Where the instruction goes, and why it matters

Position is not neutral

With a two-line prompt, order is irrelevant. With four thousand words of pasted report it stops being irrelevant, and the effect is large enough to notice without measuring anything.

Three findings, in plain terms.

Instructions after long material are followed more reliably than instructions before it. If the task sits at the top and then six thousand words follow, the model works out what to do, then reads an enormous amount of other text before producing anything. Putting the task at the end, immediately before it answers, keeps the request close to the answer.

Instructions before the material help the model read purposefully. Knowing you want parental-leave rules changes how the rest is processed.

Those two pull in opposite directions, so the practical answer is to do both. Task at the top in one sentence, material in the middle, the specific instruction repeated at the bottom. It costs twenty words.

Extract every payment deadline from the contract below.

<contract>
... 4,000 words ...
</contract>

List every payment deadline in the contract above: date, amount,
and the clause number it comes from. If a date is described rather
than stated ("30 days after signature"), quote the description.

Material in the middle of a long context gets used less well than material at either end. This is the effect usually called lost in the middle, first shown clearly in a 2023 study where the position of a relevant passage inside a long context changed accuracy substantially, with the worst results when it sat in the middle. Newer long-context models handle this better than the models in that study. Better is not solved: independent long-context benchmarks continue to show accuracy falling as the material grows, even for models advertising a million-token window.

What follows for real work

Put the most important document last. If you paste three documents and one of them is the one the question is about, it goes at the end.

Say where to look. “The payment terms are in section 9; the termination clause is near the end.” One sentence of navigation is worth more than a firmer instruction, and it costs nothing to be wrong about: a model told to check section 9 will still read the rest.

Prefer the chapter to the book. A 12-page extract that certainly contains the answer beats a 300-page manual that certainly contains the answer somewhere. This is not about the limit: it is about accuracy inside the limit.

Do not trust a big window as a filing system. A million-token context is a genuine engineering achievement and a poor substitute for choosing what to include. If you have ever pasted an entire codebase and got a confidently wrong answer about one function, you have met this directly.

Watch for the quiet quality drop. The failure here is not an error message. It is a slightly worse answer that looks exactly like a good one: a summary that misses the third of four conditions, a comparison that silently drops one candidate. Nothing announces that the middle of your document was skimmed. This is why the fix has to be structural rather than reactive: you cannot wait to notice, because the symptom is an absence.

The counter-case

There are jobs where you must give it everything: finding a contradiction between two sections you have not located, checking whether any clause covers termination, summarising a whole report. Selecting in advance would beg the question.

For those, two things help. Ask for an intermediate pass (“first list every section that mentions termination, with clause numbers; then answer”) which turns one hard retrieval into a cheap enumeration followed by a small task. And ask for quotes, so a claim about the middle of the document is checkable.

The habit

Once your prompt exceeds a page, treat it as a document with a structure rather than a message. Top: what I want. Middle: labelled material, most important last. Bottom: the precise instruction again, plus the output shape and the escape hatch.

That layout takes no longer to write and it fails much less often, especially on the day you paste something three times bigger than usual and cannot work out why the quality dropped.