PluginInstalls cleanly
Make and test your own skills
Skill Creator
Anthropic's plugin helps you write a new skill step by step, test it on example tasks, compare results and improve its description so Claude uses it.
- Works with
- Claude Code
- Good for
- Everyone
What it does
A skill is a folder with a SKILL.md file that tells Claude how to do one job your way, for example how to format a weekly report. Skill Creator is Anthropic's plugin for making these folders without knowing the format.
You say what you want the skill to do. Claude asks a few questions, writes a first draft, and suggests some test prompts. It then runs those tests, shows you the answers side by side on a review page, and lets you mark what is good or bad. It rewrites the skill from your feedback, and you repeat until you are happy.
It can also improve an existing skill. It can tune the skill's short description, which is what Claude reads to decide when to use the skill.
Read this first. Testing uses a lot of your plan's limits or API credits. There are two kinds of test run:
- Normal tests. Each test case runs twice, once with your skill and once without it, plus a grader. These run as sub-agents (helper sessions) inside your own session.
- Description tuning. Only this step starts separate
claude -psessions in the background. By default it uses 20 test queries, 3 runs each, so about 60 sessions per round. It can run up to 5 rounds, so up to about 300 sessions.
The claude -p sessions use your normal Claude Code setup: your settings, plugins and hooks. They also use your environment, so if ANTHROPIC_API_KEY is set in your shell, they use that key and its billing. While they run, the script adds a temporary command file to your project's .claude/commands folder and deletes it after (our reading of scripts/run_eval.py).
The review page is a small local web server on port 3117 that opens in your browser. Before it starts, it closes any other program that already uses port 3117. If you run something else on that port, stop it first or ask for the static review page, which writes an HTML file instead.
The plugin runs bundled Python scripts to grade results and build the review page. It writes your new skill files and a workspace folder next to the skill, named <skill-name>-workspace, with the test outputs and grades.
Who it is for
Claude Code users who repeat the same instructions often and want to turn them into a skill that works the same way each time.
Example prompts
- "I want to make a skill that turns my meeting notes into a task list."
- "Improve my weekly-report skill; it keeps missing the summary section."
- "Test my invoice skill on three example invoices and show me the results."
- "Make the description of my skill trigger more reliably."
How to install
claude plugin install skill-creator@claude-plugins-officialHow to use it
Ask in Claude Code: I want to make a skill that turns my meeting notes into a task list. Skill Creator asks what you need, writes a draft, then offers to test it on a few example prompts.
Check it works
- In Claude Code, run
/pluginand look for skill-creator in the installed list. - Ask it to make a tiny test skill. It should create a folder with a SKILL.md file.
Common problems
- The new skill does not appear — no error text — Claude Code watches
~/.claude/skillsand your project's.claude/skillsand picks up changes during the session. If the skills folder did not exist when the session started, run/reload-skills(Claude Code skills docs). - Test runs use up your limits fast — no error text — each test case runs twice (with and without the skill), and description tuning runs about 60
claude -psessions per round. Use fewer test prompts, or skip description tuning (SKILL.md, scripts/run_loop.py). - Another local program stops when the review page opens — no error text — the review page closes any program on port 3117 first. Ask for the static review page instead (eval-viewer/generate_review.py).
What to watch out for
- Each test case runs twice in sub-agents (with and without your skill), and description tuning starts about 60 claude -p sessions per round, up to 5 rounds; all of it uses your plan's limits or API credits (SKILL.md, scripts/run_eval.py, run_loop.py)
- The claude -p test sessions use your normal Claude Code setup, including an ANTHROPIC_API_KEY if one is set, and briefly add a command file to your project's .claude/commands folder (scripts/run_eval.py)
- The review page runs a small local web server on port 3117 and first closes any other program that uses that port (eval-viewer/generate_review.py)
- Runs bundled Python scripts to grade results and build the review page (SKILL.md)
- It writes new skill files and a workspace folder of test results on your computer; read a skill before you share or install it elsewhere
Licence and author
- Licence
- Apache-2.0
- Author
- Anthropic
How we tested it
We installed it on an empty test computer on 18 Oct 2026. Installed and started in a clean test system. Not yet used in a real AI session.
Install only. Test runs and description tuning not run (they need a signed-in Claude Code). Not yet used on real work.
Technical details of the test
- Date
- 18 Oct 2026
- Method
- Intake sandbox run on 2026-10-18 in a clean container with no account: cloned the repository and checked its skill, agent and plugin files; installed skill-creator@claude-plugins-official with the Claude Code command line.
Checked again on 18 Oct 2026
Alternatives
- Superpowers — a large ready-made skill set, including a skill for writing skills.
- Claude Code Setup — recommends existing skills and automations instead of making new ones.
- Writing SKILL.md by hand — Claude Code's skills docs show the format.
FAQ
Do I need an API key?
No. The test scripts use your Claude Code sign-in. If ANTHROPIC_API_KEY is set in your shell, they use that key and its billing (improve_description.py, run_eval.py).
Does it work in Codex or Gemini CLI? This plugin is for Claude Code. The SKILL.md files it makes follow the Agent Skills format.
Will my skill be shared? No. The files stay on your computer unless you share them.
Does it cost money? The README names no price. Test runs count toward your plan's limits or your API use.
Is it in Anthropic's directory?
Yes, as skill-creator@claude-plugins-official.
Sources
- github.com/anthropics/claude-plugins-official/tree/main/plugins/skill-creator
- code.claude.com/docs/en/skills
- raw.githubusercontent.com/anthropics/claude-plugins-official/main/.claude-plugin/marketplace.json