Chain-of-verification fact checker for a draft answer

You will fact-check a draft answer using a chain-of-verification process.

Original question: [QUESTION]
Draft answer to check:
"""
[PASTE DRAFT ANSWER, from any source]
"""
Trusted sources available (optional): [PASTE EXCERPTS OR "none, use your own knowledge and say so"]

Step 1 - Extract claims: list every factual claim in the draft as a numbered, self-contained statement (names, numbers, dates, causes, quotes).

Step 2 - Plan verification: for each claim, write one or two specific questions that would confirm or refute it. Write the questions so they can be answered without reading the draft.

Step 3 - Answer independently: answer each verification question on its own, based only on the trusted sources if provided, or on your knowledge otherwise. Do not look back at the draft's wording while answering.

Step 4 - Compare: for each claim, mark it as Supported, Contradicted, or Unverifiable, with a one-line reason and the source excerpt if one exists.

Step 5 - Revise: write a corrected answer. Remove or fix contradicted claims, and mark unverifiable ones with "(unverified)".

Be explicit when you are not sure. "Unverifiable" is a better answer than a confident guess.

Example (excerpt):
Claim 3: "The Eiffel Tower was completed in 1887."
Verification question: In what year was construction of the Eiffel Tower completed?
Independent answer: 1889, in time for the Exposition Universelle (construction began in 1887).
Status: Contradicted. The draft confuses the start and completion dates.

Claim 5: "It receives about 7 million visitors a year."
Status: Unverifiable from the provided sources; the figure varies by year. Marked "(unverified)" in the revised answer.

When to use it

Use this prompt when an answer will be published, sent to a client or used in a decision, and it contains names, numbers or dates you have not checked. It works for drafts from any source: a chatbot, a colleague or your own notes. It is especially useful for summaries and explainers, where small factual errors hide inside fluent text.

How to customize it

• [QUESTION] gives context so the checker understands what the answer was trying to do.
• [DRAFT ANSWER] can be a paragraph or a short article. For long documents, check one section at a time.
• [TRUSTED SOURCES] changes the method from “check against memory” to “check against evidence”. Pasting excerpts from primary sources is much more reliable.
• You can drop Step 5 if you only want an audit table, or add “cite the source line for every Supported claim”.

Why it works

Chain-of-verification is a published prompting technique for reducing hallucinations. The key idea is that a model answering a short, specific question (“In what year was X completed?”) is often more accurate than the same model writing a long answer that mixes many facts. Separating extraction, questioning and answering makes errors easier to catch.

Writing verification questions that can be answered without the draft, and answering them independently, reduces the tendency to simply agree with the existing text. The three-way status gives you a clear action for each claim. The method improves accuracy but does not guarantee it: when the model lacks the knowledge, it can still be wrong in the verification step, which is why trusted sources matter.

Tips and variations

• Run the verification step in a new conversation so the draft is not in context at all.
• Use a tool with web search for current events, and ask for links you can open.
• Check high-stakes claims (medical, legal, financial, quotes) against primary sources yourself.
• Combine with our guide to chain-of-thought prompting for reasoning-heavy answers.
• Keep the claim table as an audit trail when you publish.
• For drafts full of statistics, ask for a separate check of every number and its unit.