Claude Code fires 30 events per session you can hang your own code on any of them the 6 that changed how I work: A hook is a shell command that runs at a fixed moment. It isn't a prompt that's the whole point. The model can reason its way around an instruction. It can't reason its way around a hook 1. PreToolUse → before a tool runs, and it can block it mine refuses any write to .env, and any rm it didn't explain first 2. PostToolUse → after a tool succeeds mine runs the formatter on whatever file it just touched. I stopped asking for tidy code 3. UserPromptSubmit → after I hit enter, before the model sees it mine attaches the branch and the last commit. every prompt. I never type it 4. Stop → the turn is finished mine runs the tests. If they fail, the turn isn't finished 5. SessionStart → a session begins or resumes mine prints the three things I changed yesterday and always forget 6. PreCompact → the moment before it compacts mine saves the transcript. I've lost one. Not twice they all live in .claude/settings.json. A matcher decides which tools they fire on the other 24 are real (SubagentStop, FileChanged, PermissionDenied, WorktreeCreate). I've never needed one start with PostToolUse. it's the one you feel today Bookmark this. This is the one.