A bigger context window still needs a brief

A large upload is not a clear assignment

You can give a model a large collection of documents and still leave it guessing about the job. A folder may contain old prices, current prices, draft policies, final policies, meeting notes, and material that is only loosely related. Without a brief, the model has to infer which sources matter and what kind of answer you expect.

The first improvement is usually simple: write the question before preparing the upload. “Summarize these files” produces a very different task from “identify the current delivery commitments, show where the documents disagree, and list the information needed before we can make a promise to a customer.”

The second brief provides a decision, a source problem, and a useful output. It also gives the reviewer something specific to check. That is the starting point for making long context practical.

Give the documents an order

Create a short source index. Include the file name, date, owner if known, and its role in the task. Identify which source is authoritative when there is a conflict. If you do not know, say that the conflict must remain unresolved until someone checks it.

For example, a signed product specification may outrank an earlier sales presentation for a technical dimension. A current published price may supersede a draft planning sheet. Those are rules you establish for the task, not assumptions the model should make on its own.

Keep irrelevant files out of the first attempt. You can add them if a question requires them. This makes the work easier to inspect and reduces the chance that a useful answer gets buried under a general summary of everything in the folder.

Ask for evidence in the output

Require a source reference for each important claim. A practical format is: claim, supporting file and section, relevant passage, and any uncertainty. If the source does not support the claim, the model should leave it open rather than complete the pattern with a guess.

For numerical work, ask for the original value, unit, time period, and calculation. “Revenue increased” is not enough if one document reports a quarter and another reports a full year. The source boundary is part of the answer.

You do not need an essay-length explanation for every routine detail. Focus evidence requirements on the facts that could change a decision. A concise table with five well-supported findings is often more useful than a long narrative with a bibliography attached at the end.

Test where the answer sits

The 2024 paper Lost in the Middle found that the position of relevant information affected performance on the models and tasks it studied. It is a useful reason to test retrieval from your own document pack. It is not evidence that a particular 2026 model will fail in the same way. ^[1]

Build a small check set. Include a fact near the beginning, a fact in the middle, a fact near the end, a contradiction across files, and a question with no answer in the pack. Keep the expected answers separate from the material shown to the model.

If performance changes when you reorder the documents, record it. Try a shorter pack, a clearer index, or a staged workflow before deciding that you need a larger context window. The objective is to obtain a useful answer reliably, not to use every available token.

Use stages when the job needs them

For a complicated review, separate collection from synthesis. First ask the model to locate relevant passages. Check that selection. Then ask it to compare the evidence and produce a recommendation or summary within the agreed scope.

Anthropic’s context-engineering guidance discusses managing the information available to an agent rather than treating the prompt as the only input that matters. Our practical recommendation here is narrower: keep the task, source selection, and working notes explicit so a reviewer can see what the answer depends on. ^[2]

A staged process may add calls and review time. That can be worthwhile for an important decision and unnecessary for a rough first pass. Choose the amount of structure based on the consequence of a mistake and the effort needed to detect it.

Keep a short working brief

Use the following outline when you hand over a document-heavy task. It is deliberately small enough to read before the files.

1. Decision: what will this work help us decide?
2. Scope: which products, periods, and locations are included?
3. Sources: which files are current, and which conflicts need review?
4. Output: what format would make the result usable?
5. Evidence: which claims need a passage, calculation, or direct reference?
6. Unknowns: what should be left unresolved instead of inferred?

Finish by checking a sample of the actual references, especially those supporting the main conclusion. A model can produce a convincing source label that does not support the surrounding sentence. The reference must survive a human opening the file and reading the relevant part.

A bigger context window gives you more room. A good brief gives that room a purpose. Use both, and keep the evidence close enough that someone else can check the work without repeating the entire task.