Claude's skill creator is a skill called skill-creator that builds new skills with you: it asks what the skill should do and when it should trigger, writes the SKILL.md, runs test prompts with and without the skill, shows you the results side by side, then rewrites it from your feedback. Anthropic publishes it in its skills repository, and you start it by asking Claude to use the skill-creator skill to create a skill.
Most guides stop at "ask Claude to make a skill". The interesting part is what happens next, because skill-creator's own instructions describe a full build, test and improve loop. Here's that loop in plain English, so you know what you're signing up for and when it's worth it.
What skill-creator is, in one paragraph
skill-creator is itself a skill, sitting in Anthropic's anthropics/skills repository. Its description says it creates new skills, modifies and improves existing ones, and measures skill performance: running evaluations, benchmarking with variance analysis, and optimising a skill's description so Claude triggers it more accurately. Anthropic's announcement puts the user side simply: Claude asks about your workflow, generates the folder structure, formats the SKILL.md file and bundles the resources you need, with no manual file editing. Its instructions even tell Claude to watch for cues about how technical you are, and to explain words like JSON or assertion before using them.
The stages, in the order it works
| Stage | What skill-creator does | What you do |
|---|---|---|
| Capture intent | Asks what the skill should do, when it should trigger, what the output looks like, and whether to set up test cases | Answer, or say "turn this into a skill" after doing the task in the chat |
| Interview | Asks about edge cases, formats, example files, what success looks like and dependencies | Share a good example of the output you want |
| Write the SKILL.md | Drafts the name, the description and the instructions | Read the draft |
| Test cases | Writes 2 to 3 realistic test prompts | Confirm them or add your own |
| Test runs | Runs each prompt with the skill and a baseline without it, or against the old version when improving a skill | Wait |
| Review | Grades the runs and opens a viewer with an Outputs tab and a Benchmark tab showing pass rate, time and tokens | Click through and leave feedback |
| Improve | Rewrites the skill from your feedback and repeats | Review again until you're happy |
| Description tuning | Writes prompts that should and shouldn't trigger the skill, reviews them with you, then runs an optimisation loop | Edit the test prompts |
Two details explain why this loop exists. First, the description is the part Claude matches your request against when deciding whether to use a skill, so a vague one means a skill that never fires. skill-creator's instructions note that Claude tends to undertrigger skills, and tell it to write descriptions that are a little pushy about when to apply. Second, the baseline run is what makes testing honest: you see the same prompt answered with and without the skill, so you know whether the skill earns its place.
How it changes in Claude chat and Cowork
This is the part nobody mentions. skill-creator was written to run fully in Claude Code, and its own instructions list what changes elsewhere:
- In Claude chat, there are no sub-agents, so Claude runs the test prompts itself, one at a time. The instructions admit that's less rigorous, because the same Claude wrote the skill. There's no browser viewer, so results appear in the conversation. Description tuning needs Claude Code's command line tool, so it's skipped. Blind comparison of two versions is skipped too.
- In Cowork, there's no display, so the viewer is written as a standalone HTML file you open from a link, and your feedback downloads as a file. Description tuning does work in Cowork, but it waits until you agree the skill is in good shape.
How to start it
- Check skills work for you. On Free, Pro or Max, as of October 2026, that means Code execution and file creation is on in Settings, then Capabilities. On Team and Enterprise, the owner controls it.
- Open a new chat, or a Cowork task if you want the file-based viewer.
- Ask Claude to use the skill-creator skill, and describe the job and when it should trigger. If you've just done the task in the conversation, say "turn this into a skill" and it'll pull the steps from the chat first.
- Answer its interview questions, and give it a real example of a good result.
- Read the test outputs and say plainly what's wrong. That feedback is what the rewrite works from.
- When you're done, ask it to package the skill. On Claude.ai its packaging script produces a .skill file you can download and keep.
Copy this prompt
Use the skill-creator skill to build a skill that turns my sales call notes into a follow-up email for the prospect.
It should trigger whenever I paste call notes and ask for a follow-up.
Ask me your questions first, then test it on two example calls before we save it.
What you get at the end
A skill folder: SKILL.md with its YAML frontmatter and instructions, plus optional folders for scripts, reference documents and assets such as templates. skill-creator follows the same three-level loading Anthropic uses everywhere: the name and description are always in view, the SKILL.md body loads when the skill triggers, and bundled files load only as needed. Its own guide says to keep SKILL.md under 500 lines and to point to reference files rather than stuffing everything into one place. It also saves the test prompts, so you can rerun them after future edits.
One thing to read before you save: if every test run ended up writing the same helper script, skill-creator's instructions say to bundle that script into the skill. That's sensible, but it means your skill now runs code. Anthropic's advice for any skill is to review every bundled file, scripts included, before you enable it, and a skill you built yourself is no exception.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
When to skip it and write the file yourself
skill-creator is a lot of machinery for a ten-line skill. The Help Center's own best practices are to start simple, use examples and test incrementally, and skill-creator itself says that if you'd rather skip the evaluations and work freely, it'll go along with that. Its instructions also note that skills with subjective output, such as writing style, often don't need formal test cases at all. So I'd use skill-creator for any skill you'll rely on every week, especially one that produces files or data where right and wrong are clear. Skip it for a throwaway skill, or a short voice guide you can write by hand in ten minutes. If you're weighing plans, my breakdown of Claude's plans shows what each includes. For running it in Cowork, my Claude Cowork guide covers the basics, and the format it writes is the open one in my Agent Skills explainer.
If you'd like your team to build and test its own Claude skills with skill-creator, with someone guiding the first few, that's the kind of AI training I run. I run AI training and workshops for teams, and help businesses with AI enablement and workflow automation, so the busywork runs itself and people get time back. Most of that work is with SaaS marketing teams and small businesses. Here's how we can work together.
Frequently asked questions
What is Claude's skill creator?
It's a skill called skill-creator in Anthropic's skills repository. It creates new skills, improves existing ones, tests them against a baseline and tunes the description so Claude uses the skill at the right time.
How do I use the skill-creator skill?
With skills enabled, ask Claude to use the skill-creator skill and describe the skill you want and when it should trigger. It interviews you, drafts the SKILL.md and tests it.
Does skill-creator work in Claude chat?
Yes, with changes. In chat it runs test prompts itself one at a time, shows results in the conversation, and skips description tuning, which needs Claude Code's command line tool.
What does skill-creator test?
It runs a few realistic prompts with the skill and with a baseline, either no skill or the old version, then shows outputs and a benchmark of pass rate, time and tokens.
Do I need skill-creator to make a Claude skill?
No. A skill can be a single SKILL.md you write yourself. skill-creator is most useful for skills you'll rely on often.
