From what I see, you have to get your claude.md instructions right, and most of it is using planning mode. The CLI is exactly how you interact with it. It's less vibe code and what I'd call supervisory code.
It's less about writing code, more about building an engine that codes. It's hands off but unlike Cursor, it's not as eyes-off. You do observe and interrupt it. It is great for full on TDF - it figures out the feature, writes the test for the feature, fails the test, rewrites the code to pass. But at some point, you realize you got the instructions wrong and because you're outside the loop, you have to interrupt it.
I think it's a natural progression, but it's not for everyone. The people who hated not being able to write functions will hate claude code even more. But some people love writing their engines.