What you will be able to do
- Explain why Claude can produce confident answers that are wrong, and when rephrasing will not help
- Apply grounding techniques: permission to say 'I don't know', quote extraction, citations and restricting Claude to the documents you provide
- Stabilise inconsistent output with explicit formats and examples
- Check whether a fix worked using specific success criteria and repeated runs
1.Why a fluent answer can still be wrong
Incorrect output is a different failure from vague or brief output, and it needs a different fix. Anthropic's help article on incorrect or misleading responses says Claude can occasionally produce such answers because it is trying to be helpful. This is called hallucination, and the article describes it as a byproduct of current limitations of frontier generative AI models. It gives two forms. Claude may not have been trained on up-to-date information in some areas, so questions about recent events can confuse it. And it can produce authoritative-sounding quotes that are not grounded in fact.
This matters for diagnosis because confidence is not evidence. The article says plainly that Claude can write things that look correct but are very mistaken. If the answer depends on information Claude was not trained on, rewording the question will not create that knowledge. Rephrasing addresses an unclear prompt, and the prompt was not the problem. The article's guidance is to avoid treating Claude as a single source of truth and to scrutinise high-stakes advice.
The same article makes a related point about web search: the quality of Claude's answer depends on the sources it draws from, so you should check the original pages. Its general lesson applies whenever an answer rests on material Claude was given. If that material is wrong or out of date, better wording will not correct it. The source has to be corrected, or checked, instead.
A team building a classification prompt has zero examples in the prompt, only a description of the categories. Outputs are inconsistent on edge cases, such as tickets that could reasonably fit two categories. What is the most effective next step to reduce this inconsistency?
Correct answer: C — Add three to five diverse, relevant examples, including edge cases, wrapped in example tags so Claude can mirror the intended pattern.
- A. Incorrect. An arbitrary tie-breaking rule like alphabetical order does not reflect the actual decision criteria and can produce systematically wrong categorizations for ambiguous tickets.
- B. Incorrect. Adding more categories increases the classification surface and is likely to increase ambiguity rather than resolve the existing edge cases.
- C. Correct. Anthropic's guidance is that a few well-crafted, diverse examples (few-shot/multishot prompting) are one of the most reliable ways to steer output consistency, and including edge cases in examples helps Claude generalize to ambiguous inputs.
- D. Incorrect. Removing the category descriptions removes information Claude needs and would likely worsen consistency rather than improve it.
Sources1
2.Grounding techniques that reduce hallucination
Anthropic's guide to reducing hallucinations lists techniques that constrain what Claude is allowed to claim. They share one principle: tie each claim to something checkable, and make it acceptable for Claude to say it does not know.
| Technique | What you do | What it catches |
|---|---|---|
| Allow Claude to say "I don't know" | Explicitly give permission to admit uncertainty | Invented answers where the information is missing |
| Use direct quotes for factual grounding | For long documents, have Claude extract word-for-word quotes before doing the task | Claims that drift from the actual text |
| Verify with citations | Require a supporting quote for each claim and retract any claim without one | Unsupported claims in the final output |
| External knowledge restriction | Instruct Claude to use only the provided documents, not its general knowledge | General knowledge leaking into a document-based answer |
| Chain-of-thought verification | Ask for step-by-step reasoning before the final answer | Faulty logic or assumptions |
These techniques combine well. The guide's press-release example restricts Claude to the supplied documents, then has it check every claim against a quote and visibly mark anything it removed, so a reviewer can see where support was missing.
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.The gap becomes visible. A reviewer can see exactly where the documents failed to support a claim and decide whether to find a source or drop the point. A silent deletion would hide the fact that anything was unsupported.
Sources2
3.When the output is inconsistent rather than wrong
Some poor outputs are not wrong. They just vary in structure or tone from one run to the next, which breaks anything downstream that expects a fixed shape. Anthropic's consistency guide recommends precisely defining the output format with JSON, XML or a custom template, so Claude follows every formatting element you require.
Examples are the other lever. The prompting guide says a few well-crafted examples improve accuracy and consistency. They should mirror your real use case, vary enough to cover edge cases without teaching unintended patterns, and be wrapped in <example> tags so Claude can tell them apart from instructions. XML tags in general help Claude separate instructions, context, examples and inputs when a prompt mixes them.
4.Checking that the fix actually worked
A fix is only a fix if the output measurably improves. The hallucination guide gives two ways to check. Best-of-N verification runs the same prompt several times and compares the results: inconsistencies suggest hallucination. Iterative refinement feeds Claude's output back in and asks it to verify or expand on what it said. The help article adds a feedback route: the thumbs-down button reports an unhelpful response to Anthropic.
For prompts you rely on repeatedly, Anthropic's evaluation guide recommends starting from success criteria that are specific and measurable. Instead of "good performance", specify the result you mean, such as "accurate sentiment classification". Without a target like that, you cannot tell whether a reworded prompt fixed the problem or just changed it.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.A confident, detailed answer with authoritative-sounding quotes is probably accurate.Why is that wrong?
Hallucinated output can look correct and cite convincing quotes that are not grounded in fact, so high-stakes answers need checking.
Covered in Why a fluent answer can still be wrong
2.If Claude gets a recent event wrong, rephrasing the question more precisely will produce the right answer.Why is that wrong?
The failure is a knowledge gap, not a wording problem. Claude may not have been trained on current information, so the fix is to provide or verify the facts.
Covered in Why a fluent answer can still be wrong
3.Telling Claude it may say 'I don't know' makes its answers less useful, so you should leave that permission out.Why is that wrong?
Explicit permission to admit uncertainty is one of Anthropic's listed hallucination-reduction techniques, and the guide calls it simple and drastically effective.
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://support.claude.com/en/articles/8525154-claude-is-providing-incorrect-or-misleading-responses-what-s-going-onOfficial docs
“Claude might not have been trained on the most-up-to-date information and may get confused when prompted about current events.”
↩︎ Why a fluent answer can still be wrong“Users should not rely on Claude as a singular source of truth and should carefully scrutinize any high-stakes advice given by Claude.”
↩︎ Why a fluent answer can still be wrong“depends on the underlying sources it references, so checking original content helps you identify any information that might be misinterpreted”
↩︎ Why a fluent answer can still be wrong“You can use the thumbs down button to let us know if a particular response was unhelpful”
↩︎ Checking that the fix actually worked“Claude can write things that might look correct but are very mistaken.”
↩︎ Exam trap 1“Claude might not have been trained on the most-up-to-date information and may get confused when prompted about current events.”
↩︎ Exam trap 2 - 2.https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinationsOfficial docs
“ask Claude to extract word-for-word quotes first before performing its task. This grounds its responses in the actual text, reducing hallucinations.”
↩︎ Grounding techniques that reduce hallucination“Explicitly instruct Claude to only use information from provided documents and not its general knowledge.”
↩︎ Grounding techniques that reduce hallucination“Ask Claude to explain its reasoning step-by-step before giving a final answer. This can reveal faulty logic or assumptions.”
↩︎ Grounding techniques that reduce hallucination“Run Claude through the same prompt multiple times and compare the outputs. Inconsistencies across outputs could indicate hallucinations.”
↩︎ Checking that the fix actually worked“Explicitly give Claude permission to admit uncertainty. This simple technique can drastically reduce false information.”
↩︎ Exam trap 3 - 3.https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/increase-consistencyOfficial docs
“Precisely define your desired output format using JSON, XML, or custom templates so that Claude follows every output formatting element you require.”
↩︎ When the output is inconsistent rather than wrong - 4.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“A few well-crafted examples (known as few-shot or multishot prompting) improve accuracy and consistency.”
↩︎ When the output is inconsistent rather than wrong - 5.
“Clearly define what you want to achieve. Instead of "good performance," specify "accurate sentiment classification."”
↩︎ Checking that the fix actually worked