Put findings where synthesis can use them

Learning objective

Place important findings near the start and structure long inputs to mitigate lost-in-the-middle omissions.

Core idea

An input can contain all the required evidence and still produce an answer that overlooks part of it. The exam guide calls this the “lost in the middle” effect and asks you to put key findings summaries at the beginning of aggregated inputs, with explicit section headers for detailed results.

Treat layout as another context-management decision. Do not concatenate reports in arrival order and assume inclusion alone gives every finding equal attention. Give the downstream task an early map of the findings, keep evidence under clear source labels, and make the final question explicit.

Anthropic’s long-context prompting guidance complements this pattern: place longform data near the top, structure documents and metadata, and put the query after the documents. These are prompting practices that can improve use of evidence, not a guarantee that no middle section is ever missed.

How it works

Start the aggregate with decision-relevant findings

An aggregate combines several sources or results for one downstream task. Begin that assembled data with a concise key-findings section. Include findings that change the answer, including important exceptions and unresolved disagreements, rather than only the most convenient majority conclusion.

For a support case, the early section can identify a promised refund deadline and a separately unresolved delivery issue. For research, it can identify the conclusion supported by several sources and the contrary result that limits its applicability. The purpose is to make the answer’s dependencies visible before the reader or model reaches the detail.

Keep source references with the findings so the downstream step can inspect supporting evidence. An early claim without a pointer is a second unsupported summary. See 5.1.6 for the metadata carried by each upstream finding, and 5.1.2 for case facts that must remain exact outside a summary.

Do not turn the opening into another long report. If every paragraph is declared important, the aggregate loses the navigational value of a short findings map. Preserve the deciding facts and their references, then let the detailed sections carry the evidence and qualifications.

Give each source an explicit structure

Anthropic recommends XML tags for document content and metadata in multidocument prompts. Its documented pattern wraps each source in <document>, with <source> and <document_content> subtags. These tags are prompt text, not Messages API fields or an automatically validated document schema.

Use explicit section headers within a report or Markdown aggregate as the exam guide describes. Label each source or concern clearly enough to distinguish it from neighboring sections. A heading such as “Shipment issue — ORD-744” tells Claude what the section describes; several repeated headings called “Result” do much less.

Keep the source’s content together with its label. Separate source names from evidence only if the connection remains explicit. Otherwise a later synthesis can attach a correct fact to the wrong report even when both appear in the prompt. Existing attribution requirements still apply; layout does not replace them.

For long inputs, distinguish the key-findings section from the detailed evidence. The opening gives an overview; the detail preserves what supports, qualifies, or contradicts that overview. If a finding says a treatment improved outcomes in one cohort, the corresponding section must still contain the cohort limitation rather than only the headline result.

Put the task after the evidence

Anthropic’s long-context guidance addresses large documents or data-rich inputs (20k+ tokens). It recommends long documents near the top of the prompt, above the query, instructions, and examples, and notes that queries at the end improved response quality by up to 30 percent in tests, especially for complex multidocument inputs.

The qualifier matters: that figure is an observed improvement in the described tests, not a promised gain for your application. You do not need to reach a particular token count before using clear structure, and the guidance does not imply that a shorter prompt cannot omit evidence.

Combine the recommendations by starting the aggregated data with its key-findings map, following it with labeled longform evidence, and placing the analysis question afterward. This describes the ordering of the supplied data and task, not moving untrusted source text into the top-level system field.

Ask a specific question that directs attention to the relevant comparison. For a report with one contrary source, require the synthesis to explain which findings agree, which conflict, and what limits the conclusion. “Summarize everything” offers less guidance about the decision that the evidence must support.

Ground synthesis in relevant passages

Anthropic also recommends asking Claude to quote relevant document passages before performing a long-document task. That can help focus the response on supporting content instead of relying only on the opening summary.

Use this as an evidence check, not a request to reproduce the entire input. Select passages that support the answer or expose a conflict, then synthesize their implications. See 1.3.4 for evidence attribution. This lesson adds the input-layout choice that helps the synthesis find those passages.

Key distinctions and trade-offs

Missing evidence versus overlooked evidence: If the request omitted a finding, reordering cannot restore it. If the finding is present but buried, an early map and explicit structure address a different failure.

Key findings versus detailed source material: The opening supports navigation and decision focus. Detailed evidence supports verification and qualifications. Neither substitutes for the other.

Position guidance versus guaranteed recall: Good structure reduces avoidable friction; it does not prove that every claim will be used correctly. Check the synthesis against its required evidence.

Early data versus final query: The exam guide’s early findings and Anthropic’s end-positioned query are compatible. One organizes the evidence; the other states what to do with it.

Architecture / decision scenario

A research coordinator aggregates six reports about a support automation rollout. Most report lower handling time, but a report that finished third finds increased escalation for customers with several concerns. The final answer praises faster handling and never mentions the escalation finding, even though the source was supplied.

Rebuild the aggregate with a short opening findings section that names both outcomes and points to their source sections. Give every report an explicit heading, and preserve its evidence with its source label. With a document-style prompt, use <document>, <source>, and <document_content> for the detailed reports.

Place the task after those reports: compare handling time and escalation, identify the relevant customer groups, and quote passages supporting material disagreements before drawing a recommendation. The contradictory finding now appears in the opening map and in its labeled detail rather than relying on its arrival position alone.

If the follow-up answer still omits escalation, inspect evidence coverage instead of adding another generic instruction to “pay attention.” The required comparison has two dimensions, and both need a supported conclusion or an explicit evidence gap.

Common exam traps

• Concatenating reports in arrival order without an early findings map.
• Putting only favorable findings at the start and burying exceptions.
• Removing source labels while making the aggregate shorter.
• Treating XML tags as API-enforced schema validation.
• Turning the tested 30 percent improvement into a universal promise.
• Expecting input ordering to recover facts omitted during summarization.

Where the exam guide differs

What to remember

Start aggregated data with important findings, keep detailed evidence explicitly labeled, and put the query after longform material. Use relevant passages to ground synthesis, then verify that the answer covers the findings that decide the question.

Knowledge check

References

• Anthropic, Prompting best practices
• Anthropic, Claude Certified Architect – Foundations Exam Guide