TDD would help here, particularly if a human writes - or at least thoroughly reviews - the tests.

https://martinfowler.com/bliki/TestDrivenDevelopment.html

That’s not an impossible task with LLMs, you just have to mindfully architect the project with that in mind, hence take it slowly to design a good system, don’t outsource all thinking to LLMs.

Well they will train on my Claude Code sessions for a start. I spend a lot of time asking it to remove unnecessary code that was produced, I'm not the only one.