How to keep your team's AI skills from sprawling: 5 rules

Why skill libraries sprawl

An AI chooses a skill by its name and description. The Agent Skills format loads only those two fields at the start, and the full instructions when a task matches. Claude's authoring guide says the description is what lets the AI pick the right skill from potentially more than 100. So two skills called "proposal-writer" and "proposal-helper" with similar descriptions don't just waste space. They make the AI guess.

Copies make it worse. Someone needs the proposal skill to mention the team's discount rule, so they copy it and edit the copy. The original gets a fix next month; the copy doesn't. After a quarter, the library holds the fix in one place and the discount rule in another, and no skill has both.

Rule 1: one skill per task

Before adding a skill, search the library for the task. If one exists, improve it. Write the description so it names the task and the moment to use it ("Drafts a client proposal from call notes. Use when a prospect asked for a proposal"). Claude's guide also advises consistent naming across a collection, which makes duplicates easy to spot.

Rule 2: an owner for every skill

Each skill gets one person who answers for it: they approve changes, read feedback and decide when it's out of date. Without an owner, every teammate edits a little, and nobody can say why a step is there. In AI Skill Pod on Business, a skill can require its owners to approve edits before they go live.

Rule 3: living skills, not forks

A fork is a copy that follows nothing. It's right when your version genuinely diverges. It's wrong when you only need to add a few of your own rules, terms or examples, because every improvement to the original stops reaching you.

• Living skill: follows the latest version of its source. In AI Skill Pod, an official library skill added on Team or Business follows the latest official version; on Free, it's a snapshot of the version you added.
• Your additions on top: on Team and Business, an official skill can carry your team's own rules, terms and examples, served after the official instructions, which keep updating.
• When you do fork: on Business, when the official version moves ahead of your team's copy, an admin's AI prepares a merged update that keeps your changes, and an admin approves it.

Rule 4: test before you change

A shared skill's edit reaches everyone at once, which is the point and the risk. Keep a small eval set for each skill people rely on: Anthropic's eval guide says 20 to 50 simple tasks drawn from real failures is a great start, and Claude's authoring guide recommends building evaluations before writing extensive instructions. Run the set before an edit goes live and compare the results with the last version.

1. Collect 20 real inputs for the skill, including the ones that went wrong.
2. Write what a pass looks like for each: must include, must not include.
3. Run the current version and record the pass count.
4. Run the edited version on the same inputs.
5. Ship the edit only if it passes at least as many, and add any new failure to the set.

Rule 5: a review date

Every quarter, sort the library by use. A skill nobody ran in 90 days gets a question to its owner: still needed, merge it into another, or retire it. A smaller library is easier for people to trust and for the AI to choose from. On Team and Business, AI Skill Pod's usage log shows who used which skill and when, which makes this a ten-minute job.

Questions

What is skill sprawl?

When a team's AI skill library grows into many near-duplicate, unowned and outdated skills. The AI then picks the wrong one, because it chooses by description and several descriptions say almost the same thing.

What is the difference between a fork and a living skill?

A fork is a copy you edit; it never receives the original's updates. A living skill follows the latest version of its source. In AI Skill Pod, an official library skill in a pod on Team or Business follows the latest official version, and your team's own rules sit in an addition under it instead of in a copy.

How many skills should a team have?

As many as there are distinct, repeated tasks, and no more. A skill nobody used in 90 days is a candidate for removal.