PRODUCTION SCENARIO
A compliance reviewer feeds 40,000-token privacy policies to Claude and asks for GDPR gaps. The answers occasionally cite obligations that do not appear in the policy.
Which prompt change most directly addresses this?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Have Claude pull word-for-word quotes from the policy, then analyze those
For documents over about 20,000 tokens the hallucination guidance is to have Claude extract word-for-word quotes before doing the task, which ties every finding to text that is actually in the policy. Higher effort buys more reasoning, not a link between a claim and its source. Summarizing first puts a lossy paraphrase between the analysis and the document, so an invented obligation survives into the summary. And moving the policy to the system prompt changes how it is weighted, not whether the claims are grounded in it.