Maybe root it in reality.

RAG just puts more info into a prompt. It wouldn't effect the underlying behaviors in any way a non-RAG supported prompt would.

You can get it to be more creative with prompting.

I'm aware of that. My point is that if you have fundamental incoherence in the model behaviors, you're not going to solve it through prompting. To get performance gains through prompting, you first need model behaviors that are able to be improved through prompting. RAG and prompt engineering work on the margins, not the main.