Claude skill creator: build and test your own skills with skill-creator

Writing a skill is easy. Writing one that reliably changes Claude's results is not, and most people never find out which they have, because they never test it. skill-creator is Anthropic's answer: a skill for building skills, which interviews you, writes the files, and then measures whether the skill actually does better than Claude without it.

This guide covers how to install it, the workflow step by step, what the evaluation does, and how to get a skill from your laptop to your team. For what skills are in the first place, see Claude skills explained.

Facts are from skill-creator's own SKILL.md and the anthropics/skills repository as of October 2026.

Install

Claude Code, on its own from the official marketplace:

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

Or as part of Anthropic's example skills:

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

Claude.ai: Anthropic's example skills, skill-creator included, are available in the app. Enable it under the Skills settings; code execution must be on.

Then ask for what you want: "Create a skill for writing our release notes" or "Improve the skill in .claude/skills/incident-review and test it".

The workflow

skill-creator follows a fixed loop, which is most of its value:

1. Capture intent. It interviews you: what the task is, what good output looks like, what goes wrong today. It may research the domain.
2. Write the skill. A SKILL.md following Anthropic's writing guidance: progressive disclosure (short instructions, details in reference files) and no surprises.
3. Write test prompts. Realistic requests the skill should handle.
4. Run with and without the skill. Each test runs in parallel sub-agents, once with the new skill and once without (or against the previous version).
5. Draft assertions while the runs execute: what must be true of a good answer.
6. Grade. A grader agent scores each output against the assertions.
7. Benchmark. Pass rate, time and tokens are aggregated as mean ± standard deviation, with the difference from the baseline.
8. Analyse. An analyst agent looks for patterns in what failed.
9. Review. An HTML viewer shows the results so you can give feedback.
10. Revise and repeat.

Steps 4 to 7 are what almost nobody does by hand. They answer the question that matters: is Claude better with this skill than without it? A surprising number of skills make no difference, and some make results worse by adding rules the model already followed.

Fixing skills that never trigger

The most common skill problem is not the instructions, it is the description: Claude never loads the skill because the description does not match how people phrase requests. skill-creator has a dedicated optimisation step for this, in Claude Code:

• It writes about 20 queries, some that should trigger the skill and some that should not.
• It splits them 60/40 into training and test sets.
• It runs each query three times and iterates on the description up to five times.
• It keeps the description that scores best on the test set, so it is not just memorising the training queries.

If a skill you already have is ignored, run this before rewriting anything else. Our skills guide has the manual checklist.

Claude Code vs Claude.ai

Use Claude.ai to draft a skill quickly; use Claude Code when you want to know it works. Claude for Mac also offers a separate way to create a skill by recording yourself doing a task in Cowork, on Pro, Max and Team plans.

What it ships with

skill-creator is itself a good example of a well-built skill: the instructions are in SKILL.md, and the deterministic work is in scripts.

• Scripts for running evals, aggregating benchmarks, generating reports, improving descriptions, validating a skill and packaging it (package_skill.py builds a .skill file).
• Agents: instructions for a grader, an analyser and a blind comparator.
• An eval viewer that turns results into an HTML page.

It is Apache 2.0 licensed, so you can study and adapt it.

A worked example

Say your team writes a weekly customer-health summary from the CRM and support tool.

1. "Create a skill for our weekly customer health summary. It pulls accounts with open escalations from HubSpot and Zendesk and writes a one-page summary for the leadership meeting."
2. skill-creator asks what the leadership team cares about, what the format is, and what a bad summary looks like. Give it last month's best summary as the example.
3. It writes the skill and four or five test prompts ("summary for this week", "only enterprise accounts", "an account with no escalations").
4. The benchmark tells you how often each version follows the format and includes the right accounts, and what it costs in time and tokens. A typical outcome is that the skill version is far more consistent and somewhat more expensive per run; if it is not more consistent, the skill is not pulling its weight yet.
5. You adjust, rerun, and commit the folder to .claude/skills/ in the team repo.

The skill describes the procedure; the data comes from MCP servers. See Claude skills vs MCP for how the two fit together.

From your laptop to the team

A skill in your ~/.claude/skills/ helps you. A skill the team uses needs three more things:

• A shared home. Commit project skills to the repository, or publish them in an internal plugin marketplace.
• Review. Treat a skill like code: a pull request, a second pair of eyes on scripts and instructions.
• Evals kept with the skill. Keep the test prompts and assertions skill-creator wrote next to the skill, and rerun them when the skill or the model changes.

For a company, approved skills belong on the same list as approved MCP servers. Walma AI Hub distributes both to every employee's client under one policy, with every call logged, inside the customer's own EU region. See how it works, or browse the best Claude skills for ones worth starting from.