The Claude hooks vs skills question keeps coming up because skills get most of the attention and hooks get almost none. Both extend Claude Code. They are not interchangeable. The short version of Claude hooks vs skills: skills describe how the agent should approach a class of task, hooks fire at the execution boundary and enforce something deterministic. If the requirement is "teach the agent a workflow," reach for a skill. If the requirement is "this must happen, or must never happen, regardless of what the model decides," reach for a hook.

A skill is a descriptive bundle — a SKILL.md plus optional scripts and assets — that the agent loads into context when it matches the situation. It nudges behavior. The model still decides whether to follow it, when to invoke the scripts, and how to interpret the instructions. Skills are good at "here is the shape of this kind of task."

A hook is a shell command wired to a lifecycle event — PreToolUse, PostToolUse, UserPromptSubmit, Stop, Notification, SubagentStop. The runtime fires it whether or not the model wants it fired. A PreToolUse hook can block a tool call before it executes. A PostToolUse hook can run a linter, a formatter, or a check against the output. The decision lives in code, not in the prompt.

The pattern under the matrix: anything that depends on the model's discretion belongs in a skill. Anything that must hold regardless of discretion belongs in a hook.

Skill only — "writing release notes." A release-notes skill describes the house format, tone, required sections, and how to summarize breaking changes. There is no enforcement requirement; the model drafts, a human reviews. Wiring this as a hook would be wrong — the work is generative and there is no execution boundary to fence.

Hook only — "never commit to main." A PreToolUse hook intercepts git commit and git push, inspects the branch, and exits non-zero on main. The model cannot talk past it. The rule has nothing to teach and everything to enforce. Encoding it as a skill would mean trusting the model to remember across sessions — exactly the failure mode hooks remove.

Both — "edits to catalog data." Catalog edits need a workflow (validate schema, regenerate badges, run tests) and an enforced gate (validator must pass before commit). A skill captures the procedure; a PostToolUse hook on writes to the catalog path runs the validator and blocks on failure. Skill for the shape of the work, hook for the floor under it.

The default failure is over-using skills. Skills are easy to write and easy for the model to ignore. Hooks are slightly more work to set up and impossible to bypass. For anything load-bearing, the hook is the safer answer. For where skills sit relative to other extension surfaces, see skills vs MCP.