English or Your Own Language: Which to Write Your Prompts In

The tools understand most languages, but some cases justify English. When to switch, how to ask for a translated answer, and the test to run.

3 min read

Road sign in several languages
Photo: Julian Nyča · CC BY-SA 3.0, cropped

A master's student in Lisbon reads on a forum that "prompts in English give better results." So she spends a week laboriously translating her requests with a dictionary open beside her. Her prompts are shorter, flatter, and full of phrasings she doesn't master. The answers don't improve, and she loses the precision she had in her own language.

The advice circulates a lot and deserves some nuance. To write an email, summarize a text, or prepare a class, write in your own language.

For most uses, write in your own language

The major assistants understand most widely spoken languages well. The decisive advantage is on your side: you are precise in your own language. You add nuance, you cite an exact term, you describe your reader without hesitating. That precision weighs more than the possible performance gap between two languages, which depends on the tool, the task, and the moment.

If you want the answer in another language, ask for it. One sentence is enough: "Answer in Spanish."

The cases where English holds up

Image generators. For Midjourney or Stable Diffusion, the common practice is to write in English, and the results are more predictable. The articles on Midjourney and Stable Diffusion cover this. In ChatGPT or Gemini, other languages generally work fine.

A field's technical vocabulary. For a programming or science topic where everyone uses the English words, keep them: pull request, prompt, cache. Mixing is normal.

Sources in English. If you work on English documents, write your instructions in your own language and leave the text as is. No need to translate everything.

A project that must stay readable by others. A prompt shared with an international team, or kept in a code repository, is better filed in the common language.

What the choice of language really changes

Models have read more text in some languages than in others, which can steer the answers: an example, a legal reference, a politeness habit. A question about employment law may get an answer that follows foreign rules if you don't specify the country. So write the frame: "in Canada," "as of DATE]." The article on [AI bias touches on this reflex.

Languages other than English also use somewhat more tokens than English for the same length, as the article on tokens points out. That matters for very long texts, rarely for an ordinary prompt.

Write in your language, deliver in another

The most useful practice: keep your language for the request and set the language of the deliverable.

A nonprofit treasurer in Portugal has to write to a partner in the UK:

Prompt
Here is my email in Portuguese:

"""
[PASTE YOUR EMAIL]
"""

Write it in British English, professional but simple register, without copying the sentence structure of Portuguese. At the end, list the expressions you adapted rather than translated word for word, so I can check the meaning.

Reread the result. If you read English, notice the phrasings that are too formal. If you don't read it well, have it translated back into your language in a new conversation and compare with your original. For a document that binds you (contract, job application, official letter), have someone who masters the language review it.

The test to run on your own case

No general rule beats your own trial. Take a prompt you use often, write it in both languages, run each version three times in fresh conversations, and compare. If you see no difference, stay in your own language. If English is clearly better on your task, keep the English version for that task only. The test costs ten minutes and saves you from following forum advice.

Getting started

Dialing In the Tone of an Answer Without Spending an Hour

Describing a tone with adjectives doesn't work. Three quick methods: situate the reader, show a model, move the slider one notch at a time.

3 min read

Getting started

The follow-ups that turn an average answer around

Eight ready-to-paste follow-ups to improve a lukewarm AI answer: remove the generic, compare versions, get a critique, cut, verify.

4 min read

Getting started

Test a Prompt Before You Rely on It: A Simple Method

Tricky cases, written criteria, repeated runs: how to check in half an hour that a prompt holds up before using it for real.

5 min read