Check a draft AI answer with Chain-of-Verification

This guide explains how to run Chain-of-Verification when an AI answer needs a separate check before it is accepted as final.

Chain-of-Verification is an answer-checking method. The assistant first writes a draft answer, then creates verification questions for the important factual claims in that draft, answers those questions independently, and uses the results to write a final verified answer.

Use this guide when the first answer may sound polished but still needs claim-level verification. It is useful for research summaries, technical explanations, document reviews, reports, recommendations, public writing, and other outputs that may be sent, published, reused, or used to support a decision.

The purpose is practical: keep supported claims, correct failed claims, remove unsupported claims, and make uncertainty visible before the answer becomes final.

A fluent AI answer can still be wrong. It can include an incorrect date, an unsupported source summary, a product behavior that was not checked, a technical claim that does not match the code, or a recommendation that goes beyond the available material.

The risk is higher in professional work because polished language can make a draft look ready before it has been checked. A user may copy it into a report, send it to a client, publish it, or use it as input for another decision.

Chain-of-Verification reduces that risk by forcing the answer through a separate verification pass. The first answer is not treated as final. It is treated as draft material that exposes the claims that need to be checked.

The working rule is:

The first answer is a draft. The final answer is written only after the important claims have been checked.

Chain-of-Verification is a way to stop treating the first AI answer as the final answer.

In a normal interaction, the assistant answers the task directly. If the answer is fluent, structured, and confident, it can look finished even when some claims are unsupported or wrong.

Chain-of-Verification changes the workflow. The first answer becomes a draft. The draft is then inspected for important factual claims. Those claims are turned into verification questions. The assistant answers those questions separately, using the source material or allowed context for the task. Only after that does it write the final answer.

The method is useful because it creates distance between the answer and the check. The assistant is not just asked, “Is your answer correct?” That usually produces a weak check. Instead, the assistant is forced to identify specific claims and test them one by one.

Simple example

A draft answer might say:

The supplied policy requires approval before storing client documents.

Chain-of-Verification does not accept that sentence because it sounds clear. It turns the sentence into a question:

Does the supplied policy say approval is required before storing client documents?

Then the assistant must answer that question from the supplied policy. If the policy does not mention approval, the claim is not verified and should not appear as a confirmed fact in the final answer.

What changes in the final answer

Without Chain-of-Verification, the final answer may repeat the unsupported claim:

The policy requires approval before storing client documents.

With Chain-of-Verification, the final answer should reflect the check:

The supplied policy says client documents must be stored in the approved workspace. I did not find support in the supplied policy for the claim that approval is required before storing them.

This is the practical meaning of Chain-of-Verification: the assistant does not just produce an answer. It produces a draft, checks the claims in that draft, and rewrites the final answer so unsupported claims are corrected, removed, or marked as not verified.

The method is not a general reminder to “be accurate.” It is a concrete checking process with visible intermediate outputs:

• a draft answer;
• a list of important factual claims;
• verification questions for those claims;
• independent answers to the verification questions;
• a corrected final answer.

If those outputs are missing, the workflow is not really running Chain-of-Verification. It is only asking for a more careful answer.

Use Chain-of-Verification when all of these are true:

1. The assistant has produced, or will produce, a draft answer.
2. The draft contains factual claims.
3. The final answer needs a higher reliability standard than a first response.

Good use cases include:

• checking a research summary before reuse;
• checking a technical explanation before sending it to a developer, client, or team;
• checking a document summary before it becomes part of a report;
• checking a public article, post, or recommendation before publication;
• checking an answer that may have accepted an assumption too quickly;
• checking output where a mistake would create rework, confusion, reputational risk, or a bad decision.

Do not use it as a replacement for a stronger review process. If the task requires source discovery, full fact-checking, citation audit, legal review, medical review, security review, compliance review, or expert judgment, use the appropriate stronger workflow.

A Chain-of-Verification pass needs clear inputs. Without them, the assistant may appear to verify the answer while only restating it.

Original task

Keep the original request visible. The verification pass should check whether the draft answers the actual task, not whether the draft sounds complete in isolation.

Draft answer

The method needs a draft answer to inspect. The draft can be newly generated or copied from a previous response. It must be clearly separated from the final answer.

Source material when support is required

If the answer depends on documents, code, logs, current facts, product behavior, policy text, research papers, or other external material, provide that material before the verification pass.

If the required source material is missing, the assistant should not preserve a claim because it sounds plausible. It should mark the claim as not verified, remove it, or state the uncertainty in the final answer.

Claim selection standard

Decide which claims matter enough to check. A practical rule is to check claims that affect correctness, trust, publication, decisions, safety, or downstream reuse.

Start by asking for a normal answer, but label it as provisional. The draft should be complete enough to inspect. It should not be a loose outline unless the final answer is also supposed to be an outline.

Use wording like this:

Write a draft answer to the task. Treat it as provisional. Do not present it as final. The draft will be checked before the final answer is written.

The draft step is useful because it reveals what the assistant would normally claim. Those claims become the material for verification.

This step fails when:

• the assistant skips the draft and gives a final answer immediately;
• the draft is too vague to extract claims from;
• the draft hides uncertainty instead of making claims visible;
• the draft and the final answer are blended together.

Pass condition: there is a clearly labeled draft answer, and it has not been accepted as final.

After the draft exists, identify the factual claims that need checking. A claim is checkable when it can be tested against source material, approved references, tool results, or permitted context.

Use wording like this:

Extract the important factual claims from the draft. Focus on claims that affect correctness, trust, decisions, publication, or reuse. Ignore style-only wording unless it changes the meaning.

Prioritize claims such as:

• names of people, products, companies, tools, papers, policies, or features;
• dates, deadlines, release timing, or sequence of events;
• numbers, percentages, counts, rankings, or measurements;
• definitions or distinctions between concepts;
• claims about what a source says;
• claims about what a document, contract, policy, or codebase contains;
• technical behavior, platform behavior, or product capabilities;
• research findings or conclusions;
• recommendations that depend on evidence;
• cause-and-effect statements.

Write each claim as a short statement that can be tested.

Weak extraction:

The answer discusses the policy.

Better extraction:

The draft claims that the policy requires approval before storing client documents.

Weak extraction:

The answer explains the method.

Better extraction:

The draft claims that Chain-of-Verification answers verification questions independently before writing the final response.

Pass condition: each selected claim is specific enough that a verification question can test it.

Turn each important claim into a direct verification question. The question should test the claim, not the overall quality of the answer.

Use wording like this:

For each important claim, write one verification question that can determine whether the claim is supported, incorrect, incomplete, unclear, or not verified. Keep each question narrow enough to answer independently.

Weak verification questions are broad or self-confirming:

• Is this answer correct?
• Does this sound right?
• Can you verify the whole response?
• Is the draft good enough?

Better verification questions are narrow and evidence-seeking:

• Does the supplied document state that the deadline is June 30?
• Does the cited paper describe the method as draft, verification questions, independent answers, and final verified response?
• Does the codebase contain the function name used in the draft?
• Does the product documentation support the feature behavior stated in the draft?
• Does the policy text require approval before this action?

Use one question per claim when possible. If multiple claims depend on the same source sentence or source section, they can be grouped, but the group should stay small.

Pass condition: every important claim has a verification question, and every verification question maps back to a claim in the draft.

This is the most important step. The draft is the object being checked. It cannot be used as proof that its own claims are true.

Use wording like this:

Answer each verification question independently. Use only the supplied source material, approved references, tool results, or explicitly permitted context. Do not use the draft as evidence for itself. If the required support is missing, mark the claim as not verified.

For source-based work, each verification answer should identify the support it used. Depending on the task, that support may be a document section, file name, line range, URL, source title, code reference, transcript excerpt, policy clause, or tool result.

For tasks where explicit sources are not available and general model knowledge is permitted, the verification answer should still avoid stronger claims than the context supports.

A useful verification answer is short and structured:

Claim:
The draft claims that the policy requires approval before storing client documents.

Verification question:
Does the supplied policy say approval is required before storing client documents?

Verification answer:
Not verified. The supplied policy text discusses storage rules, but it does not mention approval.

Status:
Not verified.

Required change:
Remove the approval claim or state that the supplied policy does not specify this requirement.

Pass condition: the verification answers are produced from the allowed material or permitted context, not from the draft wording.

After the verification questions are answered, classify each checked claim. The status determines how the final answer should change.

Use these statuses:

• Supported — the available material supports the claim.
• Incorrect — the available material contradicts the claim.
• Incomplete — the claim is partly supported but needs qualification or missing context.
• Unclear — the claim cannot be judged cleanly because the wording or material is ambiguous.
• Not verified — the required support is not available.

The status should drive the revision:

• Supported claims may stay.
• Incorrect claims must be corrected or removed.
• Incomplete claims must be qualified.
• Unclear claims must be rewritten or narrowed.
• Not verified claims must be removed, marked as not verified, or replaced with a weaker supported statement.

Use a compact table when there are multiple claims:

| Draft claim | Verification question | Result | Required change |
|---|---|---|---|
| [claim] | [question] | Supported / Incorrect / Incomplete / Unclear / Not verified | Keep / Correct / Qualify / Remove |

Pass condition: every important checked claim has a status and a required change before the final answer is written.

The final answer should be written from the verification results, not from the unchecked draft.

Use wording like this:

Write the final answer using the verification results. Keep supported claims, correct incorrect claims, qualify incomplete or unclear claims, and remove or mark claims that are not verified. Do not repeat any claim that failed the verification pass unless the final answer explicitly marks it as unsupported or not verified.

The final answer should be direct and usable. It does not need to show the full verification trail unless the task requires visible review evidence.

For normal use:

Final answer:
[Corrected answer]

For professional review:

Final answer:
[Corrected answer]

Verification summary:
- Checked [number] important claims.
- Corrected [number] claims.
- Removed or marked [number] unsupported claims.
- Remaining uncertainty: [short note, if relevant].

For audit or high-risk reuse:

Draft answer:
[original draft]

Claim list:
[claims selected for checking]

Verification questions:
[questions]

Verification results:
[answers and statuses]

Final answer:
[corrected final answer]

Pass condition: the final answer no longer repeats unsupported or failed claims from the draft.

This example shows the full sequence in a source-based task.

Original task

Summarize whether the supplied policy requires approval before storing client documents.

Draft answer

The policy requires approval before storing client documents. It also says documents must be stored in the approved workspace and should not be copied to personal drives.

Extracted claims

• The policy requires approval before storing client documents.
• The policy requires documents to be stored in the approved workspace.
• The policy says documents should not be copied to personal drives.

Verification questions

• Does the supplied policy say approval is required before storing client documents?
• Does the supplied policy require storage in the approved workspace?
• Does the supplied policy prohibit copying documents to personal drives?

Verification results

| Draft claim | Result | Required change |
|---|---|---|
| Approval is required before storing client documents. | Not verified. The supplied policy does not mention approval. | Remove the claim. |
| Documents must be stored in the approved workspace. | Supported. | Keep the claim. |
| Documents should not be copied to personal drives. | Supported. | Keep the claim. |

Final verified answer

The supplied policy says client documents must be stored in the approved workspace and should not be copied to personal drives. I did not find support in the supplied policy for the claim that approval is required before storing client documents.

The final answer is stronger than the draft because it removes the unsupported approval requirement instead of presenting it as confirmed.

Use these checks before treating the Chain-of-Verification pass as complete.

Draft separation

The first answer is clearly marked as draft material and is not presented as final output.

Claim coverage

The verification pass covers the important claims that affect correctness, trust, decisions, publication, or reuse.

Question specificity

The verification questions test concrete claims. They are not broad requests for the assistant to confirm its own answer.

Independent checking

The verification answers use the supplied source material, approved references, tool results, or explicitly permitted context. They do not use the draft as proof.

Revision behavior

The final answer changes when the check finds a problem. Unsupported, incorrect, incomplete, or unclear claims are corrected, removed, or qualified.

Missing support

Missing evidence remains visible. A claim that cannot be checked is marked as not verified, removed, or rewritten with uncertainty.

If the final answer is almost identical to the unchecked draft, the verification pass was probably too shallow.

The assistant asks only one broad question

“Is this answer correct?” is too broad. It does not force claim-level checking.

The assistant uses the draft as evidence

The draft is what the workflow is checking. It cannot prove itself.

The assistant checks style instead of factual claims

Improving wording is not the same as verifying whether the answer is supported.

The assistant keeps unsupported claims because they sound plausible

A plausible claim still needs support when the task requires support.

The assistant hides uncertainty in the final answer

If a claim is not verified, the final answer should not present it as known.

The method is used when a stronger review is required

Chain-of-Verification is a prompt-based checking method. It does not replace expert review, source audit, legal review, medical review, security review, or compliance review when those are required.

Use this instruction when you want to add Chain-of-Verification to a prompt, assistant instruction, or review workflow.

Before giving the final answer, run a Chain-of-Verification check.

1. Write a draft answer and keep it provisional.
2. Extract the important factual claims from the draft.
3. Turn those claims into specific verification questions.
4. Answer the verification questions independently using only the supplied source material, approved references, tool results, or explicitly permitted context.
5. Classify each checked claim as supported, incorrect, incomplete, unclear, or not verified.
6. Rewrite the final answer from the verification results.

Rules:
- Do not treat the draft as evidence.
- Do not ask only whether the whole answer is correct.
- Do not keep unsupported claims because they sound plausible.
- If support is missing, mark the claim as not verified, qualify it, or remove it.
- The final answer must not repeat claims that failed the verification pass.

Output:
- Provide the corrected final answer.
- Include a short verification summary when the task requires visible review evidence.
- Preserve the full claim list and verification trail when the output is high-risk, reviewed, published, or reused.

This instruction is intentionally explicit. The assistant should know what to draft, what to check, how to check it, and how the check must change the final answer.