enforced-in-code

Identify whether the violation is mechanically detectable with regex or file checks, then enforce it at the lowest layer: permissions.deny for blocking tool patterns, a PreToolUse hook for custom blocking logic, or a PostToolUse hook for silent auto-correction.

Use a PreToolUse hook with exit code 2 when the action must be blocked before it happens, such as rejecting edits to sensitive files. Use a PostToolUse hook when the content is acceptable but needs automatic correction after every write, like frontmatter normalization.

CLAUDE.md loads every turn and costs tokens on all work, so it suits universal behavioral guidance. Path-scoped rules in .claude/rules/ with a paths frontmatter field load only when the agent accesses matching files, costing nothing otherwise.

No. Problems requiring judgment, like prose quality, cannot be fully automated. The recommended approach is to split the problem: automate mechanical symptoms with hooks or lint checks, and guide the judgment portion with the cheapest applicable rule layer.

Avoid automation for one-off problems, rapidly evolving standards still being figured out, and cross-file semantic checks that require understanding relationships across many files. Those belong in batch lint runs or temporary guidance until they stabilize.