How do I get ChatGPT or Claude to stop agreeing with everything I say?

Ask for the strongest case against you every time, and have it checked against what you actually know, so agreement means something when it comes. I wrote a module for this, steelman, which anyone can add from the alexandria. marketplace, and with it every idea your model gives you comes with its strongest honest objection in the same reply.

If you use one app, a line in its custom instructions asking for the best argument against your view gets you some of the way. What a line alone can’t do is check that argument against what you’ve already thought through.

Your model reads the blueprint and builds your system on your computer, and the method it comes with already pushes back in every session. Steelman carries that into every reply. Your model tests each counter against your files and the facts first, and drops any your record already answers. When one holds, it says so plainly, with the fact or number that makes it bite, instead of flattering you. When none does, it agrees and says why, and either way it says which side it lands on. Say steelman this for a full red-team of whatever’s on the table.

It stays quiet when you’re carrying out something you’ve already decided, asking a plain question, or sharing a feeling, because the aim is your thinking improved, never an argument for its own sake.

The positions you rely on most change only when you decide. Your model writes the case for a change, you talk it through with a model from a different company, and nothing is rewritten until you say yes in your own words.

Show the blueprint to your model, and it’ll tell you what you’re missing, if anything.