After using Claude Code during a cooldown I think it's great for complex tasks with plan mode — but it does require a lot of hand holding.
The biggest thing I came away with: while I read all the code, approved it, and reread the diffs, guided it and modified things, I still didn't feel like I fully understood some of the approaches it took. Maybe worded differently — the mental model of the solutions didn't feel cemented in. Solutions were quick, and quick to fix, but the comprehensive picture was hard to keep in my mind through all of it — especially when I needed to pivot due to feedback on PRs.
On Sunday I came across this article and it kind of crystalized that observation for me.
One quote that struck home:
The question isn't whether LLMs can write code — they obviously can. The question is whether we can maintain the human theoretical frameworks that transform that scattered code into coherent, domain-aware, lasting programs.