Claude Official Skill: skill-creator

What it is

Skill-creator is an official Claude Code skill from Anthropic for creating new skills, modifying existing ones, and measuring skill performance. It provides an iterative workflow: draft a skill, run test prompts, evaluate results both qualitatively and quantitatively, then refine based on feedback.

The skill targets developers and teams who want to extend their AI agent's capabilities with custom domain expertise. It works with Claude Code, GitHub Copilot, and Gemini CLI.

How it saves time or tokens

Without a structured skill creation process, writing effective AI agent skills requires trial and error. Skill-creator provides a repeatable workflow that front-loads evaluation: it generates test prompts, runs them against the draft skill, and helps you assess results with quantitative metrics. This reduces the number of iterations needed to produce a working skill. The evaluation framework catches edge cases early, preventing skills from failing in production.

How to use

1. Install the skill via Claude Code:

claude skill install anthropics/skills/skill-creator

1. Or manually place the SKILL.md file:

mkdir -p .claude/skills/skill-creator
# Copy SKILL.md content to .claude/skills/skill-creator/SKILL.md

1. Start creating a skill by describing what you want it to do. The skill-creator guides you through drafting, testing, and refining.

Example

The skill creation workflow in practice:

# Step 1: Define the skill
User: 'Create a skill for reviewing SQL migrations'

# Step 2: Claude drafts the skill SKILL.md
# Includes: trigger conditions, review checklist, output format

# Step 3: Generate test prompts
- 'Review this migration that adds a new index'
- 'Review this migration that drops a column'
- 'Review this migration with no WHERE clause on UPDATE'

# Step 4: Run tests and evaluate
# Claude runs each test prompt against the draft skill
# Generates quantitative metrics:
#   - Did it catch the dangerous UPDATE? (pass/fail)
#   - Did it flag the index naming convention? (pass/fail)
#   - Was the output format consistent? (pass/fail)

# Step 5: Refine based on results
# Rewrite sections that underperformed
# Add edge cases to the skill's instructions

Related on TokRepo

• Prompt library — Browse existing skills and prompts for AI agents.
• Featured workflows — Discover curated tools and frameworks on TokRepo.

Common pitfalls

• Writing skills that are too broad (e.g., 'review all code') leads to unfocused results. Scope each skill to a specific domain or task type.
• Skipping the evaluation step means you deploy skills without knowing their failure modes. Always run at least 3 test prompts before using a skill in production.
• Not versioning skills alongside your codebase makes it hard to track changes. Store SKILL.md files in your repository and treat them like code.