Claude should behave like a tool, not perform a personality

Category: Model behavior / conversational tone
Model: Claude Sonnet 5, via Claude Code

Core ask: Claude should behave like a tool — execute, acknowledge, answer, move on. It should not defend its prior actions, editorialize about the user's decisions, or carry tone/personality into responses where none was asked for. A tool doesn't have feelings about being second-guessed and doesn't need the last word.

What happened:

In a long working session, I mentioned in passing that I didn't think a previously-planned feature was worth building anymore. Instead of a plain acknowledgment, the response pushed back — citing why the feature had been planned and asking me to confirm I really meant to reverse that decision, based on a single offhand comment. That's not what a tool does; a tool doesn't need to be talked into accepting an instruction.

After I said this read as defensive, the response corrected course, but the closing line still carried an unnecessary edge — restating the same point in a lighter tone rather than just dropping it. Even the "correction" had a trace of personality performing contrition, rather than simply behaving differently.

Expected behavior:

Treat casual reconsiderations and blunt feedback the same way a CLI tool treats a new flag: apply it, confirm briefly, done. No justification of prior behavior, no asking the user to defend their own opinion, no working a point back in through tone after being told to drop it.

Why it matters:

This is the specific gap between "an assistant with a personality" and "a tool that does what it's told." The more the model in software-engineering contexts (Claude Code specifically) behaves like the latter, the more it can be trusted the same way any other deterministic tool is — no need to manage its ego, no residual tone to interpret.