Large context windows let you paste in long documents, but quality depends on how you do it.
Put Documents First
Place long material near the top of the prompt, with your question and instructions after it. Many models answer more accurately when the question comes after the document.
Label Everything
Wrap each document in tags with a title, date and ID so the model — and you — can refer to them precisely.
Ask for Quotes First
For accuracy, ask the model to extract the relevant quotes before answering: "First list the passages relevant to the question in <quotes> tags, then answer using only those passages." This focuses attention and makes answers checkable.
Work Section by Section
For very long documents or detailed analysis, process one section at a time and combine the results.
Beware the Middle
Details buried deep in a long context are more likely to be overlooked. If something matters, point the model to where it is.
Verify Specifics
Numbers, dates, clause references and names should be checked against the original.
Consider Retrieval
When documents are extremely long or you ask many questions about them, retrieving relevant sections may be faster, cheaper and more accurate than sending everything each time.