How to Prompt AI to Give a Shorter Answer
Brevity is a judgement and shape is a rule. Constrain the shape instead, and one format swap that does more than any length instruction.
Add "be concise" to a prompt and you will get a slightly shorter version of the same sprawling answer. Add "in one sentence" and you will get one enormous sentence with four semicolons. Neither is the model ignoring you. To prompt AI to give a shorter answer reliably you have to constrain the shape of the output rather than ask for the property of brevity, because brevity is a judgement and shape is a rule.
Why the usual instructions fail
"Be concise" fails because concise is relative and the model has no reference point. Concise compared to what? It shortens by some unspecified amount from a baseline you cannot see, which is why the effect is real but small and unpredictable.
"In one sentence" fails differently. It is a precise constraint on the wrong dimension. Sentence count is not length, and the model will satisfy the letter of the instruction by packing everything into a single clause-heavy monster. You asked for one sentence and that is exactly what you got.
The general principle: instructions that name a quality get interpreted, instructions that name a countable thing get obeyed. Length instructions work when they are countable and when the unit actually correlates with length.
Constraints that work when you prompt AI to give a shorter answer
In rough order of how reliably they work.
A word budget with a ceiling and a floor. "Between 60 and 90 words." The floor matters as much as the ceiling: without it you sometimes get a fragment, and the range gives the model a target rather than a limit to approach.
A structural template. "One line stating the answer, then at most three bullets, then nothing." This is the strongest option because there is no room left to expand into.
A named format with inherent limits. "Answer as a commit message." "Answer as a table with two columns." The format carries the length rule with it.
An explicit exclusion list. "No preamble, no restating my question, no summary at the end." These three habits are most of the excess in a typical answer.
A hard cutoff with a consequence. "If it does not fit in 50 words, say it does not fit and give me the single most important point." This prevents cramming.
Word budgets are approximate, because models count tokens rather than words and cannot see their own output length while producing it. Expect plus or minus twenty percent and set the range accordingly. If you need an exact length, generate then trim in a second pass. Vendor prompt engineering guidance makes the same general point: specific, checkable instructions outperform adjectives.
The format swap that beats all of them
If you want one change that does more than any length instruction, stop asking for prose. Prose is the format that expands, because connective tissue between ideas is what prose is made of. A table of the same content has nowhere to put the connective tissue.
Asking for a table, a checklist, a decision tree, or a list of name-value pairs typically cuts length by half or more with no content loss, because the content was never the long part. This is also why the same question answered in the chat interface and through the API can differ so much in length: the default framing differs, an effect covered in more depth in why your prompt works in chat but not in the API.
Cutting length without losing the answer
There is a failure mode on the other side. Squeeze too hard and the model drops the qualifications that made the answer correct, leaving you with something confident and wrong. Length and accuracy trade off at the margins, and a very short answer to a genuinely nuanced question is a lie by compression.
Two ways to keep both:
Split the response. Ask for a short answer plus a separate one-line flag naming anything important it had to leave out. You get brevity and you get told what it cost.
Shorten the scope, not the answer. Narrowing the question produces a short complete answer, where shortening the response produces a truncated one. These look similar and are not.
Be careful not to confuse verbosity with hedging. An answer bloated by qualifiers and disclaimers is a different problem with a different fix, and pushing on length alone will make it terse but still evasive. If that is what you are seeing, stopping the hedging is the thing to fix first, and the length usually falls out of it. If the answers are long because the request itself is muddled, fixing the prompt beats trimming the output, and the general principles in prompt engineering apply here too.
A prompt that works
Combining the above, for a question where you want a usable short answer:
Answer in 60 to 90 words.
Format: one sentence with the direct answer, then at most three bullets of supporting detail.
No preamble. Do not restate my question. No closing summary.
If something important does not fit, add a final line starting "Omitted:" naming it in under 10 words.
Question: <your question>The omission line is the part people skip and it is what makes the whole thing safe to rely on. You are not just getting a shorter answer, you are getting a shorter answer that tells you when shortness cost you something.
FAQ
Why does AI ignore be concise?
It does not ignore it, it under-applies it. Concise names a quality with no reference point, so the model shortens by an arbitrary amount from a baseline you never see. Countable constraints like a word range produce far more consistent results.
What is the best word count to ask for?
Give a range rather than a maximum, and set the floor at roughly two thirds of the ceiling. A range gives the model a target to hit; a bare maximum invites it to approach the limit or, occasionally, to return something too thin.
Will asking for shorter answers make them less accurate?
It can, if the question genuinely needs nuance. Compression drops qualifiers before it drops claims, so you end up more confident and less correct. Asking the model to flag what it omitted is the cheapest guard against this.
Does asking for a table really shorten things?
Usually by a lot. Most of the length in prose is connective tissue between ideas rather than the ideas themselves, and a table has nowhere to put it. It is the single highest-leverage change if your content suits a tabular shape.
How did this land?
About the author
Developer Advocate
Steve builds something with Swarmz every week and writes up what worked, what broke, and what he'd do differently. Tutorials and hands-on guides are his lane.
