Preview
What it is.
It pulls the process out of a conversation where you already did the work, drafts the skill, writes test prompts, runs them, helps you score the results, then rewrites from whatever failed.
A separate script tunes the description, which decides whether the skill fires on the right request at all. Your brief format and your QA gates are the things no public library will ever hold, and this is how they become repeatable.
What you get.
- A drafted skill with a name, a trigger description and instructions
- Test prompts and a run, so you see the skill behaving before you rely on it
- Quantitative evals alongside your own read of the output
- A review page for looking at the results side by side
- A description improver that tunes when the skill fires
How to set it up.
Install the example-skills plugin from the anthropics/skills marketplace.
Start from a conversation where you already did the work. It extracts the tools, the sequence and your corrections from the history, which is faster than describing the process cold.
Answer the four setup questions: what it should enable, when it should trigger, the output format, and whether the output is objectively checkable.
Say yes to test cases when the output is verifiable, such as a file transform or a set sequence, and skip them for subjective work like writing style.
Run the test prompts and read the outputs yourself before you look at any score.
Rewrite from what the evaluation showed, then widen the test set and run it again.
Pricing Plans
Free. Apache 2.0 license.
Included in Anthropic's example-skills plugin. Checked against the repo in July 2026.
Use cases
Bottling a finished workflow
You just walked Claude through your brief format by hand, again. Start the skill from that conversation and it extracts the sequence and your corrections from the history, then drafts the skill for you to test.
Testing before trusting
The drafted skill handles a file transform, so the output is checkable. Say yes to test cases and run the prompts, then read the outputs yourself before looking at any score.
A skill that never fires
You wrote a skill and requests keep going past it. Run the description improver on it as a standalone step, since the description is what decides whether the skill triggers at all.
Second-round tightening
An existing skill works most of the time. Enter the loop at the evaluation stage and widen the test set from what failed, then rewrite against the misses instead of starting over.
Best for
A process only your team runs
Your brief format and your QA gates are the things no public library will ever contain, and this is how they become repeatable.
A skill that never triggers
The description improver exists for exactly this, and it is a separate script you can run on an existing skill.
Improving something you already wrote
You can enter the loop at the evaluation stage instead of starting from a blank draft.