Hooks

A hook is a shell command Claude Code runs at a defined point in its own execution — before a tool call, after a file edit, at other documented events.

The distinction that makes hooks worth understanding: hooks are deterministic. A CLAUDE.md instruction saying “always run the formatter after editing” is a request the model may or may not honour. A hook that runs the formatter after every edit is not a request. It runs.

What people use them for

• Formatting after every file edit
• Running the linter or type checker before a commit
• Blocking edits to protected paths
• Logging what the agent did, for audit
• Injecting context at the start of a session

Hooks vs Skills

The two get confused constantly, and the difference is exactly the one above:

	Hooks	Skills
Triggered by	An event in Claude Code	The model deciding it is relevant
Guaranteed to run	Yes	No
Written as	Shell commands	Instructions, scripts, resources
Good for	Enforcement	Capability

Use a hook when the thing must happen. Use a Skill when the thing needs judgement. Full comparison.

What it does not do

• A hook is not sandboxed. It is a shell command with your privileges. A hook installed by a plugin is a shell command someone else wrote, running with your privileges.
• It is not a permission system. Hooks can block, but permissions are the mechanism designed for that.
• It does not run in the Claude apps. Hooks are Claude Code and Agent SDK.

In the Agent SDK

The Agent SDK exposes hooks as the way to intercept and control agent behaviour programmatically — the same idea, with a code interface instead of a settings file.