Choosing what not to include

The skill nobody teaches

Everyone tells you to give more context. Almost nobody tells you what to cut, and by the second month of using these tools seriously that becomes the binding constraint. You have three reports, a spreadsheet, an email thread and a policy document, and the honest question is which two pages of that belong in the prompt.

Three reasons cutting matters, and only the first is about limits.

The window is finite. Obvious, and the least important, since windows are now large.

Accuracy falls as the pile grows. From the position lesson: relevant material is used less reliably when it is buried among irrelevant material. A model given ten documents to answer a question about one of them does measurably worse than a model given the one.

Irrelevant context actively misleads. This is the underrated part. Include last year’s pricing sheet “for background” and some of last year’s prices will appear in this year’s answer, presented as current. The model has no way to know which document you consider authoritative unless you say so.

A method that takes two minutes

Before assembling a long prompt, write the answer’s skeleton. Not the answer: the shape of it. “A recommendation between supplier A and B, based on unit price, lead time and minimum order.”

Now go through your material and ask of each piece: which line of the skeleton does this feed? Anything that feeds nothing comes out.

Applied to the supplier decision: the two quotes go in. The email thread negotiating a different contract does not. Your notes from the site visit go in if they mention lead time. The company handbook does not, however relevant it feels.

You will cut about two thirds. The answer gets better, not worse, and you will have found that you were including things because you had them, not because they mattered.

When you cannot cut, do it in two passes

Some questions need the whole corpus: finding a contradiction, checking whether anything covers a topic, summarising. Do not choose in advance; narrow in the open.

Pass one, enumeration. “List every section that mentions payment terms, with its heading and clause number. Do not analyse.” Cheap and hard to get wrong.

Pass two, the real question, with only the sections pass one found.

This is the same move a competent researcher makes with a paper index, and it works for the same reason: locating and reasoning are different jobs, and doing them at once does neither well.

It also gives you something to check. If pass one lists four sections and you know there are six, you have caught the failure before it reached your conclusion, which you could not have done from a fluent final answer that quietly used four.

Rank what you do include

When several documents are in the prompt and they disagree, say which wins.

<policy2026> is authoritative. <policy2024> is included only so you can tell me what changed. Where they conflict, follow 2026 and note the difference.

Without that line, you get a blend, and a blend of two policies is a policy that does not exist. Nobody wrote it, nobody approved it, and it will read as authoritative as either source.

The signs you have included too much

• The answer mixes facts that come from different documents without saying which.
• Figures appear that are correct for some other year, region or version.
• The response gets longer and vaguer as you add material.
• The model starts hedging: “depending on which policy applies…”. That sentence is a symptom, and it means you left the disambiguation to it.

The counterweight

None of this licenses going back to describing your problem instead of showing it. The failure this lesson describes is rarer than under-supplying context, and it appears later, once someone has learned the first lesson well.

The target is not minimal and not maximal. It is everything that bears on the answer, and nothing that does not, with a stated order of authority when two pieces disagree. Getting there requires knowing what the answer looks like before you ask, which is why the two-minute skeleton is the whole method.