Source: Reduce hallucinations (Claude Docs)

Anthropic's guide to prompting techniques that reduce hallucinations: allow uncertainty, ground in quotes, cite and verify.

Contents

Summary

A practical checklist of prompt-level hallucination controls, split into basic and advanced techniques, with worked examples (M&A analysis, privacy-policy audit, press release). It ends by stating the techniques reduce but do not eliminate hallucinations, so critical outputs still need validation.

Key claims

  • Allow “I don’t know”: explicit permission to admit uncertainty substantially reduces false information.
  • Quote first: for long documents (over about 20k tokens), have the model extract word-for-word quotes, then work only from them.
  • Cite and verify: require a supporting quote per claim; after drafting, have the model find a quote for each claim and retract claims it cannot support.
  • Advanced: review thinking output for faulty reasoning; best-of-N sampling and compare (inconsistency signals hallucination); iterative refinement; restrict the model to the provided documents rather than general knowledge.

Feeds

Hallucination

Suggest changes →