The Skill Type Anthropic Trusts Most (Most Teams Skip It)

Anthropic runs hundreds of skills internally. Most teams building with Claude use two of the nine types available. That's not a knock on those teams, it's just what happens when you don't know the other seven exist.

First thing to get straight: a skill isn't a markdown file. That's the part everyone gets wrong. A skill is a folder. Instructions, scripts, and a section of gotchas Claude actually pulls from when it's working. Anthropic sorts theirs into nine categories, everything from API reference to code review to deployment.

The one that matters most isn't the flashy one. It's verification.

Here's why. Claude loves to look done. It will confidently tell you the tests pass, the deploy went fine, the thing works. Verification skills exist to check that against reality instead of Claude's own optimism.

Write for the Model, Not for You

The description field on a skill isn't documentation for your teammate. It's the trigger. Claude reads that description to decide whether this skill is relevant right now. If you write it like a Jira ticket summary, Claude won't reach for it at the right moment. Write it like you're explaining to Claude exactly when this applies.

And load up the gotchas section. Not the happy path, the docs already cover that. The gotchas are the stuff that isn't written down anywhere else, the thing that bit you last time, the edge case that broke prod.

A skill is only as good as the failure it prevents, not the task it describes.

Stop Hoarding Skills in Your Own Repo

Once you've built a few good ones, check them into your repo or stand up a team marketplace. Skills that live in one person's head or one person's Claude Code config aren't skills, they're a personal cheat sheet.

Go find the verification gaps in your own workflow first. That's where the leverage is.