The shape of a good answer

An answer’s content can be correct and well cited and still fail, because it’s four paragraphs when the question needed one sentence, or because the actual answer is buried in the middle after a paragraph of preamble. Shape is a design decision, and left unspecified the model will pick one — usually longer and more hedged than you want.

Answer first

The single most important rule: the first sentence contains the answer.

Not context, not a restatement of the question, not a description of what the sources cover. The answer.

Compare:

The question of holiday carry-over is addressed in several places in the documentation. According to the policy documents, there are provisions relating to unused days, and these have been updated recently. Specifically, the current limit is ten days.

The current carry-over limit is ten days [2]. This was raised from five in the January 2026 update.

Same information. The second respects the reader. Most people read the first sentence and stop, so the first sentence has to carry the answer.

This needs instructing, because the natural output pattern is to warm up:

Begin with the direct answer in the first sentence. Do not restate
the question or describe the sources before answering.

Length

Match it to the question. Specify a ceiling and let short answers be short.

A factual lookup — one or two sentences. “What’s the limit?” does not need three paragraphs.

A why or how question — a short paragraph or two.

A comparison — a couple of sentences per item, or a small table.

A procedure — numbered steps, one action each.

Unspecified, models produce medium-length answers regardless of question, which means lookups get padded and complex questions get compressed. Both are worse than matching.

Instruct with a ceiling, not a target: “at most three short paragraphs” permits one sentence. “About three paragraphs” guarantees three.

Structure

Prose for anything explanatory. Bullets fragment reasoning that needs connecting.

Bullets for genuinely parallel items — a list of things, not a chain of thought.

Numbered steps for sequences where order matters.

Small tables for comparisons on shared dimensions. Two or three columns; wide tables don’t survive narrow screens.

The common failure is bulleting everything, which looks organised and destroys the connective reasoning that made an explanation useful. Bullets are for lists.

Caveats and hedging

The tension: qualifications are often essential, and hedging everything makes an answer useless.

Rules that hold up:

Preserve the source’s hedging exactly. If the document says “may”, the answer says “may”. Never make a claim more definite than its source.

Don’t add hedging the source doesn’t have. Turning a clear statement into “it appears that possibly” misrepresents it in the other direction.

Put caveats after the answer, not before. A condition stated before the answer delays it; stated after, it qualifies something the reader already has.

Distinguish source uncertainty from system uncertainty. “The policy states this may vary by region” and “I’m not certain the sources fully cover this” are different claims and shouldn’t be phrased alike.

What accompanies the answer

Alongside the text:

Sources, resolvable. Not just numbers — the document, ideally the location, with the supporting passage available.

Dates on sources. Readers evaluate currency far better than any system does.

What was searched, at least on refusals and thin answers. Scope is often the missing information.

Gaps, explicitly. If the sources answered part of the question, say which part they didn’t. This is the difference between a partial answer and a misleading complete-looking one.

A conflict note, where sources disagreed.

Things to instruct against

Predictable habits that add nothing:

Restating the question. “You asked about holiday carry-over. Regarding holiday carry-over…”

Narrating the sources. “The provided documents contain information about…” The user wants the information, not a description of its container.

Closing offers. “Let me know if you’d like more detail.” Fine in conversation, noise in a question-answering system, and it appears on every single response.

Apologising. Especially on refusals. “I don’t have information on that” is better than a paragraph of regret.

Generic advice. “You should consult your HR department” appended to a well-sourced answer is padding. It’s appropriate on a refusal, where it’s an actual next step.

A shape specification

Concrete enough to use:

Begin with the direct answer in one sentence.
Then, if needed, one short paragraph of supporting detail.
Then any conditions or exceptions the sources state.
Cite each factual claim.
If the sources only partly answer the question, say which part
they do not cover.
Do not restate the question, describe the sources, apologise, or
offer further help.
Maximum three short paragraphs.

Unglamorous, and it produces answers that are noticeably better to read than the default — for the cost of a few lines in a prompt you’re already writing.