What you will be able to do
- Make a Claude draft auditable by tying each claim to a supporting quote and removing claims that have none
- Explain what citations give a reviewer, their limits, and why citations added after drafting need extra scrutiny
- Apply a validate, fix, repeat loop and plan-validate-execute to multistep or high-stakes work
- Design an independent verifier step, and explain why workflow structure matters more than model strength
1.Make every claim auditable
Fact-checking only helps if someone can follow it later. Anthropic's developer guidance calls this making the response auditable: have Claude cite quotes and sources for each claim, so any reviewer can trace a statement back to its evidence without redoing the research. The same guidance adds a second step. After Claude writes a response, it checks each claim by finding a supporting quote, and any claim without one is retracted.
Draft a press release for our new cybersecurity product, AcmeSecurity Pro, using only information from these product briefs and market reports.
<documents>
{{DOCUMENTS}}
</documents>
After drafting, review each claim in your press release. For each claim, find a direct quote from the documents that supports it. If you can't find a supporting quote for a claim, remove that claim from the press release and mark where it was removed with empty [] brackets.Two design choices matter here. The documents are named up front, so the evidence exists before any writing starts. And removed claims are marked with empty brackets instead of disappearing without trace, so a reviewer can see what could not be supported. Compare this with the reverse order: Claude drafts from general knowledge and citations are attached afterwards. In that case the claims were never grounded in the sources, and a citation found later may not actually support the sentence it is attached to. Each pairing has to be checked on its own.
A chatbot answers a user's question about a niche historical event with a detailed, confident, but entirely fabricated account, since no similar event exists in its training data. The team wants to reduce this kind of failure through prompt design alone, without adding any tools. Which instruction is most effective for reducing this type of fabrication?
Correct answer: C — Tell Claude it is acceptable, and preferred, to say it cannot verify an answer rather than guessing when unsure
- A. Encouraging a confident tone works against fact-checking goals, since it makes fabricated answers sound more credible rather than flagging uncertainty.
- B. Claude cannot selectively restrict itself to only the most recent training run, and this instruction does nothing to prevent fabrication on topics with sparse training data.
- C. Correct. Explicitly permitting Claude to express uncertainty or decline to answer when it lacks grounded information is a well-established prompting technique for reducing confident fabrication, since models often default to producing a plausible-sounding answer when not told otherwise.
- D. Limiting response length does not address the underlying cause of fabrication and may even worsen it by forcing an answer without room for caveats.
Sources1
2.What citations give you, and what they don't
On the Claude API, citations are a built-in feature. You supply documents with citations enabled, and Claude can provide detailed citations when answering questions about them, so you can track and verify the sources behind each response. Each cited text block points to a specific location in a source document, and includes the exact text it relies on.
{
"type": "char_location",
"cited_text": "The exact text being cited", // not counted toward output tokens
"document_index": 0,
"document_title": "Document Title",
"start_char_index": 0, // 0-indexed
"end_char_index": 50 // exclusive
}| Document type | Chunking | Citation format |
|---|---|---|
| Plain text | Sentence | Character indices (0-indexed) |
| Sentence | Page numbers (1-indexed) | |
| Custom content | No additional chunking | Block indices (0-indexed) |
The limits matter as much as the feature. Only text can be cited. A PDF that is a scan with no extractable text cannot be cited at all, so a missing citation may reflect the file format rather than a lack of evidence. According to Anthropic's cookbook, the feature does not return citations that point to documents or locations you did not provide. That makes it more dependable than asking for citations in the prompt, but it only proves a location exists. You still read the cited text and decide whether it actually supports the claim. In the Claude app, the AI Fluency course suggests something similar for chat users: rerun a request for specific facts in a tool with citations enabled, such as Research mode, and see whether having sources to check changes how many errors you catch.
No. The citation shows where the evidence is, and the evidence says something weaker than the claim. Checking that the claim matches the cited text word for word is the validation step. The fact that a citation exists is not.
3.Validation loops for multistep work
For longer tasks, Anthropic's Agent Skills best practices turn validation into a loop you repeat: run a validator, fix the errors, and repeat until the check passes. The validator can be a script, or a reference document such as a style guide that Claude reads and compares against. A checklist of steps keeps the checks in place, because clear steps stop Claude from skipping critical validation.
When the cost of an error is high, the guidance adds a plan-validate-execute pattern. Claude writes a structured plan first, a script validates the plan, and only then are changes applied: analyze, create plan file, validate plan, execute, verify. Validation catches problems before anything changes. The listed use cases are batch operations, destructive changes, complex validation rules and high-stakes operations. In other words, the less reversible the action, the earlier the check should come.
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.
Sources5
4.Independent verification and the human decision
Anthropic's cookbook example of fact-checking an investor update shows why structure matters. Each claim was sent to a dedicated verifier, and confirmed verdicts were then challenged by a skeptic. The subtle error it caught was a quote that had been changed from 'one of the fastest-growing' to 'the fastest-growing'. The cookbook credits the structure of the workflow for that precision, not a smarter model. Its verdicts also separated contradicted claims from unverifiable ones, which no source covered at all. Both kinds should be flagged, but they call for different fixes: correct the first, find a source for the second or drop it.
The cookbook also explains why a strict verifier is worth its occasional false alarm. An over-strict flag costs a minute of human review. An under-strict pass costs credibility with investors. That trade-off leads back to the person: the AI Fluency material for educators stresses that AI speeds work up but does not replace expertise, and that you remain the decision-maker who evaluates the output when accuracy matters.
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 makes the draft verified.Why is that wrong?
Grounding means the evidence comes first. A claim with no supporting quote should be retracted, and a citation attached afterwards has to be checked to confirm it really supports that sentence.
Covered in Make every claim auditable
2.If a sentence has a citation, the claim in it is correct.Why is that wrong?
A citation tells you where to look in a source. You still have to read the cited text and confirm it supports the exact claim, because a stronger or reworded claim can sit on top of weaker evidence.
Covered in What citations give you, and what they don't
3.The best way to improve fact-checking is to switch to a more capable model.Why is that wrong?
In Anthropic's fact-checking example, the subtle errors were caught because of how the workflow was built: a dedicated verifier followed by a skeptic.
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
“Make Claude's response auditable by having it cite quotes and sources for each of its claims.”
↩︎ Make every claim auditable“If it can't find a quote, it must retract the claim.”
↩︎ Make every claim auditable“If it can't find a quote, it must retract the claim.”
↩︎ Exam trap 1 - 2.
“Claude can provide detailed citations when answering questions about documents, helping you track and verify the sources behind each response.”
↩︎ What citations give you, and what they don't“PDFs that are scans of documents and do not contain extractable text are not citable.”
↩︎ What citations give you, and what they don't“Citations reference specific locations in source documents.”
↩︎ Exam trap 2 - 3.https://platform.claude.com/cookbook/misc-using-citationsSecondary source
“The citation feature will not return citations pointing to documents or locations that were not provided as valid sources.”
↩︎ What citations give you, and what they don't - 4.https://academy.claude.com/courses/ai-capabilities-and-limitations/next-token-predictionOfficial docs
“Re-run Probe 2 in a tool with citations enabled (like Research mode in Claude).”
↩︎ What citations give you, and what they don't - 5.
“Clear steps prevent Claude from skipping critical validation.”
↩︎ Validation loops for multistep work“When to use: Batch operations, destructive changes, complex validation rules, high-stakes operations.”
↩︎ Validation loops for multistep work“Claude performs the check by reading and comparing.”
↩︎ Validation loops for multistep work - 6.
“An over-strict flag costs a minute of human review; an under-strict pass costs credibility with investors.”
↩︎ Independent verification and the human decision“That precision comes from the structure of the workflow rather than from a smarter model.”
↩︎ Exam trap 3 - 7.
“AI accelerates, but doesn't replace expertise. You're still the decision-maker.”
↩︎ Independent verification and the human decision“Discernment isn't optional. Evaluate outputs when accuracy matters—flag claims to verify or ask for citations.”
↩︎ Independent verification and the human decision