Working across languages

The three separate questions

People treat multilingual work as one topic. It is three, and they have different answers.

1. Which language should I write my prompt in?
2. How good is the model in my language?
3. What is translation actually doing here?

Which language to prompt in

Model quality is uneven across languages, and English generally has the best coverage because it dominates the training data. That leads to common advice: prompt in English, ask for output in your language.

That advice is half right and it costs you something.

It is right that instructions in English are followed slightly more reliably by most models, and that reasoning-heavy tasks sometimes come out better. It is wrong as a blanket rule, because the material you are working with is usually in your language, and translating it into English to prompt, then back, introduces two more places to lose meaning. If the nuance you care about lives in the source text (tone in a customer complaint, a legal term with no English equivalent, a piece of family correspondence) keep it in the original.

The workable middle, and what most multilingual professionals settle on:

Instructions in whichever language you write most precisely. Material in its original language. Output in the language of the reader.

There is nothing wrong with a prompt that mixes all three. Models handle a mixed prompt perfectly well, and it is what most bilingual people do naturally once they stop worrying about it.

What quality actually looks like outside English

Be concrete rather than gloomy. Broadly:

• Major European languages, Mandarin, Japanese, Korean, Arabic, Hindi: strong, close to English for most everyday tasks.
• Widely spoken but less represented languages: Bengali, Urdu, Vietnamese, Swahili, Tamil, Telugu, Marathi and many others: usable and noticeably weaker, especially on idiom, formal register and anything culturally specific.
• Low-resource languages: often poor, and the failure mode is fluent-and-wrong rather than obviously broken, which is worse.
• Code-mixed registers: Hinglish, Taglish, Singlish, Arabizi: handled surprisingly well for casual text, unevenly for anything formal.

Two costs come with all of the above. Non-English text uses more tokens per sentence, so you hit limits sooner and pay more. And quality degrades faster in long outputs than it does in English.

Translation

Machine translation between well-resourced languages is good and has been for a while. What it still does badly:

Register. The difference between the formal and informal second person, between an office register and a family one, between polite and deferential. Say which you want: “translate into Hindi using the formal आप throughout.”

Terms of art. Legal, medical and technical terms often have an established local equivalent that a general translation will miss. If you know the term, give it.

Things with no equivalent. Kinship terms, food, administrative categories, honorifics. The useful instruction: “where there is no good equivalent, keep the original word and add a three-word gloss in brackets.”

Names and numbers. Check them by eye, every time. Transliteration of names is inconsistent, and number formats differ: the Indian lakh and crore system, decimal commas in much of Europe, different digit grouping.

The back-translation check, which costs nothing

Take the translated output, start a fresh conversation, and ask for it to be translated back. Compare with your original.

This does not prove correctness (a symmetric error survives it), but it catches dropped clauses, reversed negations and shifted register, and those are the errors that matter. It takes a minute and it is the only free verification most people have for a language they do not read.

For anything with consequences (a contract, a medical instruction, a public notice) a human speaker checks it. No prompt substitutes for that, and the cost of the error is not symmetrical with the cost of the check.

One thing worth doing today

Write your standing brief in the language you work in, and include a line about register: which form of address, how formal, whether English technical terms stay in English or get translated.

Most bilingual professionals leave English terms in place (“invoice”, “deadline”, “server”) because that is how their colleagues actually speak. A model with no instruction will translate them all, producing text that is technically correct and reads like a government form.