Why moving the length setting changes so little

Replies run to six or seven paragraphs in a panel the size of a phone screen, and dropping the length setting either changes nothing or starts cutting sentences in half.

In rough order of how often each one turns out to be the answer. Work down rather than across: each carries a way to tell whether it is yours before you change anything.

• 1

  The cap is a ceiling and the reply is not hitting it

  Why

  Brief allows 500 output tokens, Standard 1,200 and Thorough 2,500. Nothing about those numbers asks for a shorter answer. If a reply finishes under the ceiling, lowering the ceiling changes nothing at all until it drops below what the model was going to produce anyway, at which point the reply stops mid sentence.

  How to confirm it is this one

  Read the end of a long reply. If it finishes properly, the cap is not involved and moving it will not help. If it stops mid sentence, the cap is already being hit and the length you are seeing is the maximum rather than the intent.

  Fix

  Leave the cap where it is and change what feeds it, using the causes below. Move the cap only when replies are being cut off, and then upwards.

• 2

  Too many passages are being put in front of the model

  Why

  Passages it reads decides how many pieces of your material are shown. Six is the default. More material is more for the model to summarise, and it is also more chances for a marginally relevant passage to earn a paragraph it did not deserve.

  How to confirm it is this one

  Behaviour tab, the Passages it reads box. Anything meaningfully above 6 is the most likely cause of length on an assistant with a lot of indexed material.

  Fix

  Bring it back to 6, or to 4 if your material is repetitive. The setting's own note in the dashboard puts it plainly: more gives it context and crowds out the best match.

• 3

  Your own instructions are pulling against the brevity rule

  Why

  The instructions you write are placed ahead of the standing rules, and the model follows both. Asking it to be thorough, to explain fully, to give examples, or to walk somebody through something is a direct instruction to write more, and it wins over a general request to be concise.

  How to confirm it is this one

  Behaviour tab, read your own instructions back. Look for words like thorough, detailed, comprehensive, step by step, and for any instruction to always include something, which is the quiet one, because always means in every answer.

  Fix

  Cut those words. Instructions about tone and scope work well here, instructions about volume fight the rules and produce exactly this. If you want detail on one topic only, say so about that topic.

• 4

  The material itself is long-form and gets quoted at length

  Why

  The model answers from the passages it is shown. A passage taken from a page written as an essay arrives as an essay, and asking for a short answer built out of long undivided source text is asking the model to do editing work it is not reliably good at.

  How to confirm it is this one

  The widget shows a numbered source under every grounded reply, linking to the page it used. Open it. If the cited page is one long section with no headings, that is where the length is coming from.

  Fix

  Break the page into headed sections that each answer one thing. This helps twice: shorter answers, and better retrieval, because a heading worded like the question is the strongest signal the matcher has.

• 5

  Brief was set and replies got worse rather than shorter

  Why

  The reply ceiling covers everything the model produces on its way to an answer, not only the words that reach the visitor. A tight ceiling can be consumed before the visible answer is finished, which is not a shorter answer, it is a truncated one.

  How to confirm it is this one

  On Brief, look for replies that come back very short, stop abruptly, or arrive close to empty on questions that used to be answered properly. That pattern is the ceiling, not the material.

  Fix

  Go back to Standard and shorten the answer through the passage count and your own instructions instead. Brief suits an assistant answering opening hours and delivery times, not one explaining a policy.

One thing that will not shorten anything is worth knowing about, because it looks like a lever. The material put in front of the model is capped at 4,000 tokens, and when the passages exceed it the last one is truncated rather than dropped, so a citation can never point at something the model was not shown. Past that point, raising the passage count adds no more material: it only means the final source arrives half-read. Conversation history is bounded the same way, at the last three turns, so a long conversation is not what is inflating a reply either.

Why is there no minimum length setting?

Because a floor would be a licence to pad. A one-line question deserves a one-line answer, and the only honest way to make replies longer is to give the assistant more worth saying.

Can I ask for bullet points instead of paragraphs?

Yes, in your own instructions on the Behaviour tab, and it is one of the most effective things you can put there. Bullets sit better in a narrow panel than paragraphs do, and they make an over-long answer obvious at a glance rather than after reading it.

The answer is long because it covers three things I did not ask about. Why?

Usually too many passages, occasionally material that mixes three topics on one page. Check the numbered sources under the reply: if they point at three different pages, lower the passage count. If they all point at one, that page is doing too much.