard task working in conversation, then extract it into a skill
• The description field is the most important part: It determines when Claude loads your skill
• Skills are living documents: Plan to iterate continuously, not write once
• Leverage dynamic context: !`command` lets skills inject real-time data
• Use subagent isolation: context: fork prevents complex skills from polluting main conversation context
• Mind the description budget: All skill descriptions share 2% of context window — keep descriptions concise