What it is
A length is a number and a unit: words, sentences, bullets, items, minutes, pages, slides. On the ten checks it is Scope, the fourth check. Pick the unit the destination uses. A speech is measured in minutes, a form field in characters, a slide in bullets, a board pack in pages.
Words such as “short”, “brief” and “detailed” are sizes too, but each model reads them differently, and a model update can change how they are read. A number is the version that survives the trip.
The useful part is often the second half of the instruction: what to keep when the answer has to be shorter than the material. “Under 120 words” tells the model how much; “keep the decision and the deadline, cut the background first” tells it what.
When it helps
It helps whenever the answer has a container. Subject lines, text messages, slides, abstracts with a word limit, a two-minute update in a stand-up, a form with a character count. In each case, an answer that does not fit is a second round of editing.
It helps when you will read many answers, such as replies to triage or summaries of every ticket in a queue, because a fixed size makes them quick to scan and easy to compare. And it helps with cost: output is usually the expensive side of a model call, so a size is often the simplest way to keep a prompt that runs every day cheap. The guide to what an LLM call costs has the arithmetic.
When it does not
A word limit is a target the model aims at, and it may not land exactly on it. Count before you paste into a field with a hard limit; the Token Counter counts words as well as tokens, and the cut to a word limit template trims an existing text to fit. Counts of sentences, bullets or items are easier to check by eye than words.
A count can invite padding. Ask for ten ideas when there are four good ones and you may get six weak ones to make up the number. Add “if there are fewer, list only those”, which is one case of the broader pattern of saying what to do when unsure.
A tight limit with no priority cuts at random. If 100 words is not enough room for everything, the model will decide what goes, and it may drop the one fact the reader needed. Say what must survive.
And some tasks need everything. Pulling every action item out of a transcript, or every clause that mentions liability, should not be capped at five. Set a length for each item instead, such as one line per action, and let the count be whatever it is.
Before and after
Three requests where the after version adds a size. Each score comes from the Prompt Grader’s free ten-check estimate, written for any chat model.
Write a speech for my team's end-of-project lunch. We shipped the new billing system last month after more than a year of work.
Write a speech for my team's end-of-project lunch, one I can say in about two minutes, so roughly 250 words. We shipped the new billing system last month after more than a year of work.
Left to itself, a model asked for a speech can write one long enough for a conference. Given two minutes and a word count, it has room for one story and a thank-you, which is about what a lunch can hold.
List the risks in the project plan below. """ [the project plan, pasted in full] """
List the five biggest risks in the project plan below, most serious first, one sentence each. If you find fewer than five real risks, list only those. """ [the project plan, pasted in full] """
The first version gets every risk the model can think of, at equal weight, from the vendor going bankrupt to someone being ill. The second gets a ranked shortlist you can take into a meeting, and the last sentence keeps it from inventing a fifth risk to fill the slot.
Turn these notes into a slide for the all-hands meeting. Notes: [your notes, pasted in full]
Turn these notes into one slide for the all-hands meeting: a title of at most eight words and no more than four bullets of ten words each. If the notes hold more than that, keep what affects the most people. Notes: [your notes, pasted in full]
A slide with no size can come back with every note on it, which is a page, not a slide. The second version fits on the screen, and the last sentence tells the model what to keep when the notes do not fit. This pair gains two points because bullets also count as a format.
The line to add
Length: {{number}} {{unit}}. If that is not enough room, keep {{what_must_stay}} and cut {{what_can_go}} first.
The unit is whatever the destination counts: words, characters, bullets, items, minutes or pages. If you only add the first sentence, you still get a size. The second sentence is what makes the size safe to set tight.
In ChatGPT, Claude, Gemini and Copilot
The assistants have different default lengths, which is the best argument for setting your own. Google’s guide for Gemini says Gemini 3 models give direct and efficient answers by default, and that you have to ask explicitly for a more detailed or conversational one. Its own example of a constraint is a length: summarize this text in one sentence.
Anthropic’s guide describes its latest Claude models as less verbose than earlier ones, with Claude Opus 5 as an exception whose default answers run longer; for that model it says to prompt explicitly for conciseness. OpenAI’s guide to its latest models notes that recent GPT models are more concise by default, warns that broad instructions such as “be concise” can make answers too brief, and recommends saying what a short answer must keep and what it can leave out, the second sentence of the line above.
Microsoft’s guide to Copilot prompts uses the same move in one of its sample prompts, a thank-you note to be kept under 80 words. A number with a unit is the length instruction that carries across all four.
Key takeaways
- Give a number and a unit, in the unit the destination counts.
- A size is a priority: “the five biggest” makes the model rank.
- Say what must survive when the answer has to be short.
- Treat word counts as a target and check them before pasting into a hard limit.
- Do not cap tasks that need everything. Set a length per item instead.
Questions
How do I make ChatGPT answers shorter?
Give a number and say what must stay: “Under 100 words. Keep the decision and the deadline; cut the background first.” A bare “be brief” is read differently by each model, and OpenAI’s own guidance warns it can make answers too short. For text you already have, the cut to a word limit template trims it to a set count.
Can AI write to an exact word count?
Close to it, but not reliably exact. Ask for the count, then check it before you paste into a form with a hard limit; your word processor or the Token Counter will count it. Counts of sentences, bullets or items are easier to check by eye than words, so use them when the exact word count does not matter.
Why are Gemini’s answers short and Claude’s sometimes long?
Each model has its own default length, and the defaults change between releases. Google says Gemini 3 answers tersely unless asked for detail; Anthropic says one of its recent models runs longer and should be asked for brevity explicitly. Setting a length in the prompt takes the default out of it, whichever assistant you use.
Prompts to use with this
- Cut a text to a word or character limitYou need a text to fit a form, abstract or post limit without losing the main point.