[ManuelKiessling]

To be honest, it feels like you are contradicting yourself.

If LLMs are, as you stated, really good at catching potential issues, then they are, almost by definition, really good at producing code without potential issues, if guided correctly: all they need to do is inspect and iterate, until they do not find any more potential issue in the code they produced.

[layer8]

That isn’t necessarily true. LLMs could be good at finding issues that human miss when coding, while being bad at catching issues that they themselves miss while generating code.