In my experience with Claude Code and Sonnet, it is absolutely possible to have architectural and design-oriented conversations about the work, at an entirely different and higher level than using a (formerly) high-level programming language. I have been able to learn new systems and frameworks far faster with Claude than with any previous system I have used. It definitely does require close attention to detect mistakes it does not realize it is making, but that is where the skill comes in. I find it being right 80% of the time and wrong 20% of the time to be a hugely acceptable tradeoff, when it allows me to go radically faster because it can do that 80% much quicker than I could. Especially when it comes to learning new code bases and exploring new repos I have cloned -- it can read code superhumanly quickly and explain it to me in depth.

It is certainly a hugely different style of interaction, but it helps to think of it as a conversation, or more precisely, a series of individual small targeted specific conversations, each aimed at researching a specific issue or solving a specific problem.