Give it the brief, not the archive

Extra material is not neutral

There is a comfortable assumption that more background can only help, because anything unnecessary will simply be ignored. It is not ignored. Everything you supply is treated as relevant by default, since you supplied it, and material that has no bearing on the question competes for attention with the material that does.

The visible symptom is an answer that addresses something adjacent to what you asked. You paste a long thread and ask what to do about one item in it, and the response covers the whole thread, weights the loudest part heaviest, and treats your actual question as one item among many.

A second symptom is subtler and worse. A stray document in the pile establishes a fact that is out of date, and the answer is built on it confidently, because nothing in the material was marked as superseded and nothing about the format distinguishes current from historical.

Write what a competent stranger would need

The useful test is to imagine briefing a capable freelancer who starts tomorrow and has an hour. You would not send the entire shared drive. You would write a few paragraphs covering what the thing is, who it is for, what has already been decided, what is off the table, and what you actually want back.

That brief is usually between one and three short paragraphs, and writing it takes about four minutes. It outperforms twenty pages of pasted material with striking regularity, partly because it is selective and partly because the act of writing it forces you to work out what you are asking.

The parts people leave out of the brief are consistent. Constraints that feel obvious to you because you have lived with them for months. Decisions already taken, which the tool will happily reopen. And the audience, which changes almost everything about a useful answer and is stated in maybe one prompt in five.

Keep the material and the instruction visibly apart

When source text and instructions are mixed into one block, the boundary is guesswork. Sentences from your material get read as directions, and your directions get read as part of the subject matter, which produces answers that summarise your own instruction back to you.

Marking the boundary explicitly fixes most of it. A labelled section for the material, another for what you want done with it, and a plain line saying that the first section is reference only. It is unglamorous and it removes a whole category of confusion.

The same separation matters for anything arriving from outside your control. Text pulled from a web page or a forwarded message can contain sentences that look like instructions, and if the boundary is unmarked they can be followed. Labelling untrusted material as material is a small habit with a real security benefit.

Reference material should be cited, not just supplied

If you do attach a long document, add the requirement that any claim drawn from it must quote the sentence it came from. This changes the character of the output considerably. It becomes checkable, because you can search for the quoted line, and it becomes shorter, because unsupported assertions have nowhere to hide.

It also surfaces the case where the document simply does not contain the answer. Without the quoting requirement, that case is filled in from general knowledge and reads exactly like the supported claims. With it, the gap becomes visible, provided you also say what to write when nothing in the source applies.

None of this guarantees the quote is real. Verifying that the cited sentence exists in the document is a mechanical check and worth doing on anything consequential, particularly when the source is long and you are unlikely to read it in full yourself.

When to hand over everything after all

There are genuine cases for volume. Searching a body of text for a mention, comparing two versions of a contract clause by clause, or asking what a long thread actually decided are all tasks where the material is the point and pre-selecting it would beg the question.

The distinguishing feature is whether you know in advance which part matters. If you do, select it. If the whole job is to find out, supply the lot and accept that the middle of a very long input gets thinner treatment than the ends, which is a good reason to ask a narrow question rather than an open one.

What rarely works is the middle position, where a large quantity of half-relevant material is supplied as insurance. That combines the cost of volume with none of the benefit of a search, and it is the most common shape a briefing takes when nobody has decided what the question is.