Does this actually work? I always find it strange when a system prompt tells an LLM to do or not do something that requires domain knowledge to understand. Like, how does the underlying base model at all understand what it means to verify documentation?

we all know that it doesn't work 100% of the time. That's the nature of llms.