Forum · Claude

How do you roll out AGENTS.md, skills and hooks to hundreds of developers?

Asked on Reddit · rewritten without usernames ·

A company is putting Claude Code in front of more than 500 developers on Windows and Mac. The setup is an AGENTS.md file, about fifty skills, guardrail hooks, slash commands and a subagent, all shipped through Git when people clone the repo. How do you push updates without asking everyone to pull, and how do you stop the guardrail hooks from being switched off locally?

Split the setup into two layers, because they have different owners. Project rules like AGENTS.md and project skills belong in the repo, where they change with the code and go through review. Company guardrails belong in a place developers cannot edit, so they should not live in the same Git folder anyone can change in a branch.

For the guardrails, Claude Code reads a managed settings file from a system location that needs admin rights to change. Hooks and permission rules placed there apply to every session and sit above user and project settings. Your device management tool can push that file to Windows and Mac machines the same way it pushes other config, which also solves the automatic update part for that layer.

For skills and commands, package them as a plugin in an internal marketplace instead of copying folders. Developers install it once and get updates from one place, and you can version it, so a broken skill can be rolled back without touching 500 repos.

Fifty skills is a lot to ship to people who did not write them. Test each one before it goes out: give it a few real tasks, check that it triggers when it should and stays quiet when it should not, and drop the ones nobody uses. A smaller set people trust beats a big set they learn to ignore.

Listings mentioned

  • Skill Creator · skill by anthropics
    Runs evals on a skill and tunes its description so it triggers reliably before you ship it.
  • Writing hookify rules · skill by davila7
    Shows how to write hook rules that warn on or block risky commands.

Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.

More in Claude