Skill creator: how Anthropic's skill-creator works, plus other ways to build a skill

What is the skill creator?

The skill creator is Anthropic's official skill for making other skills. You install skill-creator, tell Claude what you want a skill to do, and it interviews you, drafts the SKILL.md, runs test prompts, and helps you improve the skill until it works. Its own description sums it up: "Create new skills, modify and improve existing skills, and measure skill performance."

It ships in two places. The source lives in anthropics/skills, Anthropic's public skills repo, and it's also packaged as a plugin in the official Claude Code marketplace. Claude Code's skills documentation points to that plugin as a way to run evals on a skill.

What sets it apart from writing a skill yourself is testing. skill-creator doesn't stop at a draft: it compares Claude's output with the skill and without it, grades the results, and can tune the skill's description so it triggers on the right requests.

Other tools now use the same name for something lighter. Codex has a built-in $skill-creator and Cursor has /create-skill. Both are covered further down, along with writing a skill by hand.

How do you install skill-creator in Claude Code?

The quickest way is the plugin from Anthropic's official marketplace. Start Claude Code by running claude, then enter this at its prompt:

/plugin install skill-creator@claude-plugins-official

Claude Code's docs list two fixes if that fails. If you see Marketplace "claude-plugins-official" not found, add the marketplace and retry:

/plugin marketplace add anthropics/claude-plugins-official

If you see that the plugin isn't found in the marketplace, refresh your local copy of the catalog with /plugin marketplace update claude-plugins-official. In the VS Code extension or the desktop app, install it from the plugin screen instead.

You can also get it from the anthropics/skills repo, which works as its own marketplace. Its README gives these commands, and the repo's marketplace file lists skill-creator inside the example-skills plugin, next to skills like frontend-design and mcp-builder:

/plugin marketplace add anthropics/skills
/plugin install example-skills@anthropic-agent-skills

Once it's installed, ask for what you want in plain words. Claude Code's docs use the example "evaluate my summarize-changes skill with skill-creator". For a new skill, describe the job, such as "make a skill that writes our release notes."

How do you use skill-creator in the Claude app?

According to the anthropics/skills README, the example skills, skill-creator included, "are all already available to paid plans in Claude.ai." Skills need code execution. Anthropic's help article on using skills in Claude says to turn on "Code execution and file creation" under Settings, Capabilities, then manage skills under Customize, Skills.

The skill's own instructions include a section for Claude.ai. The core loop is the same (draft, test, review, improve), but some steps change because the app has no subagents:

• Test cases run one at a time in the conversation, not in parallel.
• There is no browser viewer. Claude shows each test prompt and its output in the chat and asks for feedback.
• The baseline comparison and benchmark are skipped.
• Description tuning is skipped, because it needs claude -p, which only exists in Claude Code.
• Packaging still works. The package_skill.py script produces a .skill file you can download.

To add the finished skill to your account, the help center describes zipping the skill folder and uploading it from Customize, Skills, using the + button and Upload a skill. One limit to watch: Anthropic's help article on creating custom skills lists a 200-character maximum for descriptions in the Claude app, shorter than the 1,024 characters the open standard allows.

What does skill-creator's workflow look like?

The SKILL.md describes a loop, and it tells Claude to work out where you are in it and jump in there. If you already have a draft, it can go straight to testing. If you say you don't want evaluations, it is told to drop them and just work with you.

1. Capture intent. It asks what the skill should enable, when it should trigger, what the output looks like, and whether test cases make sense. If you say "turn this into a skill" mid-conversation, it pulls the answers from what you already did.
2. Interview and research. It asks about edge cases, formats, example files and success criteria before writing tests.
3. Write the SKILL.md. It fills in the name and description, plus an optional compatibility field. The description carries all the "when to use" information, and the skill tells Claude to make it "a little bit 'pushy'" because Claude tends to undertrigger skills.
4. Write test cases. Realistic prompts go into evals/evals.json inside the skill folder.
5. Run them twice. Each prompt runs with the skill and without it (or against the old version when you're editing a skill), in separate subagents so every run starts clean. Token count and duration are recorded.
6. Grade and benchmark. A grader checks each assertion and writes pass or fail with evidence to grading.json. Results are aggregated into benchmark.json with pass rate, time and tokens for each configuration.
7. Review. It opens an HTML viewer with an Outputs tab and a Benchmark tab, where you read each result and leave comments. Your comments are saved to feedback.json for the next round.
8. Improve and repeat. It rewrites the skill from your feedback and runs the next iteration, then grows the test set once things look good.

There's also an advanced blind comparison mode, which Claude Code's docs describe as "a blind A/B between two versions of the skill," so you can confirm an edit is actually better.

The skill also carries writing advice you can use without it: keep SKILL.md under 500 lines, move detail into references/, scripts/ and assets/ folders that load only when needed, and explain why a rule matters instead of stacking up all-caps MUSTs.

How does skill-creator tune a skill's description?

The description decides whether a skill gets used, so skill-creator has a separate step for it. It writes about 20 test queries, roughly half that should trigger the skill and half that shouldn't. The should-not queries are meant to be near misses: requests that share words with the skill but need something else. You review and edit the list in an HTML page before anything runs. The format is a plain list:

[
  {"query": "the user prompt", "should_trigger": true},
  {"query": "another prompt", "should_trigger": false}
]

Then a script splits the set into 60% for training and 40% held back for testing, runs each query three times to get a reliable trigger rate, and asks Claude for a better description based on the misses. It repeats this up to five times and picks the best description by its score on the held-back queries, which guards against a description that only fits the training set. This step runs claude -p, so it works in Claude Code (and, per the skill, in Cowork) but not in the Claude app.

What are the other ways to create a skill?

skill-creator is thorough, which also makes it slow for a small skill. These are the main alternatives.

Write it by hand. A skill is a folder with a SKILL.md file: a name (lowercase letters, digits and hyphens, matching the folder) and a description, then instructions in Markdown. Save it in ~/.claude/skills/<name>/ for all your projects or .claude/skills/<name>/ for one repo. Our guide on how to write a SKILL.md walks through the rules and a full example. The minimal shape, from the anthropics/skills README:

---
name: my-skill-name
description: A clear description of what this skill does and when to use it
---

# My Skill Name

[Add your instructions here that Claude will follow when this skill is active]

Codex's built-in creator. OpenAI's Codex skills docs say to invoke $skill-creator. It asks what the skill does, when it should trigger, and whether it should stay instruction-only or include scripts. More in our Codex skills guide.

Cursor's built-in creator. Cursor lists /create-skill among its built-in skills; it "creates Agent Skills, including their structure and SKILL.md files." See Cursor skills.

Kipt's SKILL.md generator. If you already have a prompt you reuse, the free SKILL.md generator turns it into a skill. It drafts a name and a starting description with a "Use when" sentence, and turns blanks like {{language:TypeScript}} into plain instructions. It checks the name (lowercase letters, digits and hyphens, up to 64 characters) and the description (up to 1,024 characters). It runs in your browser.

Which skill creator should you use?

A reasonable path: start with a hand-written skill or the generator to get a working draft fast, use it for a while, and bring in skill-creator once the skill matters enough to measure. Its description tuning is worth running on any skill that triggers too often or not at all.

Wherever you create them, keep your skills somewhere you own. Kipt stores your prompts and skills as Markdown in a private GitHub repo and installs every skill into Claude Code or Codex with one npx skills add command. For which skills to try first, see the best Claude Code skills.