Chain-of-Verification Fact Check Prompt

Draft an answer, generate verification questions, answer them independently, then revise, reducing factual errors.

Based on Chain-of-Verification (Dhuliawala et al., 2023).

Steps

1. Baseline: answer the question.
2. Plan verifications: list factual claims as questions.
3. Execute independently: answer each question without showing the draft (separate calls), ideally with retrieval/search.
4. Revise: rewrite the answer, dropping or correcting claims that failed verification.

Prompts

# Step 2
List each factual claim in the ANSWER as a standalone yes/no or short-answer question.
Return JSON: {"questions": ["..."]}
ANSWER: {{draft}}

# Step 4
Original question: {{question}}
Draft: {{draft}}
Verification results: {{qa_pairs}}
Rewrite the draft so it is consistent with the verification results.
Remove any claim that could not be verified. Do not add new unverified claims.

Failure modes

• The model can 'verify' a wrong claim with the same wrong knowledge; pair with retrieval for facts.
• Verification questions that just restate the claim add cost without catching errors.
• Long drafts produce too many checks; cap the number of questions.

Markdown version: /blueprints/chain-of-verification.md

Related blueprints