If you have to have a prompt monkey to figure out the best prompts, that defeats the whole point of having an LLM powered knowledge system.