Prompts & Writing

Prompting for Multilingual Work Without Losing Nuance

By Jim Vernon, Editor, AI Intelligence International · Published 31 January 2026 · Reviewed against our editorial standards · About the author

Machine translation is strong enough for most business communication and still fails in the places that matter: register, legal precision, humour, and terms of art that differ by market.

The reliable improvements come from supplying context rather than from switching tools or language pairs.

Key takeaways

  • Supply the register explicitly: Languages with formality distinctions will default unpredictably.
  • Give it your glossary: Every organisation has terms that must be translated a specific way — product names, legal entities, regulatory categories, internal jargon.
  • Translate meaning, then check direction: For important text, ask for the translation, then in a fresh context ask for a back-translation into the original language and compare.
  • Localisation is not translation: Currencies, date formats, address structures, examples, holidays, legal disclaimers and payment methods all need replacing rather than translating.

Supply the register explicitly

Languages with formality distinctions will default unpredictably. State the relationship: writing to a long-standing supplier, to a government office, to a customer who complained, to a colleague you know well.

Register errors are the most common cause of translated business communication landing badly, and they are entirely preventable.

For languages with honorific systems, name the specific level you want rather than saying 'polite'.

Give it your glossary

Every organisation has terms that must be translated a specific way — product names, legal entities, regulatory categories, internal jargon. Paste the glossary into the prompt or the system prompt.

Without it, output is inconsistent across documents, which looks careless to customers and creates real problems in contracts.

Maintain the glossary as a shared file with the source term, the approved translation, and a note on what not to use.

Translate meaning, then check direction

For important text, ask for the translation, then in a fresh context ask for a back-translation into the original language and compare. Divergences reveal where meaning shifted.

This is cheap and catches the majority of substantive errors, especially negations, conditionals and quantities.

It does not catch cultural inappropriateness, which needs a native reader.

Localisation is not translation

Currencies, date formats, address structures, examples, holidays, legal disclaimers and payment methods all need replacing rather than translating.

Ask explicitly for localisation with the target market named, and list the elements to adapt. A translated page with foreign payment methods converts badly regardless of language quality.

Have a person from the market review anything customer-facing before launch, even briefly.

Learning and practice use

For language learning, the model is strongest as a patient conversation partner with correction, and weakest as an authority on nuance and current slang.

Set the scenario, the level and the correction style explicitly: correct after every message, or only at the end, or only grammar. Uncorrected practice reinforces errors.

Ask for corrections with explanations in your stronger language, which is the pattern learners retain best.

Where to keep a human

Contracts, regulatory filings, medical or safety instructions, and marketing where tone carries the brand. In each case the cost of a subtle error exceeds the translation saving by orders of magnitude.

Use the model to prepare the draft and the glossary compliance check, and pay a professional to review. That combination is usually cheaper than full human translation and better than either alone.

Register and formality are the usual failure

Most models default to a neutral, slightly formal register that reads as translated in languages with strong formality distinctions. Specify the level explicitly — the polite form in Japanese, tu or vous in French, the appropriate honorific level in Korean — and give one example line.

In languages where code-switching is normal in business, say so. A Tagalog business email that refuses to use any English terms reads stranger than one that uses the terms everyone actually uses.

State the audience relationship rather than a formality label where you can: writing to a long-standing supplier is a clearer instruction than 'semi-formal'.

Regional variety and local conventions

Name the variety: Brazilian rather than European Portuguese, Simplified for mainland China, Latin American rather than peninsular Spanish. Unspecified, output drifts toward whichever variety dominated the training data.

Conventions travel with the language. Date formats, currency symbols and placement, decimal separators, address order, name order and honorifics all differ, and getting them wrong marks the text as foreign faster than any grammatical slip.

For anything customer-facing, have a native speaker read the first ten outputs. Ten is enough to catch the systematic errors, after which the prompt can be corrected once rather than every time.

Translation versus writing natively

Translating an English draft preserves English sentence rhythm and idiom even when every word is correct. For short, high-visibility text — headlines, taglines, error messages — brief the model in the target language and let it compose natively.

For long documentation where consistency matters more than elegance, translation is the right approach, with a glossary of terms that must not vary.

Keep that glossary in the prompt. Terminology drift across a long document is the complaint reviewers raise most often, and it is entirely preventable.

Frequently asked questions

Which languages are weakest?

Lower-resource languages and regional variants remain noticeably weaker, especially in idiom and formal register. Test with your own material before committing.

Should I write in English and translate?

For most business use yes, provided you write plainly. Idiomatic English translates poorly, so simplify the source first.

Is back-translation reliable?

It reliably reveals meaning shifts and unreliably reveals tone problems. Use it as a screen, not as approval.

How do I keep terminology consistent at scale?

A shared glossary in the system prompt plus a mechanical check that approved terms appear. Consistency is a process problem, not a model problem.

Are models equally good in every language?

No. Quality tracks training data volume, so widely written languages get noticeably better output than lower-resource ones. Increase review accordingly.

Should the prompt be in the target language?

Often yes for creative or idiomatic work. For complex instructions, a clear English prompt requesting output in the target language is usually more reliable.

Tools mentioned in this article

More in Prompts & Writing

← All articles