How to stop skill version drift on a team

How drift happens

The owner writes a "client proposal" skill and sends the ZIP to five people. Three upload it in Claude, one pastes it into a ChatGPT project, one forgets. A month later the owner changes the fee wording and sends a new ZIP. Two people re-upload; three don't. Now there are four versions, and nobody can tell which one wrote the proposal a client queried.

Nothing is broken; this is what copies do. Every route that hands out files ends the same way.

Three ways to keep one version

From the Claude, Claude Code and OpenAI help pages, read 5 November 2026. More detail: seven ways to share AI skills compared.

A worked example: moving off ZIPs

Illustration: a made-up business. A 9-person marketing agency has eight skills in circulation as ZIPs, in an unknown number of versions. The operations lead does three things in one afternoon:

1. Collects every version people actually have and keeps the newest of each skill, with its owner's sign-off.
2. Puts the eight in one shared place every person's AI reads, and asks people to delete their uploaded copies.
3. Writes ten test requests for the most-used skill, the client report, with what a good answer looks like.

Two weeks later the account lead edits the client report skill to add a "next month" section. Before publishing, she runs the ten test requests: one comes back with the new section in the wrong place. She fixes the instruction, re-runs, publishes. Every report from then on has the section, in the right place, for all nine people.

Roll out a change safely

1. Keep a short test set for each skill people rely on: 10 to 20 real requests and what a good answer contains.
2. Edit, then run the tests before you publish.
3. Publish to the one shared copy, not as a new file.
4. Say what changed in one line, where the team will see it.
5. Remove old copies from people's uploads and notes, so the shared one is the only one.

For rules that keep the library itself tidy, see five rules for keeping team skills from sprawling. AI Skill Pod is one such shared library, from Free (up to 3 members and 10 skills).

Questions

Why do shared skills go out of date?

Because a ZIP file is a copy. Each person uploads the version they received, and when the owner improves the skill, nothing reaches the copies already uploaded. After a few edits, the team runs several versions.

How do I make sure teammates always have the latest skill?

Keep one copy that everyone's AI reads: a git repository for Claude Code users, skills shared inside a Claude Team or ChatGPT Business workspace, or a skill library every AI connects to over MCP.

How do I roll out a change to a shared skill safely?

Test it first on a small set of real inputs with known good outputs, then publish it, and tell people what changed. With one shared copy, a mistake reaches everyone, so the test matters more.