I reach for Claude Code hooks when I’ve told an agent the same safety rule twice and I want the repository to enforce it the third time. Hooks let you run command, HTTP, LLM prompt, or agent hooks around the session lifecycle. They are powerful because they can influence tool use and decisions. That also makes them risky.

Common Events

Claude Code exposes many hook events, including:

• Session start and end.
• User prompt submission and expansion.
• Pre-tool and post-tool use.
• Permission requests and denials.
• Tool failure.
• Subagent start and stop.
• Task creation and completion.
• Pre-compact and post-compact.
• File changes.
• Working directory changes.
• Configuration changes.

Use the narrowest event that fits. A broad hook that fires on every tool call should be very small and very fast.

Hook Locations

Hooks can come from user, project, local, managed, plugin, skill, and agent configuration. That means a hook might be inherited from a place other than the repository you are currently reading.

When diagnosing surprising behavior, inspect the hook sources before blaming the model.

Good Hook Uses

Good hooks are deterministic:

• Block edits to generated files.
• Add a required instruction when a file pattern is touched.
• Log tool use for local audit.
• Deny shell commands that target production.
• Run a fast formatter check after a specific tool action.

Bad hooks try to run the whole development process invisibly. Put real quality gates in scripts and continuous integration.