ClaudeSummarizationLevel 2: Context

How Do You Get Claude to Cite Its Sources?

It depends on where you're working. Web search and Research cite sources for you. The Word and Google Docs add-ins cite within the document. For a file uploaded to a normal chat, ask for it: have Claude quote the exact words and give a page number for each claim, and label anything it inferred.

Prompt to copy 
<document>
[attach or paste the source]
</document>

First, find the word-for-word quotes in this document most relevant to [your question], each with its page or section. Put them in <quotes> tags, numbered. If you can't find any, say "No relevant quotes found."

Then, using only those quotes, list the key claims. For each claim:
- The claim, in one line
- The quote number that supports it
- [stated] if the document says it directly, or [inferred] if it's your interpretation

Don't use outside knowledge. Finally, check each claim against its quote, and remove any claim the quote doesn't fully support. Tell me how many you removed.
  • Web searchcites on its own
  • Researchlinks each claim to a source
  • Claude for Wordclickable section citations
  • A file in a chatask for quotes and pages
From Anthropic's help pages. They don't promise page citations for files in a regular chat, so that's where you ask.
  1. 1Word-for-word quotes first
  2. 2Claims built on the quotes
  3. 3Unsupported claims removed
  4. 4You spot-check the quotes
Based on Anthropic's guide to reducing hallucinations. It says these methods reduce errors and don't eliminate them.

Try it now

  1. 1Click +, choose Add files or photos, attach a report you know well, and paste the prompt.
  2. 2Pick three quotes at random and find them in the original, using Ctrl+F or Cmd+F.
  3. 3Read all claims marked [inferred], since those are where Claude has gone beyond the text.
  4. 4You'll know it worked when the quotes you checked appear word for word, on the page Claude gave.

Use it when

Reports you'll quote from, legal or medical reading, literature reviews, due diligence, and any summary that someone else will rely on. Use it whenever a wrong claim would cost you more than the two minutes it takes to check a quote.

Skip it when

Skimming an article to decide whether it's worth reading. A plain summary is fine, and quotes make it slower. Also note the method's limit: scanned PDFs with no extractable text are poor candidates for page-level checking, because Anthropic's API docs say those PDFs aren't citable. Get a text version if you can.

Why it works

Anthropic's guide to reducing hallucinations recommends this. For long documents, meaning over 20,000 tokens, it says to ask for word-for-word quotes first, which grounds the task in the actual text. It also suggests having Claude find a supporting quote for each claim after drafting and retract any claim without one. Anthropic notes these methods reduce errors but don't eliminate them.

The help pages read for this card don't promise clickable page citations for files uploaded to a regular chat, which is why the prompt asks for quotes. Elsewhere, citations are built in: Anthropic says each web search response includes them, Research links claims to sources, and Claude for Word gives clickable section citations that jump to the passage.

Example

Say you have a 60-page market report and need to know what it says about churn. Claude returns five numbered quotes with page numbers and four claims built on them. One claim, "churn is rising in small accounts", is marked [inferred] since the report only gives two separate figures for different years. You search for one quote, find it on page 34 as stated, then go look at that inferred comparison yourself before repeating it to your team.

Going further

On the API, you don't need the quote prompt. Add "citations": {"enabled": true} to a document block. Anthropic's docs say PDFs and plain text are split into sentences, and PDF citations return page_location with the cited_text, which doesn't count toward output tokens. Because the API extracts that text directly, Anthropic says the pointers are guaranteed to be valid. Two traps: citations with structured outputs (output_config.format) return a 400 error, and images in PDFs can't be cited.

For web search through the API, citations are always on, and cited_text holds up to 150 characters.

Common mistakes

  • Assuming a page number is right because it's there. In a chat, Claude writes it as text, so spot-check a few.
  • Asking for a summary first and citations afterward. Quotes first is what grounds the task, per Anthropic's guide.
  • Trusting a web citation without opening it. A linked source only tells you where to check, and the page may not say quite what the sentence implies.
  • Using a scanned PDF and expecting exact quotes. Without extractable text, what you get depends on Claude reading page images, so verify carefully.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Mistake How Do You Check a Claude Summary Against the Source? Technique How Do You Summarize a Very Long PDF With Claude? Best practice How Do You Get Claude to Put the Conclusion First? Mistake Which Decisions Shouldn't You Leave to Claude?

Beyond the course

Glossary Hallucination Pricing guide Claude pricing, plans and models