Module 8 of 10~35 minutes

Multilingual Prompting

Language, audience and local convention are context you state deliberately — and fluency arrives long before accuracy.

Objectives

State the output language explicitly rather than relying on the language of the prompt.
Describe an audience specifically, without substituting a country for a description.
Explain the difference between translating and localizing.
Identify the conventions — units, dates, currency, formality, terminology — that a task depends on.
Work deliberately when input and output languages differ.
Explain why fluent output can be less reliable than it appears.
Name what must be checked by a person who actually reads the target language.
Decide when a translation carries enough risk to require professional review.

Introduction

Every earlier module quietly assumed one thing: that you and your reader share a language and a set of conventions. For most of the world, and for most of the people taking this course, that is not true.

This module removes the assumption. It teaches language, audience and locale as things you state deliberately rather than leave to be guessed — and it teaches the failure mode that makes multilingual work genuinely risky.

Fluency arrives before accuracy. An assistant can produce text that reads beautifully in a language while being wrong about the facts, the conventions, or the reader.

Lesson 1

Say Which Language, and For Whom

Module 3 said context is what the assistant needs and cannot infer. Language is exactly that. Say the output language explicitly — do not assume it follows the language you wrote the prompt in. You may write in one language and need output in another, and leaving it implied invites the wrong guess.

Describe the audience; do not name a country. "Write for people in country X" is too broad to act on and invites a stereotype in place of a description. What actually helps is who they are, what they already know, why they are reading, and what they need to do next.

Weak: "Write this for a French audience." Better: "Write this in French, for parents of children starting school this year, most of whom have not used this enrollment system before. They need to know what to bring and by when." The second says nothing about nationality and everything about the reader.

Lesson 2

Translation Is Not Localization

Translation carries the meaning across. Localization makes the result work for the reader — conventions, examples, tone, and the things that would simply be done differently there.

A translated notice can be perfectly accurate and still land badly because it keeps the original's date format, an unfamiliar unit, a holiday that does not apply, an example nobody recognizes, or a level of formality that reads as rude or as distant.

Ask which one the task needs. A quoted policy may need faithful translation. A message to members almost always needs localization.

Lesson 3

The Conventions That Carry Meaning

Where quiet errors live — each is worth stating as context when it matters.

Units — distance, weight, temperature. A converted number is a claim.
Dates — day/month order changes meaning silently and is rarely flagged.
Currency — symbol, placement, decimal and thousands separators.
Names — order, honorifics and forms of address vary widely.
Formality — the choice between formal and familiar address exists in many languages and carries real social weight.
Terminology — a domain term may have an established local equivalent, or may be left as-is.

Names and proper nouns deserve their own check. They are the thing most often mangled and least often noticed, because they look plausible either way.

Lesson 4

Working Across Two Languages

Three ordinary situations, each handled by stating the intent.

Source in one language, output in another. Say both: "The notes below are in Portuguese. Write the summary in English."
Think in your strongest language. You may write the prompt in whichever language you reason most precisely in, and request output in another. The reasoning taught in this course does not change with language.
Mixed-language material. Say what to do with it — translate everything, keep quoted terms in the original, or preserve names as written.

The rule underneath all three: an unstated language decision becomes a guess.

Lesson 5

Fluency Is Not Accuracy

This is the module's central warning, and it is genuinely hard to act on. An assistant's output can be grammatically excellent and idiomatically smooth in a language while being wrong about a fact, a convention, or the reader.

Worse, fluency actively suppresses your scepticism: text that reads well feels checked, and in a language you do not read confidently, you have no signal at all.

The practical consequence: you cannot judge output in a language you do not read. Not "should not" — cannot. Reading it back through a machine translation checks your understanding of it, not its correctness.

Where the stakes are low — an informal note, a draft to be reviewed anyway — this is a manageable risk. Where they are not, it is the reason the next lesson exists.

Lesson 6

What a Person Must Still Check

Someone who genuinely reads the target language should check, at minimum:

Names and proper nouns, exactly as they should appear.
Numbers, dates and units against the source.
Anything with legal, financial, medical or safety consequence — always, without exception.
Tone and formality, which are invisible to a non-speaker and immediately obvious to a reader.
Anything that will be published or sent to many people.

Involve a professional translator for contracts and legal text, medical or safety instructions, regulated communications, formal notices to the public, and anything where being misunderstood carries real cost. An assistant can produce a strong first draft for a professional to work from. That is a legitimate role, and it is not a substitute for one.

Safety Across Languages

Everything from Module 7 applies unchanged when you work across languages, and one risk gets sharper: material you cannot read is material you cannot screen. A pasted document in an unfamiliar language may contain personal information you would have removed had you recognized it — or an instruction addressed to the assistant that you could not have spotted.

Minimize before translating, not after. And treat a translated result about health, money, law or safety exactly as Module 7 requires: verified by a qualified person before anyone acts on it.

Human + AI

You know who the reader is; the assistant does not. It can produce fluent text in many languages, and it cannot tell you that a date will be misread, that a form of address will offend, or that a name has quietly become a different word. Those are judgments made by people who know the reader and the language.

Fluent does not mean correct, and translation is not a solved problem.
It will not reliably keep your formatting or your conventions.
Re-translating back does not prove accuracy — it checks your understanding, not the text.
You cannot judge output in a language you do not read.

Practical exercise

≈7 min

One message, two readers. A short practical notice — a change of date and place for a community meeting — must reach two different groups of readers who do not share a first language.

1. Write the audience description for each group: who they are, what they already know, what they must do. No country names as a substitute for a description. 2. State the language decision explicitly for each: source language, output language, and what to do with any names or quoted terms. 3. List the conventions the notice depends on — the date, the time, any address format, the level of formality — and mark which must be localized rather than translated.

4. Write the two prompts: same underlying facts, two genuinely different requests. 5. Write the check list — what a reader of each language would need to confirm before this is sent, and which single item you would not send without a native check.

No knowledge of any particular language is required, and no AI account is needed: the reasoning is about decisions, not vocabulary. Work it against three sample versions of the same notice — one that is accurate but keeps the source's ambiguous date format, one that reads beautifully but has "translated" the venue name so the address no longer identifies a real place, and one that is correctly localized with a single formality choice flagged for a native reader. The second is the module's lesson in one artifact: the best-reading version is the broken one. Most learners find step 3 is where the real work is — the facts translate easily and the conventions are what actually break. Nothing is submitted, stored or graded.

Key takeaways

Say the output language explicitly — an unstated language decision becomes a guess.
Describe the audience: who they are, what they know, why they are reading, what they must do next. Never a country instead.
Translation carries meaning; localization makes it work for that reader.
Units, dates, currency, names, formality and terminology are where quiet errors live.
Fluency is not accuracy, and you cannot judge output in a language you do not read.
A native reader checks names, numbers, dates, tone — and anything with legal, financial, medical or safety consequence.
Material you cannot read is material you cannot screen: minimize before translating, not after.

Practice in Prompt Lab

Optional

Want to try what you learned with real prompts? Prompt Lab is an optional practice environment, separate from this course.

Try adapting one prompt for a different language, audience or locale, and note what you would be unable to verify yourself.

Practice Prompting in the Real World

Opens in a new tab. Optional practice — never required for this module, the checkpoint, your progress, the Final Assessment, the certificate or Reward Points.

Before you move on

Every skill in this course now exists separately: forming a request, structuring it, supplying context, reusing what works, judging the result, understanding where fetched material comes from, handling it safely, and adapting it for real readers.

Separately is the problem. Real work uses several at once, and the only way to know whether you can do that is to build something that has to hold together.

Putting it together — a small set of patterns you would genuinely use — is Module 9.