Mixing languages on purpose

Most multilingual people do not speak one language at a time. A sentence starts in Hindi and the technical noun is English. Tagalog with English verbs. Arabic with French administrative vocabulary. Yoruba with English numbers.

This is not sloppiness — it is how bilingual communities actually speak, with its own consistent grammar. And it is how you will naturally type into an assistant, so it is worth knowing what happens when you do.

Mixed input mostly works

Good news first. Models generally handle code-switched input well, because their training data contained plenty of it — the internet is full of people writing exactly this way.

You can type naturally. You usually do not need to “clean up” your prompt into one language before sending it.

The failure is the output language

The problem is not comprehension. It is that mixed input leaves the output language ambiguous, and an ambiguous choice defaults to English.

So say it:

Reply entirely in Hindi, Devanagari script.

Reply in Tagalog. Keep technical terms in English.

Reply in the same mix I used.

That third one is worth knowing about. If you write Hinglish and want Hinglish back — because that is how your colleagues actually communicate — ask for it explicitly. Models can do it, and will not choose it on their own.

Which terms stay untranslated

Every field has vocabulary that lives in English regardless of the surrounding language. Translating those makes text harder to read for the intended audience, not easier.

A doctor writing in Bengali does not want hypertension rendered into classical Bengali. A developer writing in Spanish does not want pull request translated. The English term is the term.

So supply the list:

Keep these in English: [terms]. Translate everything else.

This is the single most effective instruction in this lesson for professional writing, and it is the thing generic translation gets most reliably wrong.

Think in English, answer in yours

A genuinely useful technique that follows from L1.

If the model reasons better in English, you can use that without accepting an English answer:

Work through this problem in English first. Then give me only the final answer, in Marathi.

Or, when you want to see both:

Answer in English, then give the same answer in Marathi.

You get English-level reasoning quality with an output you can use. For anything analytical — comparisons, planning, working something out — this often beats asking in your language directly.

The caveat: what you get is a translation of English-shaped reasoning. For anything where local context is the point — local law, local convention, what is normal in your city — this is exactly the wrong technique, because the English reasoning may not know those things. Ask directly in your language there, and check the facts as L2’s test taught.

Scripts and romanisation

Many languages are written in a non-Latin script and typed in Latin characters: Hindi as kaise ho, Arabic as kayfa haluk, Greek in Latin letters.

Models generally understand romanised input. Two things to specify:

Which script you want back. Romanised input often produces romanised output, which may not be what you want to send anyone.

Which romanisation, if it matters. There are competing conventions for most languages, and consistency matters for names and searchable text.

Try it now (6 minutes)

Write a prompt the way you would actually type it, mixing languages if that is natural for you.

Add one line specifying the output language, the script, and which terms stay in English. Compare that against the same prompt without the line — the difference is usually larger than expected.

Check your understanding

1. Mixing languages in a prompt usually causes problems with: The model understanding you The output language, which defaults to English when unspecified The length of the response

2. The most effective instruction for professional writing in your language: Ask for simpler vocabulary List the terms that must stay in English Request a longer answer

3. 'Reason in English, answer in my language' is the wrong technique when: The task is analytical Local context is the point — local law, convention, what is normal where you are The answer needs to be long

Recap

Type the way you actually speak — mixed input is usually understood — but always specify the output language, the script, and which terms stay in English. Use “reason in English, answer in mine” for analytical work where the reasoning quality matters, and avoid it whenever local context is the substance of the question.