I've stumbled on the same workflow. Except for one thing: If I just do as OP does, Claude Code will tend to overengineer. For example it'll build complex solutions to super rare race conditions that have trivial fallout. But I've found that all it takes is a "skeptical pass". Here's how it goes: After having a bunch of specialist subagents review the (plan/implementation), after doing the deduplication/synthesis of their findings, the main agent will bucket them into A) Trivial/obvious fix B) there's multiple possible resolutions, but the LLM had a strong lean, so it went with it on its own C) Genuine ambiguity, where it asks me what to do (and presents its lean) and D) Wontfix. Crucially, after doing this, I have it run a "skeptical pass" where it takes a hard look at these findings and see if maybe some of them deserve to be downgraded. Generally, a lot of things make their way into wontfix this way. I find, I don't need to push back against overengineering, I can have the LLM do so itself, and it'll actually do a decent job of it.

This sounds harder than just writing the code.