How to Prompt AI in Your Own Language (Not Just English)

You do not need to write in English

A common assumption is that AI only works properly in English, so people translate their request, get an English answer, then translate it back — losing nuance twice. For widely spoken languages that detour is usually unnecessary.

Anthropic publishes one useful reference point. On a professionally translated version of a standard multiple-choice knowledge test (MMLU), its Claude Sonnet 4.5 model with extended thinking scored, relative to its own English score: 98.2% in Spanish, 97.8% in Brazilian Portuguese, 97.5% in French, 97.2% in Arabic, 96.9% in Simplified Chinese, 96.7% in Hindi and 95.4% in Bengali. In lower-resource languages the gap widens — 91.1% in Swahili and 79.7% in Yoruba — and it widens more for the smaller Claude Haiku 4.5 model (52.7% in Yoruba).

Two cautions about reading those numbers. They are one provider’s results on a knowledge test, not a measure of how natural the writing sounds; and other assistants will differ. The practical lesson still holds across tools: in major languages, write in your own; in less widely used languages, prefer a larger model and check the output more carefully.

1. Name the output language — and the variety

Assistants usually reply in the language you write in, but not always: an English system prompt, an English document in the conversation, or a mixed-language message can pull the answer into English. Anthropic’s guidance is that stating the input and output languages explicitly improves reliability, even though its models can usually detect them.

Be specific about the variety, because they differ in vocabulary, spelling and formality: Brazilian or European Portuguese; Spanish for Mexico, Spain or Argentina; Simplified or Traditional Chinese; Modern Standard Arabic or a regional dialect; Canadian or European French; formal Hindi or everyday Hindi with English loanwords.

• Weak: "Write a welcome email for new customers."
• Strong: "Write a welcome email for new customers in Mexican Spanish, using usted. Natural and warm, as a native speaker in Mexico City would write it."

2. Use the native script, not transliteration

Anthropic recommends submitting text in its native script rather than transliterated into Latin letters — Devanagari for Hindi, Arabic script for Arabic or Urdu, and so on. Transliteration has no single standard spelling, which makes the input more ambiguous.

If you actually want a mixed or romanised style — Hinglish for a social post, for example — say so explicitly: "Reply in Hinglish, written in Latin script, the way people text in Mumbai." The model will not reliably guess that you want it.

3. Ask for localisation, not translation

Translating word for word produces text that is technically correct and obviously foreign. Localising adapts it for the reader: idioms that exist in their language, local currency and number formats, date order, units, examples and names that fit, and the right level of formality — tu or vous, tú or usted, aap or tum.

Anthropic’s own suggestion for fluency is to ask for "idiomatic speech as if it were a native speaker", and it notes that effective communication often needs cultural and regional awareness beyond translation.

• "Adapt this, don’t translate it: rewrite the landing page below for small-business owners in Indonesia, in Indonesian. Replace the US examples with local ones, show prices in rupiah as [PRICE], and keep the tone friendly but professional."

4. Write instructions in whichever language you think most precisely in

The instructions and the output do not have to share a language. If you think most precisely in English but need the result in Tamil, write the instructions in English and name Tamil as the output. If you think best in Portuguese but need an English report, do the reverse. What matters is that both languages are named.

For translation tasks, name both languages, as Anthropic’s documentation does: "Translate the text from German to Korean. Respond with only the translation."

5. Show one example in the target language

Examples are among the most reliable ways to control tone and format, and that is doubly true across languages, where "formal" and "friendly" mean different things in different places. Paste one short piece written the way you want — your own message, a competitor’s page, a past newsletter — and ask for the same register and rhythm.

6. Check what you cannot read yourself

• Back-translate: ask a second chat to translate the result into a language you know, and compare it with what you meant.
• Ask the model to list anything it was unsure about — idioms, names, technical terms, formality choices.
• Have a fluent speaker review anything public, legal, medical or financial. A smooth sentence can still be wrong.
• Re-check numbers, dates and names after any translation or rewrite; they are the details most likely to drift.

A reusable multilingual prompt

Copy this, fill in the brackets, and delete what you do not need. The instructions can stay in English even when the output is not.

Task: [WHAT TO PRODUCE]. Output language: [LANGUAGE AND VARIETY, e.g. Latin American Spanish]. Reader: [WHO, WHERE THEY LIVE, WHAT THEY KNOW]. Register: [FORMAL / FRIENDLY; e.g. use usted / use tu]. Localise, don’t translate: use local currency ([CURRENCY]), date format, units and examples that fit [COUNTRY]. Write in native script, idiomatically, as a native speaker would. Example of the style I want: [PASTE A SHORT SAMPLE]. Afterwards, list any words or choices you were unsure about.

PromptVibe’s free generator writes the prompt in the language you type your goal in, so you can describe what you need in Spanish, Hindi or Arabic and get a structured prompt back in the same language.

Related articles