Make AI show its sources

What it is

This is a set of lines you add to a request so Claude shows where its answer comes from. Anthropic's guidance on reducing hallucinations (answers that sound right but are wrong) describes three habits: let Claude say "I don't know", ground the answer in direct quotes, and check each claim against a quote. You do not install anything. You only change what you type.

Who it is for

It helps anyone who asks AI to work from a document: a manager summarising a report, an HR lead checking a policy, a marketer drafting a press release from product briefs, or an operations person reading a contract. It is most useful when someone else will read the result and you need to point to the source.

What you need

• Access to Claude.
• The text you want Claude to work from, pasted into your request.
• Anthropic's documentation does not name a plan, admin right or other app for this technique. It only describes how to write the prompt.

Set it up

Add these lines to your request, in this order.

1. Write your task first. Then paste your document between simple tags, such as <documents> and </documents>, so Claude can tell your text from your instructions. Anthropic's examples use tags in this way.
2. Tell Claude to use only your documents. Anthropic calls this "external knowledge restriction": Claude should rely on the text you provide and not on its general knowledge.
3. Give Claude permission to not know. Anthropic's example wording: "If you're unsure about any aspect or if the report lacks necessary information, say "I don't have enough information to confidently assess this."" Anthropic says this can drastically reduce false information.
4. For long documents, ask for quotes first. Anthropic recommends this for very long documents and gives more than 20k tokens as the size. Ask for exact, word-for-word quotes before the task, and add: "If you can't find relevant quotes, state "No relevant quotes found.""
5. Ask Claude to base its analysis only on those quotes and to refer to them by number.
6. After a draft, ask Claude to check itself. For each claim, it should find a direct quote that supports it. If it cannot find one, it should remove the claim and mark the spot with empty [] brackets.
7. If an answer still looks wrong, ask a follow-up that asks Claude to verify or expand on its earlier statements. Anthropic calls this iterative refinement and says it can catch inconsistencies.
8. Optional: send the same prompt more than once and compare the answers. Anthropic says that differences between runs could point to made-up details.

A first thing to try

Take a product brief or a report you already have and ask for a short summary for your team. Paste this, and replace the bracketed part with your text:

Write a short summary for my team using only information from these documents.

<documents>
[paste your documents here]
</documents>

1. First, extract exact word-for-word quotes that are most relevant to the summary, and number them. If you can't find relevant quotes, state "No relevant quotes found."

2. Write the summary using only those quotes, and refer to them by number.

3. If any point is not covered by the documents, say "I don't have enough information to confidently assess this."

4. Finally, review each claim in your summary. If you can't find a direct quote that supports a claim, remove it and mark where it was removed with empty [] brackets.

Then read the numbered quotes next to the summary and check that each one really says what the summary claims.

Limits and privacy

• Anthropic says these techniques significantly reduce made-up answers but do not eliminate them. It says to always validate critical information, especially for high-stakes decisions.
• Anthropic describes the quote-first step for tasks with long documents. Its documentation does not say what the step adds for short ones.
• Asking for quotes in a prompt is not the same as Anthropic's separate built-in citations feature. That feature is set up by developers through Anthropic's API (the way software connects to Claude). Anthropic says it is more reliable in some ways and, in its own evaluations, significantly more likely to cite the most relevant quotes than prompt-only approaches. Anthropic also says it cannot cite images yet.
• Anthropic's documentation on this technique does not describe how your pasted text is stored or used. Check Anthropic's privacy policy and your company's rules before pasting sensitive documents.

Ask Claude to show its quotes, read them yourself, and treat any claim without a quote as unconfirmed.

Where these steps come from

Steps from Anthropic's documentation, checked 13 Nov 2026.

• Reduce hallucinations
• Citations
• Prompting best practices

This guide was written by AI from the pages above. Menus and plans change, so open them if a step looks different.