@li9292: Thanks to Teacher Kazik for introducing the Hook approach. Let me also share two of my daily use cases: **Case 1: Passive Observation** In my thought project, I just want to know what Claude is doing. ```json "PostToolUse": [{ "matcher": "Agent|Skil… ```

Thanks to Teacher Kazik for introducing the Hook gameplay. Let me also share my two daily use cases: **The First: Passive Observation** In my thinking projects, I just want to know what Claude is doing. ```json "PostToolUse": [{ "matcher": "Agent|Skill", "hooks": [{ "command": "jq -r '...' >> log.md", "async": true }] }] ``` Every time Claude calls an Agent or Skill, append a line to `log.md`: timestamp, tool name, description. At the end of the session, append a `SessionStop` entry. That's it. No intercept, no intervene, no block. Pure observability. Through Hooks and logs, I solved a long-standing problem: What did the AI actually do? Not what it says it did, but what it actually called. The two are often different. **The Second: Active Intervention** In learning projects, I want to force myself to actually learn, rather than letting Claude write code for me. Five hooks, chained into a state machine: - `UserPromptSubmit`: Check if the session is in "learning mode"; if so, activate subsequent interception - `PreToolUse (Write|Edit)`: Directly intercept writing code, `exit 2` to exit, tell Claude not to write now, teach first - `PostToolUse (Bash)`: After each command run, force a follow-up question: "Do you understand what just happened?" - `Stop`: Before Claude prepares to end, check if this round truly guided the user to think; if not, continue - `SessionEnd`: At session end, extract concepts learned this time and write to memory file **What Are Hooks** Hooks are essentially interfaces for injecting shell commands into Claude's lifecycle. ``` UserPromptSubmit → PreToolUse → [Tool Execution] → PostToolUse → Stop → SessionEnd ``` Each node can do: - `exit 0`: Allow, say nothing - `stdout`: Inject text into Claude's context (it will read it) - `exit 2`: Block, force Claude to reconsider The difference is who has the final say. Normally, Claude has the final say; with Hooks, you have the final say. **Why I Think Hooks Are Underrated** Prompts can be forgotten. CLAUDE.md rules can be bypassed. But Hooks are shell commands: they won't be "misunderstood", won't be "flexibly interpreted based on context", and won't fail due to long conversations. They are hard constraints, not soft suggestions. My current understanding of "custom AI workflows" is: Prompts take care of intent, Hooks take care of discipline.

"PostToolUse": [{
  "matcher": "Agent|Skill",
  "hooks": [{
    "command": "jq -r '...' >> log.md",
    "async": true
  }]
}]

UserPromptSubmit → PreToolUse → [Tool Execution] → PostToolUse → Stop → SessionEnd