I’ve been building text editor as notion from the ground up. Recently i’ve been uploading the full codebase to Gemini to discuss different tasks. I can fully agree that the LLM can simply not decide on how to do things! If you slightly tweak your prompt it will suggest an entirely different architecture. You have to be very careful about pulling out the right stuff from the cargo cult best practice stuff baked in to the model. Another thing, a really obvious optimization might be just staring it in the face and it won’t suggest it unless you say “have you considered XYZ?” and then all of a sudden it won’t stop talking about XYZ! But most of all you just have to be strong in your will power to guide it in the direction you want.