What you will be able to do
- Set up a task so that Claude's output can be checked against the documents it was meant to use
- Run a claim-by-claim check that removes any claim without a supporting quote, and leaves a record someone else can audit
- Use repeated runs, step-by-step reasoning and follow-up prompts to bring inconsistencies to the surface
1.Make the output checkable before you check it
Validation is much easier when the output was built to be checked. Anthropic's developer guide on reducing hallucinations describes several ways to set this up in the prompt. They are general good practice from prompt-engineering material, not part of the help centre's list of recommendations.
Allow uncertainty. Explicitly give Claude permission to say it doesn't know. The guide's example tells Claude to reply that it doesn't have enough information to assess something confidently, rather than guess. An honest gap is easy to check. A confident invented answer is not.
Restrict the knowledge base. Tell Claude to use only the information in the documents you provide, and not its general knowledge. Then every claim has a known place it should have come from, and you know exactly where to look.
Extract quotes first. For long documents (the guide says over 20,000 tokens), ask Claude to pull out word-for-word quotes before doing the task. Then have it base its analysis only on those quotes, referring to them by number. Each conclusion can then be traced to a passage you can read.
Checkpoint 1 of 4· Check yourself
You want to be able to trace every claim in a Claude-drafted compliance summary back to the policy document you supplied. Which instruction does most to make that possible?
Restricting Claude to the supplied document gives every claim a single known source to check against. Tone, length and disclaimers do not make any claim traceable.
“Explicitly instruct Claude to only use information from provided documents and not its general knowledge.”Source: platform.claude.com
Sources1
2.Verify each claim, and retract what you can't support
The central technique is to verify with citations. The aim is a response that can be audited, because every claim cites the quote and source that supports it. The developer guide goes one step further. After Claude drafts a response, have it go back through each claim and find a supporting quote. If no quote can be found, the claim must be retracted.
The guide's worked example is a press release drafted only from the product briefs and market reports provided. Claude reviews each claim and looks for a direct quote that supports it. Any claim without one is removed, and the gap is marked with empty [] brackets. The brackets matter. They show a reviewer what was taken out, so the gap stays visible instead of being quietly smoothed over.
This also explains why adding citations after the fact is risky. If a draft was written from general knowledge and you then ask Claude to add citations, the claims came first and the sources second. Some claims may have no real support, and a citation attached to them can give false confidence. The safe method is the reverse: test every claim against a source and drop the ones that fail.
A claim-by-claim record also makes your fact-check reviewable later. A note saying "checked" can't be audited. A list of each claim, its source and the exact supporting quote can be.
Checkpoint 2 of 4· Put it in order
Put the steps of the developer guide's press-release verification in order
- 1.Review each claim in the draft
- 2.Find a direct quote from the documents that supports the claim
- 3.Remove any claim without a supporting quote and mark the gap with empty [] brackets
- 4.Draft the press release using only the provided product briefs and market reports
Claude drafts from the supplied documents first. Then it checks each claim for a supporting quote, and removes and marks any claim that has none.
“remove that claim from the press release and mark where it was removed with empty [] brackets”Source: platform.claude.com
Checkpoint 3 of 4· Exam question
A policy officer has one hour before a board meeting and a Claude-written briefing containing roughly twenty factual claims. She cannot check them all. What should she do first?
Correct answer: B — Identify the few claims the board's decision actually rests on and verify those ones first.
- A. Incorrect. Order of appearance has no relationship to importance, so the claims that matter most may sit in the part she never reaches.
- B. Correct. When verification time is scarce it has to be spent where an error would change the decision, so triage by consequence comes before any checking begins.
- C. Incorrect. Claude has no independent view of which of its own statements are unsupported, so this ranking is generated the same way the claims were.
- D. Incorrect. Maximising the count optimises for the wrong thing; easy claims are often the low-stakes ones.
Sources1
3.Cross-checking: repeated runs, reasoning and follow-ups
Some errors don't show up when you check claims against sources. The logic may be wrong, or an answer may be unstable. The same developer guide lists three cross-checks for these cases. Each one exposes a different kind of problem.
| Technique | What you do | What it reveals |
|---|---|---|
| Chain-of-thought verification | Ask Claude to explain its reasoning step by step before giving a final answer | Faulty logic or assumptions |
| Best-of-N verification | Run the same prompt several times and compare the outputs | Inconsistencies across runs, which may indicate hallucination |
| Iterative refinement | Feed Claude's output into follow-up prompts asking it to verify or expand on its statements | Inconsistencies that can then be corrected |
Best-of-N works because of how Claude generates text. Answers are sampled, so asking for the same specific facts in a fresh conversation can produce different details. A detail that keeps changing between runs is a warning sign. A detail that stays the same is not proven true, though: consistency is a reason to keep checking, not a substitute for a source.
All of these techniques are versions of what the Anthropic Academy calls the generator-verifier pattern, and they are tools of Discernment: evaluating what AI gives back before acting on it. The Academy also offers a useful test of yourself. Practise on a topic where you are the expert, because you can only score fabrications you are able to recognise. Then ask whether you would have caught the same errors in a field you know less well. If not, those are the outputs that most need an outside source.
Checkpoint 4 of 4· Match them up
Match each cross-checking technique to the problem it is designed to expose
Tap a term, then the definition that fits it.
Step-by-step reasoning exposes faulty logic. Comparing repeated runs exposes unstable details. Follow-up prompts test earlier statements and catch inconsistencies.
“Run Claude through the same prompt multiple times and compare the outputs. Inconsistencies across outputs could indicate hallucinations.”Source: platform.claude.com
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Asking Claude to add citations to a draft it already wrote from general knowledge validates the draft.Why is that wrong?
Validation means testing each claim against a source and retracting any claim that has no supporting quote. Attaching sources to existing claims skips that test.
Covered in Verify each claim, and retract what you can't support
2.A fact-check is auditable as long as the reviewer notes that the output was checked.Why is that wrong?
An auditable check ties each claim to the quote and source that support it, so someone else can repeat the check later.
Covered in Verify each claim, and retract what you can't support
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinationsOfficial docs
“Explicitly give Claude permission to admit uncertainty. This simple technique can drastically reduce false information.”
↩︎ Make the output checkable before you check it“ask Claude to extract word-for-word quotes first before performing its task”
↩︎ Make the output checkable before you check it“You can also have Claude verify each claim by finding a supporting quote after it generates a response.”
↩︎ Verify each claim, and retract what you can't support“Make Claude's response auditable by having it cite quotes and sources for each of its claims.”
↩︎ Verify each claim, and retract what you can't support“Ask Claude to explain its reasoning step-by-step before giving a final answer. This can reveal faulty logic or assumptions.”
↩︎ Cross-checking: repeated runs, reasoning and follow-ups“Use Claude's outputs as inputs for follow-up prompts, asking it to verify or expand on previous statements.”
↩︎ Cross-checking: repeated runs, reasoning and follow-ups“If it can't find a quote, it must retract the claim.”
↩︎ Exam trap 1“Make Claude's response auditable by having it cite quotes and sources for each of its claims.”
↩︎ Exam trap 2“Explicitly instruct Claude to only use information from provided documents and not its general knowledge.”
↩︎ Checkpoint“remove that claim from the press release and mark where it was removed with empty [] brackets”
↩︎ Checkpoint“Run Claude through the same prompt multiple times and compare the outputs. Inconsistencies across outputs could indicate hallucinations.”
↩︎ Checkpoint - 2.https://academy.claude.com/courses/ai-capabilities-and-limitations/next-token-predictionOfficial docs
“Product features like citations, uncertainty signaling, constrained generation, and generator-verifier loops exist specifically to push this limitation further out.”
↩︎ Cross-checking: repeated runs, reasoning and follow-ups“Would you have caught fabrications in a domain you didn't know well?”
↩︎ Cross-checking: repeated runs, reasoning and follow-ups