Config Files Set AI Agent Intent, Hooks Enforce It — Here's the Difference

A growing distinction in AI coding agent setups separates prose configuration files, like CLAUDE.md, from executable hooks that enforce hard rules at runtime. While config prose shapes what an agent attempts to do, PreToolUse hooks can deterministically block dangerous actions such as force-pushing to main or modifying protected files like .env. Developers are advised to use a reliability ladder: placing linters and CI at the base, hooks for blocking unsafe commands, and prose only for judgment-based guidance that requires context. Hooks handle checkable, repeatable rules without consuming token budget, whereas prose is reserved for nuanced decisions like defining what 'done' means or explaining legacy workarounds. The framework recommends versioning both hook scripts and config files in git to maintain consistent, auditable agent behavior across a codebase.