Question

A team builds a Claude-powered internal help desk that answers questions from a company knowledge base. During testing, they notice that when a user asks about a topic the knowledge base does not cover, Claude confidently invents a plausible-sounding but incorrect policy. The retrieval step is working correctly and only returns relevant passages. Which change to the prompt would most directly reduce these fabricated answers?

Accepted Answer

Explicitly instruct Claude that if the provided knowledge base passages do not contain the answer, it should say it does not know rather than guess.

Answer

Increase the temperature so Claude produces more varied responses across attempts.

Answer

Raise the max_tokens limit so Claude has more room to fully explain each policy.

Answer

Add a role prompt telling Claude it is a highly experienced, authoritative policy expert.