You don't need to know the whole solution before you go to AI. Tell it the problem and work it out together. The solution shows up through the back and forth.
AI won't get it right the first time, and it doesn't know it's wrong until you say so. Just tell it plainly what's off and it corrects fast. That's how you actually get to the thing you wanted instead of settling for the first draft.
The interesting part is that correcting it is also how you find out what you actually think. Saying "no, not like that" forces you to know what you meant. Talking to AI is already thinking out loud, so the raw material is free.
But there's so much communication happening with AI now, and most of it just disappears into chat history. The principles are in there, but they don't pull themselves out. You have to be thoughtful about distilling them, otherwise it's just a transcript and you keep re-deriving the same lesson.
I've seen at work that distilling principles takes real time, which is exactly the part AI can help with. You say what's wrong, it fixes it, and then it can help you turn the correction into something you keep.
It only counts if it changes a decision though. People love sayings that are true but weak in practice. Don't Repeat Yourself is the classic one. A principle you pulled out of a real correction is different, because you already know the decision it changes.