AI translates with tone, not just words

Machine translation used to be recognizable from the first sentence: stiff word order, literal phrases, lost nuance, and text that made sense but no native speaker would have written that way. Today's models handle this far better — but only when you tell them more than “translate.” The word “translate” by itself is an instruction for the most boring possible output: convert the meaning, nothing more.

Translation isn't a search for equivalents — it's a decision about how the sentence would sound if someone from the target culture had written it from scratch. And that's a decision someone has to hand the AI. Who's reading the text, what's your relationship to them, what do you want from them, how formal is your field in their country, how do you write yourself — every one of those details moves the result further than any dictionary would.

This guide is about handing over those details in practice. We'll go through why literal translation sounds stiff, how to steer tone with your own samples instead of adjectives, where translation ends and localization begins, how to write business correspondence in a language you don't fully command, how to check the result when you can't judge it yourself — and how to set the whole process up once so that next time it takes a minute. The prompts are ready to copy — just fill in the brackets.

A typical scenario

Laura freelances writing project documentation for construction firms, and she's sending a proposal to a German client for the first time. The text is finished in English, so she drops it into a translator and hits send. The reply comes back four days later — cold, one sentence, just a question about price. The job drags on and eventually goes to someone else.

What happened: the translation was grammatically correct and socially off. German business correspondence likes a formal, structured opening (“Sehr geehrte Damen und Herren,” “Wir erlauben uns, Ihnen unser Angebot vorzulegen”); Laura's English original was breezy, direct, and closed with a casual “let me know what you think.” The literal conversion read as underdressed for the occasion — like someone who wasn't quite sure the job was worth the effort. On top of that, it carried over details that meant nothing to the recipient: a domestic business registration number, standards specific to her home country, and a date written in a format that reads differently abroad.

The second time, Laura does it differently. She pastes three of her own German emails that had worked well in the past into the AI, describes the recipient and the industry, and says this is a first contact and what the next step should be. She gets back text that sounds like her — just in German. Then she has it translated back into English and compares the two: one sentence about the scope of work shifted from “includes” to “may include,” which would have cost her money. She fixes it and sends. The reply comes the next day, and it's specific.

The difference between the two attempts wasn't an hour of work — it was about ten minutes. The rest of this guide is about organizing those ten minutes so that next time they take two.

Phase 1: why literal translation sounds stiff

Before you start writing better prompts, it's worth knowing exactly what goes wrong in the worse ones. The mistakes come in three layers, and each one gets fixed differently.

Layer one: sentence rhythm and length

Every language has its own natural sentence length and pace. Many languages — German, French, Spanish — are perfectly comfortable with long compound sentences carrying several embedded clauses; English business writing cuts them into shorter units, one idea per sentence. Translate a long compound sentence one clause at a time, in the same order, and you get a sentence that's grammatically fine and reads like a legal document.

You can spot it by a suspicious pile-up of commas and conjunctions in the translated text. The fix is simple — just ask for it: “break up long compound sentences, one idea per sentence.”

Layer two: politeness register

This is where the most damage happens, because the mistake is invisible. The level of formality different cultures consider normal varies dramatically. German business correspondence is formal and structured, American is direct and friendly from the very first contact, British is polite but businesslike, and too much formality in it reads as standoffish.

A literal translation of a very deferential German opening line (“Hiermit erlaube ich mir, Ihnen unser Angebot vorzulegen”) into English produces a sentence that sounds like it's from the nineteenth century. Go the other way, and a breezy, direct American opening translated word-for-word into German can come across as presumptuous. Register doesn't get translated — it gets set.

Layer three: phrases with no counterpart

Idioms, set phrases, wordplay, and cultural references. The English idiom “kick the bucket” has no literal German equivalent — but German has its own idiom for the same idea, “den Löffel abgeben,” literally “hand over the spoon.” Models handle this well — if you let them. Ask for “translate” and they stick close to the original wording; ask them to “carry across the meaning, even at the cost of a different image,” and they'll swap the image for one that actually works.

A single test that catches all three at once

Before you send anything important, have the result critiqued — in the role of the recipient, not the translator.

I'm attaching a text I translated into [language] and I'm about
to send it to [who: a new client / a long-standing partner /
a government office].

[paste the translated text]

Don't evaluate grammar. Read it through the eyes of a native
speaker in [country] and answer this:
1. Can you tell it's a translation? Where exactly — list the
   sentences and say what gives it away
2. Is the level of formality right for this situation? Where
   is the text too formal, and where is it too familiar
3. Which phrasings would a native speaker never use
4. What would the recipient think about the writer after the
   first three sentences
5. Are there details in the text that mean nothing to the
   recipient

Don't rewrite anything — just list the findings with a citation
of the passage in question.

Point 4 is the uncomfortable, most useful one. The ban on rewriting is there on purpose: if you let the model fix the text directly, you get a different version without learning what was wrong — and you'll make the same mistake next time.

Phase 2: sample-driven translation

This is the core of the whole guide, and the technique that moves the translation the most. Describing tone in words is nearly impossible: “professional but friendly” means something different to every person, and to every model too. Showing tone is easy — paste two or three of your own texts in the target language and say “like this.”

Why samples beat adjectives

A sample carries information you'd never be able to describe: how long your sentences are, whether you use bullet points or continuous prose, how you open and close, whether you're formal or casual, how many pleasantries feel normal to you, whether you write in first person singular or plural. Three of your own emails carry more information about your voice than a paragraph of instructions.

Where do you get samples if you don't already write in the target language? Three options, ranked by quality: your own older texts that worked (even if AI helped write them and you edited them afterward); texts from a colleague or partner in the target country that you know are solid (with their permission); texts from companies you want to sound like in that market — their public offers, their website, their newsletters.

The basic sample-driven prompt

Translate this text into [language] for [description of the
recipient: who they are, where they're from, what field, my
relationship to them].

First, read these texts of mine in [language]. This is how
I write:

SAMPLE 1:
[paste your own text]

SAMPLE 2:
[paste your own text]

SAMPLE 3:
[paste your own text]

Text to translate:
[paste the text]

Rules:
- match the tone, sentence length, and formality of the samples,
  not of the original
- don't translate literally; write what I would have written if
  I'd written it directly in [language]
- use industry terminology from [field] the way it's actually
  used in [country]
- where you're not sure how to translate a term, leave it in
  parentheses in the original alongside your translation and
  flag it as TODO

At the end, add a short list of places where you departed from
the original, and why.

That last paragraph is the reason to use this prompt even when you speak the language yourself. The list of departures shows you where the model made a decision on your behalf — and that's exactly where a shift in meaning can hide without you noticing it in the result.

When you don't have samples: description instead of a model

Before you write “formal,” try describing specific, observable traits instead. A model responds far better to “no polite preamble, first sentence gets straight to the point, no exclamation points, no superlative adjectives” than it does to “professionally.”

Translate into [language]. I don't have a sample, so here's a
description instead:

Recipient: [who, country, position, how well we know each other]
Situation: [first contact / ongoing collaboration / complaint /
rejection / apology]
What the recipient should do after reading: [specific action]

Define the tone like this:
- form of address: [first name / last name with title / no
  direct address]
- sentence length: short, one idea per sentence
- pleasantries: [minimum / the usual amount for this culture]
- no superlatives, no exclamation points, no phrases like
  “don't hesitate to reach out”
- closing: a specific next step, not generic politeness

Text: [paste the text]

Give me two versions: one a bit more formal and one a bit more
relaxed, so I can compare and pick.

Two versions to compare is a trick worth using. In comparison, you'll recognize the right register instantly, whereas with a single version you have nothing to measure it against.

Your own voice, across languages

If you write in a foreign language regularly, it's worth doing one extra thing: having your own style described, and saving that description.

I'm attaching [5] of my texts in [your language] and [3] texts
in [language] that I wrote myself or was happy with.

[paste the texts]

Describe my writing style in a way detailed enough that someone
could write a new text from the description that sounds like
me. Focus on:
- typical sentence length and structure
- how I open and how I close
- level of formality and how it shows up concretely
- words and phrases I use repeatedly
- what I tend to avoid
- how I structure text (paragraphs, bullet points, headings)

Separately, note where my [your language] style and my [language]
style differ — and whether that difference is deliberate or a
loss in translation.

Output this as instructions for another model, not as an essay
about me.

Save the result — it's a text you'll paste into every future request, or store in persistent instructions. The last paragraph is usually the surprising one: most people unintentionally write more formally and more blandly in a foreign language than in their own, because they reach for safe, well-worn phrasing. Working with samples in general is covered in show AI a sample.

Phase 3: localization, not translation

Translation converts sentences. Localization converts the text so it makes sense in the target country — and that's a different job, one that reaches into content itself.

What gets localized

• Domestic institutions and terms. Sole-trader registration numbers, local business IDs, home-country certifications, references to a specific regulatory body — a recipient elsewhere won't know what they are. Either replace them with the local equivalent, add a parenthetical explanation, or drop them when they're not essential to the message.
• Formats. Dates (the order differs between British and American notation, and mixing up April 3rd and March 4th is a common mistake), numbers and decimal separators, units, phone formats, addresses.
• Measures and standards. Meters versus feet, kilograms versus pounds, clothing sizes, paper formats.
• Cultural references and humor. A joke built on pop culture from your home country doesn't translate — either replace it with a functional equivalent or cut it.
• Legal and business conventions. Payment terms, warranties, invoicing practices, data-protection language. This is where localization ends and a specialist's work begins.
• Directness. How “no” gets said, how “we won't make the deadline” gets said, how a complaint gets phrased. This varies the most, and it matters the most.

The localization prompt

This text was written for a [home country] audience. I want it
for [country], audience [who], purpose [sales page / proposal /
how-to guide / social media post].

[paste the text]

Don't translate it — localize it. Go through this process:

1. First, list everything in the text that's specific to
   [home country] and would be unclear or inappropriate for
   the target audience: institutions, abbreviations, standards,
   formats, cultural references, units of measure
2. For each item, propose a solution: a local equivalent, an
   explanation, or dropping it — and say why
3. Only then write the localized version
4. Finally, list what I should have decided myself that you
   decided for me

Anywhere this touches law, taxes, or warranties, don't replace
anything — just flag that it needs review by someone familiar
with local regulations.

Point 1 is valuable on its own: you'll often discover the text is unclear to a foreign reader for reasons that never would have occurred to you. Always check point 4 — localization is a series of small decisions, and some of them carry business consequences.

Where localization shouldn't reach

Legal text, terms and conditions, warranty clauses, anything with deadlines and penalties. A model can translate a contract in a way that reads smoothly and simultaneously means something different in the target jurisdiction. For binding documents, AI produces a working draft for a translator, not a translation. Use it to understand what's in the text and to build a list of questions — leave the binding wording itself to a specialist.

Phase 4: business correspondence in a foreign language

The most common real-world use. Short texts, high stakes, no time.

First contact

The first email decides whether there's a second one. The most common mistake is translating your home language's usual opening paragraph verbatim — two sentences introducing yourself, one apologizing for the interruption — habits that read as padding once they land in another language.

Write a first-contact email in [language] for [who: role,
company, country].

Context I know:
- how I got this contact: [referral / website / trade show /
  cold outreach]
- what I'm offering: [briefly]
- why this specific person: [a concrete reason, not a generic
  phrase]
- what I want them to do: [a short call / a reply to one
  question / review of a proposal]

Requirements:
- maximum [120] words, this gets read on a phone
- the first sentence says why I'm writing to them specifically
  — not who I am
- no superlatives about our company
- a specific, easy-to-act-on call to action at the end
- politeness at the level normal in [country] for a first
  contact in [field]

Write three variants with different levels of directness, and
for each, one sentence on who it suits and what the risk is.

Three variants with commentary beat one “best” version. The difference between them shows you the scale you're deciding on — and next time you'll already know it.

Bad-news messages

Rejections, delays, price increases, complaints. This is where culture plays the biggest role: what sounds like a reasonable explanation in one country reads as an excuse in another.

I need to write a message in [language] that carries bad news.

Situation: [what happened]
Recipient: [who, country, how long we've worked together]
Degree of my responsibility: [my mistake / circumstances /
their mistake]
What I want to preserve: [the relationship / the deadline /
the price / trust]
What I'm offering as a fix: [specifically]

Write the message the way someone in [country] who has handled
this situation a hundred times would write it:
- how much space to give the apology versus the fix, based on
  norms in that culture
- no excuses and no passive constructions that obscure who did
  what
- clearly state when and what happens next

Below the message, note what you did differently than would be
standard in [your language] business communication, and why.

The note below the message is there to help you learn: you'll typically find that other languages handle the apology in one clipped line and spend more space on the fix, while your first draft in your own language carried two paragraphs of explaining. And the rule that applies to everything still applies here: AI proposes, a human approves — the person who read the whole text is always the one who hits send.

Calls, meetings, and things said out loud

A written translation is one thing, speaking is another. Before a meeting in a foreign language, it's worth preparing not just the content but specific sentences for the awkward moments.

Tomorrow I have a [call / meeting] in [language] with [who,
country] about [topic]. My level is [describe it, e.g. I can
get by, but I'm not confident with technical terms].

Prepare me a cheat sheet:
1. 10 sentences I'll need, in both languages — opening, handing
   over the floor, disagreeing, asking someone to repeat
   something, closing
2. 15 technical terms from [field] with their translation and
   a rough phonetic spelling I can read at a glance
3. 5 polite phrases for buying myself time to think
4. 3 sentences for admitting I didn't understand without
   sounding incompetent
5. What to avoid in a meeting in [country]

Keep it to one page, so I can have it in front of me.

Point 4 addresses the most common source of trouble: people in a foreign language would rather nod along than admit they didn't understand — and end up agreeing to something other than what they thought.

Phase 5: checking your work when you can't judge the language yourself

A translation you can't evaluate is a risk. But there are three checks even someone with only partial command of the target language can run.

Back-translation

The simplest and most effective. You have the text translated back into your own language and compare it with the original. Shifts in meaning jump out immediately.

Translate this text back into [your language]. Translate it
literally and faithfully to what the text actually says — NOT
to what you think I meant to say. Don't improve it, don't
smooth it out.

[paste the translated text]

Then compare it with my original:

[paste your original text]

List every place the meaning shifted in a table: original
wording | back-translation | what the difference is | how
serious it is (business, legal, relational).

Check specifically: commitments (what I'm promising), conditions,
deadlines, numbers, degree of certainty (will / should / might).

That last paragraph is the most important one. The most dangerous shift in a translation isn't a dictionary mistake — it's a change in the degree of commitment — “we will deliver” becoming “we will try to deliver,” “includes” becoming “may include.” Both sound perfectly acceptable and mean completely different money. The “don't improve it” instruction is necessary — otherwise the model quietly fixes, during back-translation, something that was never wrong in the text.

A check from a second model

Have text produced by one tool reviewed by a different one. The model that wrote the translation judges it leniently — the same phenomenon as with texts you have peer-reviewed; see AI as a reviewer.

This text in [language] started out as a translation from
[your language]. The [your language] original is attached too.

[paste both]

You're an experienced proofreader who has handled text for a
company in [country]. List findings, don't rewrite:
1. Errors: grammar, prepositions, articles, collocations
2. Places where the text sounds like a translation even though
   it's grammatically correct
3. Terms used incorrectly or unusually for [field]
4. Meaning shifts against the [your language] original
5. Register: where the text is too formal for the situation,
   and where it's too familiar

For each finding, note how serious it is: blocks sending /
worth fixing / cosmetic.

Sorting by severity is practical — for an email that has to go out in ten minutes, you fix the first category and let the rest go.

A human, for anything that matters

The third check is boring and irreplaceable: for contracts, official communication, and anything that's going to be published, have a native speaker read the text. AI shortens their job from translating to proofreading, which is an order-of-magnitude difference in time and cost, but it doesn't replace them. This matters especially for marketing text, where the impression is the point, and for anything legal, where the consequence is.

Sensitive data and translation

Before you paste text in for translation, go through it the same way you would for any other request. A contract with names and account numbers, a client complaint with their details, an internal note about an employee — the translation itself usually doesn't need that data. Replace it with placeholder names (“Client A,” “amount X”) and add it back once the text is finished. Tone and meaning don't get lost in the process. Where you have no choice but to work with sensitive data, the site-wide rule applies: only in a paid account with a contractual data-protection agreement.

Phase 6: a workflow for translations you do again and again

When you're translating the same kind of thing over and over — proposals, invoices, product descriptions, answers to routine questions — there's no point reinventing the request every time. The goal of this phase is to shrink every future translation down to pasting in the text.

A Project as persistent context

Set up a Project (in Claude or ChatGPT) for each language pair and type of communication that comes up often. Your project instructions should include the style description from Phase 2, a description of the typical recipient, and rules that should never be broken. Your uploaded files should include sample texts and the glossary. From there, you just paste in the text and write “translate according to the project's rules.” How to build and maintain Projects is covered in projects and persistent context.

A glossary: terms that must never change

The most common complaint about machine translation on recurring text isn't the quality of any one sentence — it's inconsistency: the same term translated a different way every time. The fix is a glossary — a table that always translates the same way.

I'm attaching [10] of my texts in [your language] and their
existing translations into [language].

[paste the texts]

Build a glossary from them for consistent translation:

1. A table: [your language] term | approved translation |
   a note on when it's used | variants to avoid
2. Separately, list terms that were translated inconsistently
   across my texts — for each, list all the variants you found
   and recommend one
3. Separately, list names that should NOT be translated
   (product names, company names, legal entity types) — and how
   they should be inflected or written
4. Phrases I repeat that should have a fixed translation
   (greetings, closings, standard terms)

For point 2, don't decide on your own — flag it as a question
for me.

Point 2 typically turns up five to ten terms you've translated three different ways for years. Deciding which variant wins belongs to you, or to someone who knows the language and the field; the model's job is just to gather the variants and recommend one.

The everyday working prompt

Once your style is described and your glossary is built, a routine translation looks like this:

Translate into [language] per the project's rules.

Text type: [proposal / reply to an inquiry / product description
/ newsletter]
Recipient: [who]
What's different this time: [e.g. this is a repeat client, so
go a bit less formal; or: mention the delay right in the opening]

[paste the text]

Output:
1. The translation
2. Glossary terms you used
3. Terms not in the glossary that should be
4. Places where you weren't sure — with a question for me

Point 3 is what turns a one-off tool into a system: the glossary fills itself in with terms you actually run into in practice. Review and approve the additions once a month.

Templates for text that repeats verbatim

For text that only changes in the details — order confirmations, reminders, answers to a routine question — there's no point translating it every single time. Have a set of fill-in-the-blank templates built once, and just fill them in from then on.

Build a set of templates in [language] from these [your language]
texts of mine:

[paste 5-8 typical texts]

For each type:
- a template with [brackets] where information gets filled in
- a more formal variant and a less formal variant
- one sentence on when to use which
- what should NEVER be changed in the template (legally or
  commercially binding wording)

Write the templates so that once the brackets are filled in,
you can't tell it's a template — no generic filler.

Save the finished templates somewhere you can find in five seconds — text-expander snippets, drafts in your email client, or a prompt library as covered in a prompt library. A template sitting in a folder you never open gets rewritten from scratch anyway.

Common mistakes

• Asking for just “translate.” Without a recipient, a situation, and a purpose, you get the most neutral possible version — the one that sounds like a translation. One extra sentence of context changes the result more than picking a different tool would.
• Describing tone with adjectives instead of samples. “Professional but friendly” gets interpreted however the model likes. Two of your own texts in the target language solve that instantly.
• Not checking the degree of commitment. The most expensive translation mistakes aren't stylistic: “we will deliver” turns into “we will try to deliver” and goes unnoticed until something goes wrong. Back-translation catches this.
• Translating a legal text and relying on it. A readable translation of a contract is a great aid for understanding it and building a list of questions — it is not the binding wording. That's a specialist's job.
• Pasting in sensitive data the translation doesn't need. Replace names, account numbers, ID numbers, and internal data with placeholders; meaning and tone survive just fine without them.
• Letting the text get “improved” into phrasing you don't understand. A sentence you can't translate back into your own language is a foreign object sitting in an email with your name on it — and if someone asks about it, you won't have an answer.

The best tools

• Claude or ChatGPT with samples and context — the main tool in this whole process: they handle tone, register, localization, and can explain what they changed and why. Persistent rules and a glossary can live in a Project, so you never enter the context twice.
• DeepL — a specialized translator that keeps very natural output for many language pairs even without a long prompt; a good choice for a fast first pass on a longer text, which you then refine to match your own style.
• Google Translate — for getting the gist of a foreign text and simple, informal messages. Not for anything going out under your name.
• Dictionaries with usage examples — a check on whether a term is really used the way the model translated it in that field; the best collocation check is seeing it in ten real sentences.
• A native speaker or professional translator — for contracts, official documents, and published text. AI shortens their job from translating to proofreading; it doesn't replace them.

What you get out of it

• Time: the first translation with a proper prompt takes about as long as hunting for the right phrasing by hand, but every one after that is a matter of a minute. For someone writing in a foreign language several times a week, that's hours a month — and a glossary and templates cut it down further.
• Money: a caught shift in the degree of commitment is the one thing from this guide you can put a direct number on. One proposal that doesn't sound machine-translated, and one commitment corrected before it went out, usually pays for the entire time invested in setting this up.
• Peace of mind: you know what you're sending, even in a language you don't fully command — because back-translation showed you what the text actually says.
• Quality: consistent terminology across every text and a stable tone, even when several people on a team are writing them. The glossary doubles as documentation you can hand to a new colleague.

Pro tip

The fastest route to a translation that doesn't sound like one is a detour: don't translate, retell. Instead of “translate this text,” give the model “here's what I want to say, to whom, and why — write it in [language] the way someone from there would write it,” and attach your original text as reference material, not as a template to follow. The model then doesn't tie the result to your original sentence structure and doesn't have to untangle itself from it — it writes directly in the logic of the target language. On shorter business text, the difference is audible immediately.

And the closing rule: you check the meaning, not the words. A beautifully worded translation that quietly shifted one commitment is worse than a stiff translation that says exactly what you meant. A back-translation before you hit send costs two minutes, and it's the cheapest insurance you have in this whole area.