Language: 日本語English

skill-creator Skill

Author
Anthropic
Topics
Claude Code Customization & Workflow · Code Quality, Review & Testing
First seen
2026-07-09
Last confirmed
2026-07-13
Explanation last updated
2026-07-12

The explanation below is AI-generated. Please verify it against the sources.

skill-creator is a Claude skill that helps users create new skills, improve existing ones, and measure their performance through structured testing. According to the SKILL.md, the workflow involves interviewing the user to draft a SKILL.md, writing test prompts, running the skill and a baseline (or old version) on those prompts, and reviewing results both qualitatively and via quantitative benchmarks. It also includes a separate description-optimization loop that uses train/test eval queries to improve a skill's triggering accuracy, and a packaging step that produces an installable .skill file. The SKILL.md gives adapted instructions for different environments (Claude Code, Claude.ai, Cowork) since availability of subagents and a browser differs across them.

Overview

skill-creator is not a commercial SaaS but a Claude Agent Skill (a SKILL.md file with bundled scripts and reference docs) used within Claude Code, Claude.ai, or Cowork to assist in building, testing, and refining other skills.

What you can do with skill-creator

  • Interview the user to capture what a skill should do, when it should trigger, expected output format, and whether test cases are needed, then draft the SKILL.md
  • Write test prompts and run them with the skill and against a baseline (no skill, or an old version) to compare outputs
  • Draft quantitative assertions, grade runs against them, and aggregate results into benchmark.json/benchmark.md with pass rates, timing, and token usage
  • Launch eval-viewer/generate_review.py so the user can review outputs and benchmark data in a browser or static HTML file
  • Iterate on the skill across multiple rounds based on user feedback captured in feedback.json
  • Run scripts.run_loop, a description-optimization loop that tests skill-triggering accuracy on a train/test split of eval queries and proposes an improved description
  • Package the finished skill into an installable .skill file via scripts.package_skill

Sources

Original description (English)

Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.

History of skill-creator

Back to list