Claude Code hooks turn prompt requests into hard rules

The article frames hooks as a shift in how rules are enforced when working with an AI coding agent. Asking in a prompt for the agent to run a formatter after every edit is a request; the guide's point is that it may or may not be honored. A hook instead runs a user-defined command whenever a specified event occurs, so the behavior is no longer left to the model's judgment.

The three hands-on examples map to three different moments. PostToolUse fires right after a tool runs, so editing a file can trigger a formatter such as prettier for JavaScript and TypeScript, ruff for Python, or gofmt for Go, keeping the diff aligned with the formatter. PreToolUse fires just before a tool runs and is the one that can block: the sample script greps the incoming command and rejects destructive or irreversible patterns, exiting with code 2, which sends its stderr message back to Claude so the agent knows why it was stopped and may try another approach. Stop fires when the agent finishes responding and carries no matcher, making it a place for lightweight checks like lint or a type check.

The guide is candid about what comes with this power. Hooks execute shell commands with the user's own permissions, so it warns against using an untrusted repository's .claude/settings.json without reviewing it. It also notes that a hook can hang and stall the session, that a PostToolUse formatter rewriting files can create loops if the hook itself calls the agent's tools, and that hook specifications may change by version. Whether these hooks pay off in practice is likely to hinge on how carefully teams keep them fast, scoped, and reviewed — and on whether the enforcement they buy is worth the shell access they grant.