Chapters — Skills
Claude Code Skills·Advanced·7 min read·Sep 6, 2026

Testing and Refining a Skill

How to actually tell whether a skill is working: it triggers when it should, stays quiet when it shouldn't, and stays easy to read.

A skill is never really "done" on the first write

Writing a skill's first version is the easy part. Knowing whether it's actually any good takes watching it work, or fail to work, on real tasks. This chapter is about closing that loop: how to tell a working skill from one that only looks finished.

Three questions worth asking after every real use

Did it trigger when it should have?
If a task that clearly matched didn't pull the skill in, the description from earlier chapters is the first place to look, it's almost always a wording problem, not a deeper one.
Did it stay quiet when it shouldn't have fired?
A skill that fires on unrelated tasks usually has a description that's broader than the instructions actually justify.
Was the main file still fast to use?
If following it started to feel slow or cluttered, that's the signal from an earlier chapter to split something out into a reference file.

A skill that keeps passing all three, use after use, doesn't need more work. One that's failing any of them has a specific, fixable problem, not a reason to throw the whole thing out.

Real tools built for exactly this

Claude Code has purpose-built ways to check a skill rather than just guessing from a handful of manual tries. A /skill-doctor style report can evaluate a skill's setup and flag likely problems directly. For skills packaged as part of a plugin, claude plugin eval runs a proper eval suite against it, in a sandbox, the same way you'd want any real behavior change tested before trusting it. And when a skill was born informally, built up over a real working session rather than written from a blank page, /run-skill-generator can capture what actually worked into a proper, reusable skill file.

None of this replaces watching it work

Automated checks catch real problems, but the most reliable signal is still simple: use the skill on a real task, in a real conversation, and pay attention to whether it actually helped. A skill that consistently makes a recurring task faster and more consistent is doing its job, whatever a report says.

Bringing the whole guide together

Across this guide: a skill is a saved packet of expertise, reached either by name or by a description that has to earn its match on its own (chapters 1 to 3). A good description states its trigger precisely, and a good file stays lean by pushing depth into reference files only opened when needed (chapters 4 and 5). Not every task belongs in a skill at all, some are better as a subagent, or just asked directly, and once a skill exists, its scope and its boundary against similar skills both need to be deliberate, not accidental (chapters 6 to 8).

None of this is really about Claude Code specifically. It's the same discipline behind any good reusable process: write down what actually works, make it easy to find only when it's relevant, and keep checking that it still does what you meant it to.

Write the real one

Go back to the skill-shaped task from the very first chapter, the one you were asked to keep in mind. Write its name and description now, for real, using everything from this guide. That's genuinely the hardest part done. The instructions underneath it are just writing down what you already know how to do well.

Part of a free guide

Skills

A simple guide to Claude Code Skills: what they are, and how to build your own.

Browse All Guides →