I’ve been giving the LLM input prompt sentence structures and sample conversations that it seems to follow pretty deterministically and consistently since my use case is fairly simple requests