P

Take A Deep Breath: Definition and Examples

Prompt engineering technique that asks the language model to "take a deep breath" before answering, encouraging it to slow down its reasoning and produce more thoughtful and accurate responses.

Full definition

"Take A Deep Breath" is a prompt engineering technique popularized in late 2023, notably by a Google DeepMind study, which demonstrated that adding this simple phrase to a prompt could significantly improve language model performance on complex reasoning tasks. This expression acts as an emotional and cognitive signal that modifies the model's behavior.

The principle relies on the fact that LLMs have been trained on vast corpora of human texts where the phrase "take a deep breath" is associated with moments of concentration, calm, and deep reflection. By including this instruction, the model activates text generation patterns linked to more methodical and structured thinking, resulting in higher quality answers.

This technique is part of the broader category of "emotional prompting," which exploits the fact that language models are sensitive to affective and motivational signals in instructions. The DeepMind study showed improvements of up to 8% on math benchmarks like GSM8K when this phrase was combined with Chain-of-Thought prompting.

Although the exact mechanism remains debated in the scientific community, the prevailing hypothesis is that this instruction acts as a "meta-prompt" that encourages the model to allocate more tokens to intermediate reasoning before producing its final answer, thus reducing hasty errors.

Etymology

The expression "Take a deep breath" comes from common English language, used to calm someone before mental or physical effort. Its application in prompt engineering was formalized in September 2022 by the Google DeepMind paper titled "Large Language Models as Optimizers," where researchers discovered that this phrase, automatically generated by an optimizer LLM, outperformed more elaborate manual instructions on mathematical reasoning tasks.

Concrete examples

Solving a complex mathematical problem

Take a deep breath and work through this step by step. A train leaves Paris at 2:00 PM at 120 km/h. Another leaves Lyon at 3:00 PM at 150 km/h heading toward Paris. The distance between the two cities is 465 km. At what time do they meet?

Critical analysis of an argument

Take a deep breath and think carefully. Here is a sales pitch. Identify logical fallacies, unsubstantiated claims, and any manipulation techniques: [TEXT]

Debugging complex code

Take a deep breath and analyze this code carefully before answering. This Python program produces an intermittent segmentation fault. Examine each function and identify the root cause: [CODE]

Practical usage

Add "Take a deep breath" at the beginning of your prompts when asking questions that require complex reasoning, such as math problems, logic, or analysis. Combine this technique with Chain-of-Thought ("step by step") to maximize its effectiveness. This method is particularly useful when the model tends to answer too quickly and makes careless errors on multi-step tasks.

Related concepts

Chain-of-Thought PromptingEmotional PromptingStep-by-Step ReasoningZero-Shot Prompting

FAQ

Does "Take a deep breath" really work, or is it a placebo effect?
The results are measurable and reproducible. The Google DeepMind study demonstrated statistically significant improvements on standardized benchmarks. The model does not actually "breathe," but this instruction activates more methodical reasoning patterns from its training data. However, the effect varies depending on the model and task.
Should the phrase be written in English, or can it be translated into French?
The English version "Take a deep breath" is the most tested and documented. The French translation "Prends une grande inspiration" may work, but results are generally better in English because models are mostly trained on English-language corpora. You can write the rest of your prompt in French while keeping this opening in English.
Is this technique still effective with the latest models like GPT-4 or Claude?
Effectiveness varies depending on the model and version. The latest models already incorporate more advanced reasoning mechanisms, which may reduce the marginal impact of this technique. However, it remains useful as an additional signal, especially when combined with other methods like Chain-of-Thought. Test on your specific use case to evaluate the actual gain.

See also

How to use this prompt

  1. Copy the prompt with the button above.
  2. Paste it into ChatGPT, Claude or your favorite AI assistant.
  3. Replace the bracketed variables with your details, then refine the result.

About Prompt Guide

Prompt Guide is a free library of 2500+ ready-to-use prompts for ChatGPT, Claude and other AIs, with guides to learn prompting and tools to build and optimize your own prompts.

More definitions

Get new prompts every week

Join our newsletter.