A long document: better to split it up than to hand it all over at once.

The instinct is to paste everything in at once. That's what produces the most disappointing results, and for a specific reason.

Why the middle gets lost

Even when a document technically fits within the context window, the model's attention isn't evenly distributed. The beginning and end are handled well, the middle gets skimmed over. The phenomenon is well-known and measured: in a long text, information placed in the centre is retrieved significantly less well than at the extremes.

Direct consequence: "summarise this 120-page report" produces a summary that mostly reflects the introduction and conclusion. That's often exactly what the author wanted you to take away, not what you're looking for.

The method that works

One chapter at a time, in separate conversations or with clear sequencing. You get a summary per section, and you assemble them afterwards. It takes ten minutes longer, and the result is on another level.

For a document whose structure you don't know, start with: "Just give me the detailed outline, without summarising the content." You'll then know what to break up.

Search rather than summarise

Often, you don't want a summary but an answer. "In this document, what does it say about termination deadlines? Quote the exact passages." The "quote the exact passages" is essential: it lets you verify, and it sharply reduces fabrication.