# /skill-creator: Create, test and tune your own agent skills

Source: https://trendhub.net/skills/anthropics-skills/skill-creator
Skill folder: https://github.com/anthropics/skills/tree/main/skills/skill-creator
Pack: Anthropic Agent Skills (https://trendhub.net/skills/anthropics-skills)
Works with: Claude Code

skill-creator helps you write a new skill or improve an existing one through a loop of drafting, testing and revising. Claude runs the draft against realistic sample prompts, shows you the results for review, and can benchmark the skill and refine its description so it triggers on the right requests.

## When to use it

- You want to turn a workflow you just did with Claude into a reusable skill
- An existing skill gives uneven results and you want measured improvements
- A skill rarely triggers, or triggers on the wrong requests, and its description needs tuning

## How it works

1. **Capture intent**: Claude clarifies what the skill should do, when it should fire and what good output looks like.
2. **Draft the skill**: It writes a SKILL.md and any supporting scripts or reference files.
3. **Run test prompts**: It runs a few realistic prompts with the skill and presents outputs and metrics for your review.
4. **Revise and repeat**: Your feedback and benchmark results drive rewrites until the results hold up.
5. **Optimize the description**: An optional loop scores trigger accuracy on a train and held-out query set and picks the best description.

## What it steers the agent away from

- Shipping a skill that has never been tried on real prompts
- Vague descriptions that cause a skill to be skipped or to fire on unrelated tasks
- Overfitting a description to the same queries used to tune it

## What it can touch

Writes skill folders and test workspaces on disk, runs bundled Python scripts (including a description-optimization loop that calls Claude repeatedly), and can generate and open local HTML review pages in a browser. In Claude.ai it adapts by running tests inline and skipping the browser viewer.

## Pairs well with

- `/mcp-builder`
- `/doc-coauthoring`
- `/webapp-testing`

## Questions

### Do I have to run formal evaluations?

No. The skill is meant to adapt to where you are, and it can work informally if you just want help drafting and iterating.

### Does it work outside Claude Code?

Yes, with reduced features. In Claude.ai there are no subagents, so tests run one at a time and quantitative benchmarking is skipped.
