// how to
How to Fact-Check Claude's Answers and Avoid Made-Up Details
Ask Claude to answer only from documents you provide, or turn on web search so answers come with citations. Ask it to flag which claims it’s unsure of, and verify specific facts (names, numbers, dates, quotes, rules) at the original source before you rely on them.
The short version
Three habits make Claude far more reliable: give it the source, let it say “I don’t know”, and verify specifics yourself. Anthropic’s own advice on reducing made-up answers includes allowing the model to say it doesn’t know and grounding answers in quotes from the documents.
Step by step
- Give it the source. Attach the document, policy or data and say “answer only from this”.
- Allow “I don’t know”. Tell it to say so rather than guess.
- Ask for quotes from the document that support each answer.
- Use web search for anything public or recent, so answers carry citations. See research with Claude.
- Ask what to double-check: Claude is good at flagging its own weak spots.
- Verify specifics at the source: the law, the company’s own page, the official statistic.
Prompts to copy
Answer from my document only:
Answer using only the attached document. Quote the sentence that supports each answer.
If the document doesn't answer something, say "not in the document" instead of guessing.
Flag uncertainty:
Mark each claim in your answer as [FROM SOURCE], [GENERAL KNOWLEDGE] or [UNSURE].
Then list the [UNSURE] ones and how I could check them.
Check its own answer:
Review your previous answer for factual errors and unsupported claims. Be critical. What would you change?
What to check first
- Quotes and citations: do they exist and say this?
- Numbers: prices, statistics, dates.
- Rules and laws: they vary by place and change.
- Recent events: anything recent needs a visible source.
- Names: people, products, companies.
Mistakes to avoid
- Asking open questions about facts you could provide. Give it the document.
- Not allowing “I don’t know”. Without permission, models tend to answer anyway.
- Treating fluent as correct. Confidence isn’t evidence.
When it really matters
For health, legal, tax and money decisions, use Claude to understand and prepare questions, then confirm with a professional or the official source.
Related: how to use Claude.
Frequently asked questions
Does Claude make things up?
Like every AI assistant, it can produce confident answers that are wrong, especially for specific facts, recent events and niche topics. Asking it to say when it's unsure, and to answer from sources you provide, reduces this a lot.
Are Claude's citations reliable?
They make checking much easier, but open the important ones: a source may be outdated or not say exactly what the answer suggests.
Free download
Get the AI Coding Starter Kit
A PDF you can use with any AI coding assistant: a project-rules template, 25 tested prompt patterns, a team adoption checklist and an assistant comparison sheet. The editable Markdown sources are included.
- Works with Copilot, Claude Code, Codex, Cursor and more
- Double opt-in, one-click unsubscribe
- No trackers, never shared