From my experience Claude 3.7 seems to make less mistakes when you provide detailed specs. It saves a lot of back and forth, especially at the beginning, when you're trying to 'de-noise' the idea. I find that the best way is using a hybrid approach: going between iterating over specs to iterating over code, and once I have the MVP I update the specs one last time. I pass these specs to another session and continue working. It solves multiple problems at once: context getting full, LLM getting lost etc.