[[Claude hallucination]] that broke convention on 2 different worktrees today, as follows: ![[Claude code and vault convention.png]] > **Claude Code:** You're right - I shouldn't have planned a commit. Stopping there. > > **On the commit:** you didn't ask, and there's no convention requiring it. `[[Vault Git Policy]]` says the agent "stops and leaves the changes for me to review and merge by hand" - "leaves the changes" is consistent with leaving them uncommitted in the worktree for you to inspect and stage yourself. I invented the commit step. Deleting that task and saving the rule. ![[Claude code and convention 2.png]] > **Me:** Is it only in memory or is there a convention for that? > > **Claude Code:** Yes, there's a convention - I just wasn't following it. `[[Task note]]` says (inside the `## Log` section): > > > *One entry is appended when a **session finishes**, or once when an **entire worktree run finishes** - never one per question or exchange.* > > So the rule already exists. The memory I just saved is redundant with that convention. Want me to delete the memory and rely on the convention, or keep it as a reinforcement note? What I learned is that my conventions are weaker than internal [[Claude Code]]'s memory, so I should enforence updating it each time I update my conventions.