The short version

Long-document work combines two different problems: finding the relevant evidence and reasoning from it. Anthropic recommends asking Claude to extract relevant passages first. That creates a visible grounding stage before interpretation and helps unrelated parts of a large context carry less influence.

The pattern is documents, evidence selection, analysis, answer.

Structure the source pack

Place long documents before the final question and give each one a stable source label. XML tags or another consistent delimiter can distinguish document content, source metadata, and instructions. Put the task after the documents so Claude encounters the full evidence before deciding what matters.

Then describe the evidence categories to find. “Extract relevant quotes” is better when relevance is defined: contractual obligations, approval thresholds, stated assumptions, numerical results, or passages that contradict the proposed conclusion.

Separate evidence from interpretation

Ask Claude to preserve the source label beside every extracted passage. The analysis should refer back to that evidence and clearly distinguish what the documents state from what Claude infers.

This separation makes missing support easier to see. When the source pack does not answer a material question, the response should name the gap instead of filling it with a plausible assumption.

Example

Suppose Claude receives a lengthy agreement, policy guide, and technical standard to assess a proposed vendor arrangement. A one-step request to decide whether the arrangement is acceptable encourages retrieval and reasoning to blur together.

A stronger prompt first asks for passages concerning authorization limits, liability, data handling, service commitments, and exceptions, with each passage labeled by source. It then asks Claude to analyze the arrangement using only that selected evidence, distinguish direct statements from interpretation, and list unresolved questions.

The extracted passages are not the finished product. They form a compact evidence layer that readers can inspect before considering the conclusion.

Preserve context without copying everything

Select enough surrounding text for a passage to retain its meaning. A sentence may depend on a heading, definition, exception, or nearby qualifier. Ask for concise extracts, but include any language needed to avoid changing the source’s scope.

For very large collections, the process may run in stages: find candidate sections, select the strongest evidence, then synthesize. Keep source labels stable across stages.

Practical checklist

  • Put large source material before the question.
  • Label every document consistently.
  • Define the evidence categories to retrieve.
  • Preserve enough context for each passage.
  • Keep source statements separate from interpretation.
  • Identify contradictions and missing support.
  • Base the final answer on the selected evidence.

Try it

Take a multi-document question and split the prompt into two stages. First specify the passages Claude must find and label. Then define the decision or explanation it should produce from those passages and the evidence gaps it must disclose.