Ground the Answer

This is the technique that matters most for research, legal, compliance, analysis, and any work where a confident-but-wrong answer is worse than no answer. It directly targets hallucination — the model’s habit of producing plausible, fluent claims that aren’t actually supported by the source.

The move:

Answer the question using only the document below.

First, in <quotes>, copy the exact sentences from the document that
support your answer. If the document does not contain the answer, write
"NOT FOUND" in <quotes> and stop.

Then, in <answer>, answer the question using only those quotes.

<document>
{the source}
</document>

Question: {the question}

You’ve forced a sequence: find the evidence, then answer from the evidence. The model can’t fluently invent a conclusion when it has to point at the exact sentences that back it up first.

Why quoting first works

Left to its own devices, the model generates the most plausible-sounding answer and may fill gaps with invention — it’s optimizing for a fluent response, not a sourced one. Requiring quotes before the answer changes the order of operations:

1. It anchors the response in the actual text instead of the model’s general impression of it.
2. It makes gaps visible — if there are no supporting quotes, that’s now obvious to both of you, instead of being papered over with a confident paragraph.
3. It gives you something to verify. You can scan the quotes against the source in seconds, instead of re-reading the whole document to check the answer.

This is the same instinct as the reason-before-answering technique from the thinking lesson, pointed at a different failure: there, you wanted logical steps before a conclusion; here, you want textual evidence before a claim.

Give it permission to come up empty

Half of grounding is the quoting. The other half is explicitly allowing “I don’t know.” Models hallucinate partly because nothing in the prompt told them that “not found” is an acceptable answer — so they assume they must produce something.

If the answer is not in the document, say "The document does not
address this." Do not use outside knowledge. Do not guess.

Those three sentences prevent a large share of confident fabrications. “Do not use outside knowledge” is important too: without it, the model may quietly blend the document with its training data, and you lose the guarantee that the answer came from your source.

Long documents: where to put what

Grounding gets more important as the source gets longer, because there’s more room for the model to drift. Anthropic’s long-context guidance gives a concrete structure:

• Put the long document near the top, above your question and instructions. In Anthropic’s testing, placing the query at the end — after the document — improved response quality by up to 30% on long inputs.
• Wrap the document in tags (the structure technique from the last lesson) so its boundaries are unambiguous.
• Ask for quotes first when grounding matters, so the model re-reads the relevant parts before answering instead of relying on a fuzzy impression of a 50-page input.

Putting it together for a long source:

<document>
{the long document — placed first}
</document>

Using only the document above:
1. In <quotes>, copy the exact passages relevant to the question.
   If none exist, write "NOT FOUND".
2. In <answer>, answer using only those quotes. Do not use outside
   knowledge.

Question: {the question — placed last}

Key takeaways

• Quote first, answer second. Forcing the model to extract supporting passages before it concludes is the strongest single defense against hallucination.
• Permit “not found.” Explicitly allow “the source doesn’t address this” and forbid outside knowledge — much of fabrication comes from the model assuming it must answer.
• Quotes are your audit trail. Verify the claim by checking the quotes against the source — far faster than re-reading everything.
• For long sources: document first, query last, wrapped in tags. Query-at-the-end placement alone can lift quality by up to 30%.
• This is the technique to reach for on anything where wrong-but-confident is the costly failure — research, legal, compliance, analysis.

Next: Chain Your Prompts