One trick I do to avoid this: When it's doing something wrong, and I see some pattern based on which I'd do it differently, I tell it the observation rather than my conclusion. If I tell it stuff like what about X, it will just agree that X is a good idea. But if I say, I see that Y, it might come to the same conclusion that it needs to do X, or something completely different. It will still say I'm absolutely right, but at least it's to something where I am absolutely right. I mostly use Claude Code to fill the blanks in my thinking, so this works better for me.