learn

ClaudeWatch — learn

ClaudeWatch watches every command and records its decision; this skill is the learn step that turns those recordings into suggestions. You vet a window's worth of accumulated prompts once, as a batch, instead of pressing enter on each one. It is the replacement for the generic transcript-scanning approach: it works from the hook's own decisions, so it knows what ClaudeWatch allowed, asked, and blocked — not just what looked read-only.

0. Prerequisite: a decision log to learn from

Logging is on by default — the hook writes to ~/.claude/claudewatch/decisions.jsonl unless CLAUDEWATCH_LOG is set to an opt-out value (off, 0, false, none, or empty). Each Bash record stores the command shape (git push, not the full command with its arguments), never the raw command, so inline secrets stay out of the plaintext log; the file is owner-only (0600). Two states block a learn pass; distinguish them before doing anything else.

First, check whether logging has been turned off:

echo "CLAUDEWATCH_LOG=[${CLAUDEWATCH_LOG:-<unset>}]"

• off / 0 / false / none / empty — logging is disabled. Warn the user loudly that disabling logging disables this skill: /ClaudeWatch:learn has no data to work from and stays useless until logging is re-enabled. To re-enable, remove the opt-out from the env block in their settings.json (unset it, or set it back to the default path) and restart their sessions:

  { "env": { "CLAUDEWATCH_LOG": "~/.claude/claudewatch/decisions.jsonl" } }
  

  Then stop — there is nothing to learn from until logging is back on and sessions have run.

Otherwise logging is on; check whether the log has any records yet:

ls -la "${CLAUDEWATCH_LOG:-$HOME/.claude/claudewatch/decisions.jsonl}"

If the file does not exist, logging is on but no sessions have run with the hook active yet. Tell the user to run some sessions, then stop — there is nothing to learn from until then.