One word in the system prompt cuts AI verbosity: “laconic”, and word count drops by up to 86%

🗣️ We tell the model “be brief” and we keep getting walls of text.

Changing a single word in the system prompt brought verbosity down by 62%, while keeping the information that matters.

The word that changes the result

The word is “laconic”.

It means expressing a complex idea in a way that is short, clear and direct. It sounds like a nuance, and yet it changes the result quite a bit. And someone measured it.

How the test on Claude, ChatGPT and Gemini was run

The test: three prompts built on purpose to generate long answers, given twice to Claude, ChatGPT and Gemini. Once without the laconic directive and once with it.

Every combination was repeated three times, in incognito, with reasoning turned on to push each model to ramble. A score based on 11 criteria also checked how much of the important information survived.

How far the word count drops

The results on word count:

• ChatGPT: −82.7%
• Gemini: −86.4%
• Claude Opus 5: −62.1% (still the wordiest of the three)

The A/B test is documented in the Laconic_tests repository by High Sierra Labs: three models (ChatGPT 5.5, Claude Opus 5, Gemini 3.1 Pro) with extended thinking on, three decision prompts and three seeds per cell, for 54 responses scored blind by two scorers.

Fewer words, same substance

The interesting part goes beyond the number of words removed.

The content score stayed high: the directive mostly cut the filler, keeping the substance of the answer.

In that rubric the verdicts hold at 9 out of 9 on ChatGPT and Claude and 6 out of 9 on Gemini; across the 33 tracked details, Claude keeps 29, ChatGPT 22 and Gemini 18.

Why this is different from the internal flag in Claude Code

What we’re talking about here is different from the internal flag in Claude Code, the environment variable that switches the preset.

This technique is far more portable: one line in the system prompt of any model, in chat or via API. And it works outside code too, for emails, analyses and documents.

Four rules for a laconic system prompt

💡 How to reduce verbosity without losing substance:

1. Ask the model to be laconic.

Define what you mean, too: expressing complex ideas in a way that is short, clear and direct. A plain request for brevity can push the model to cut useful details as well.

2. Explicitly ban filler.

No preambles, no recaps of what you just asked, no lists of alternatives when you only need one.

3. Put the directive in the system prompt.

That way it shapes the whole session instead of just the next answer.

4. Plan an exception.

When you want to go deeper on a new topic, explicitly ask for an extended version. Brevity stays the default, with the option to expand when you need it.

The goal is the same substance in less space.

Verbosity adds words. Quality adds information.

How do you keep models concise today? Did you know this trick?