AI Translation Prompts: 8 Templates to Translate and Localize Your Content
A good AI translation prompt gives the tool three things: the target audience, a glossary of terms that must not change, and the expected tone. Without that framing, ChatGPT or Claude produce text that reads smoothly but sounds generic. With it, you get a usable first draft that a reviewer signs off before publication.
This article is for SME owners and startup founders who sell, hire or communicate in several languages without an in-house translation team: a product sheet to put into English, a web page to adapt for Belgium or Quebec, a customer reply to write in German. You will find a preparation method, eight ready-to-use prompts, a review routine and the legal rules to keep in mind. For the general principles of writing prompts, the prompt engineering guide for SMEs is the starting point of this series.
Translation vs localization: what AI does well, and less well
Language models translate the general meaning of everyday text quite accurately; their mistakes come mostly from industry terminology, tone and cultural references. That is exactly where translation and localization part ways.
Translating means conveying the same message in another language. Localizing means adapting that message to a market: units, date formats, currency, examples, level of formality, legal notices. "Free delivery from €50" does not carry over word for word for a US customer: you need to decide whether to keep euros, show a conversion, and whether the threshold still makes sense in that market. No LLM will make that business decision for you.
What works well with a general-purpose assistant:
- everyday text (emails, articles, FAQs, descriptions) in widely spoken languages;
- rewriting in a given tone, more direct or more formal;
- consistency across a long document, provided you supply the glossary;
- explaining translation choices, which a classic machine translation tool does not do.
What always needs a human check:
- technical, legal or product-specific terms;
- product names, slogans and wordplay;
- figures, dates and units, which the tool may convert without saying so;
- less widely represented languages, where quality drops noticeably.
The main risk is not the unreadable sentence, which is easy to spot. It is the perfectly fluent sentence that says something different from the original. We explain this mechanism in our article on AI hallucinations.
Before the first prompt: glossary and style sheet
Twenty minutes of preparation saves most of the corrections: a glossary of 15 to 30 terms and a half-page style sheet are enough for an SME. Both documents are reused for every translation and pasted at the top of the prompt.
The glossary lists terms with their required translation, or the note "do not translate". Example for a software publisher translating from French:
| French term | Required English | Note |
|---|---|---|
| Espace client | Customer portal | Never "client space" |
| Devis | Quote | "Estimate" only for non-binding figures |
| Tableau de bord | Dashboard | |
| Offer name "Pilote" | Pilote | Brand name, do not translate |
| Organisme de formation | Training provider | Add "certified" when Qualiopi is mentioned |
The style sheet sets out the audience (executives, technicians, general public), the level of formality (formal "vous", neutral "you", "Sie" in German), format conventions (dates, decimal separator, currency) and two or three sample sentences that have already been approved. Those samples make a big difference to consistency: it is few-shot prompting applied to translation.
If you translate often, save these rules as standing instructions (a Claude project, a custom GPT or workspace instructions). What belongs in standing instructions versus the one-off request is explained in our article on system prompts and user prompts.
The 8 translation and localization prompts
Each template follows the same logic: context, audience, constraints, then output format. Replace the items in square brackets. The tool translates into whichever language you specify. To understand why this structure works, see the anatomy of a good prompt.
1. Framed translation of everyday text
You are translating for [company], which sells [offer] to [target audience] in [country]. Translate the text below from [source language] into [language and variant: British English, European Spanish...]. Tone: [formal / neutral / direct]. Keep the structure (headings, lists, paragraphs). Do not translate: [brand names, product names]. If a sentence is ambiguous in the source, translate it as closely as possible and flag it after the text, with your question. Text: [paste the text]
The last instruction is the most useful one: it turns the tool's uncertainties into visible questions instead of silent choices.
2. Translation with a mandatory glossary
Here is our glossary (source term, required target term, note): [paste the table]. Translate the following document into [language], applying this glossary strictly. After the translation, list every glossary term found in the text and how you rendered it, then flag any technical term not in the glossary for which you had to choose a translation. Document: [paste]
The control list at the end speeds up review: you check the list instead of rereading every sentence hunting for terms.
3. Localizing marketing copy
Here is promotional copy written for the French market: [paste]. Adapt it for [target market] rather than translating it word for word. Keep the key message: [message in one sentence] and the call to action: [action]. Adapt: examples, cultural references, formats (dates, currency, units), the level of formality usual in that market. Present the result in two columns: adapted text, and an explanation of each adaptation compared with the original.
The explanations are there to help you decide, not to be approved blindly. If the tool swaps a French example for a local reference you do not know, ask for a more neutral option. The same tone principles apply to social posts, covered in our LinkedIn content prompts.
4. Web page or product sheet with length limits
Translate this product sheet into [language] for an online store aimed at [country]. Constraints: title of 60 characters maximum, short description of 150 characters maximum, long description of similar length to the original. Target search phrase in the target language: [phrase you have validated]. Work it naturally into the title and the first sentence, without forced repetition. Convert measurements into [system], keeping the original value in brackets. Sheet: [paste]
Do not let the tool pick the keyword in the target language on its own: it often suggests the literal translation, which is not necessarily what people search for. Check the phrase with a keyword research tool or a speaker from that market.
5. Replying to a customer in their language
A customer wrote in [language]: [paste the message]. First, summarize in [your language] what they are asking, in three lines maximum, and flag anything ambiguous. Then write a reply in [language] that says: [reply points in your language]. Tone: [polite and direct]. Length: [number] words maximum. Below the reply, add its translation into [your language] so I can check what I am sending.
This prompt avoids the classic trap of answering what you think you understood. The initial summary and the check translation take ten seconds to read. For prospecting emails in another language, our sales email prompts work with the same language instruction.
6. Working translation of a legal or HR document
Translate this [contract / policy / procedure] from [language] into [your language] for internal understanding. This is a working translation, not a binding version. Translate as closely as possible, without rephrasing or simplifying. Keep the original term in square brackets next to each legal concept ([example: "consideration"]). After the text, list the clauses whose meaning depends on the governing law and that should be checked by a lawyer.
For a contract or any binding document, AI translation helps you understand it and prepare the conversation with your adviser. It is not something to sign on. The legal section below explains why.
7. Critical review of a translation
Here is a source text in [language] and its translation into [language]: [paste both]. Act as a professional reviewer. Compare sentence by sentence and list in a table: omissions, additions, mistranslations, errors in figures or units, tone shifts, terms that do not match the glossary [paste or "none"]. For each issue, quote the sentence concerned and suggest a correction. If you find no issue in a category, write "nothing to report"; do not invent errors.
Run this review in a new conversation, ideally with a different tool from the one that translated: a model rereading its own work in the same thread tends to approve it. Asking the tool to walk through its reasoning sentence by sentence, as in chain of thought prompting, makes discrepancies easier to see.
8. Back-translation to check a language you cannot read
Here is a text in [language]: [paste the translation]. Translate it literally into [your language], without improving it or fixing awkward phrasing. Then compare it with this original text: [paste] and list the differences in meaning.
Back-translation does not guarantee style, but it catches mistranslations, altered figures and promises whose scope has changed ("up to 30%" turned into "30%"). It is the minimum check when nobody in-house reads the target language.
ChatGPT, Claude or a dedicated tool: which one for which text
For text that needs a tone adjustment or an explanation of choices, a general-purpose assistant such as ChatGPT or Claude is more flexible; for large, repetitive volumes with stable terminology, a dedicated machine translation tool such as DeepL is often faster. The right choice depends on the text, not on a general ranking.
| Need | General-purpose assistant (ChatGPT, Claude) | Dedicated translation tool (DeepL) |
|---|---|---|
| Email, post, web page to adapt | Good fit: tone, context, explanations | Fine, needs tone edits |
| Marketing localization | Good fit, with review | Weak on its own: translates more than it adapts |
| Large repetitive volume (catalog, documentation) | Possible, slower, risk of length drift | Good fit, especially with a glossary feature |
| Explaining a translation choice | Yes | No |
| Critical review and back-translation | Yes (prompts 7 and 8) | Not designed for it |
Differences between assistants (accepted text length, file handling, data settings) are covered in our ChatGPT vs Claude vs Gemini comparison. On a long document, keep the token limit in mind: beyond a certain size, translate in coherent sections and paste the glossary each time, rather than sending fifty pages at once.
Checking an AI translation before publishing
The simple rule: the more a text commits your company, the more the review must be done by a human who is competent in the target language. A blog post and a safety notice do not call for the same level of control.
A practical three-level grid:
- Internal use (understanding an email or a document you received): prompt 1 or 6 is enough, with a critical read on your part.
- Low-stakes publication (article, post, routine customer reply): framed translation, then critical review (prompt 7) and back-translation (prompt 8) if nobody reads the language.
- Binding text (terms of sale, contract, user manual, safety sheet, pricing page): the AI translation is a first draft, and a translator or lawyer who knows the language approves the final version.
Three checks to run every time, in under five minutes: all figures, dates, prices and units are identical or correctly converted; brand and product names are intact; no commitment has been added or removed (warranty, deadline, refund).
What the law says before you publish a translation
Using AI to translate is not prohibited, but some documents must exist in French when you operate in France, and the company remains responsible for what it publishes, whatever the tool. Three texts come up most often for an SME selling or employing in France.
French Law No. 94-665 of 4 August 1994, known as the Toubon Law, requires French in the designation, offer, presentation and instructions for use of goods and services, in invoices, and in written, spoken or audiovisual advertising. In practice, if you translate a manual or product sheet written by a foreign supplier into French, the published French version must be accurate: a rough translation of user instructions is a compliance issue, not just an image problem.
The French Labour Code also requires French for written employment contracts (Article L1221-3), internal rules and documents that set obligations for employees or that they need to do their job (Article L1321-6), with exceptions, notably for documents received from abroad. If a document from a foreign parent company arrives in English, AI translation can help prepare the French version, but the version circulated to employees must be carefully reviewed.
Finally, the EU AI Act (Regulation (EU) 2024/1689) sets transparency obligations in Article 50. For text published to inform the public on matters of public interest, the deployer must disclose that the text was generated or manipulated by AI, unless the content has undergone human review or editorial control and a person holds editorial responsibility for the publication. A translation reviewed and owned by the company generally falls within that exception, which is one more reason to formalize your review step. For other uses, check your situation with your adviser: the European Commission continues to clarify how Article 50 applies.
Data and confidentiality: what you paste into the tool
Text you want translated is still company data: contracts, customer emails or HR files should only go through a properly configured business account, in line with your AI policy. Translation is one of the uses where people most readily paste whole documents without thinking.
A few simple habits:
- remove or replace names, addresses and numbers when the translation does not need them;
- check in your plan settings whether your conversations can be used to train models;
- for documents containing personal data, the GDPR applies as it does to any other processing;
- avoid unknown free translation extensions and websites for sensitive documents.
These rules are detailed in our article on data confidentiality with AI.
Frequently asked questions
What is the best prompt for translating with ChatGPT?
There is no single formula, but a good translation prompt specifies the language and variant, the audience, the tone, the terms not to translate, and asks the tool to flag ambiguities. Template 1 in this article is a reusable starting point for most everyday text.
Does ChatGPT translate better than DeepL?
It depends on the text. A general-purpose assistant is more flexible for adapting tone or explaining its choices; a dedicated tool is often faster and more consistent on large volumes with stable terminology. Test both on one of your real texts before choosing.
What is the difference between translation and localization?
Translation conveys the same message in another language. Localization also adapts examples, formats (dates, currency, units), level of formality and cultural references to the target market. Marketing copy almost always needs localization.
Can you publish an AI translation without review?
For internal use, yes, with caution. For published text, a review is strongly recommended, and it is a must for text that commits the company: terms of sale, contracts, manuals, documents for employees. The company remains responsible for what it publishes.
How can I check a translation in a language I do not speak?
Combine two checks: a critical review in a new conversation (template 7) and a literal back-translation into your language (template 8). Then check figures, names and commitments yourself. For binding text, have a competent speaker approve it.
Do you have to disclose that a text was translated by AI?
Article 50 of the EU AI Act requires disclosure for certain text published to inform the public, with an exception when the content has been reviewed and a person holds editorial responsibility for it. For a reviewed product sheet or sales page, the question arises differently; check your specific case with your adviser.
How GrowthPerf can help
Translating with AI depends less on the tool than on the method: a glossary, a style sheet, framed prompts and a review routine matched to what is at stake. GrowthPerf, a Qualiopi-certified training provider, trains leaders and teams at SMEs, startups and non-profits in these practices, using their own content.
The Operational AI course includes workshops on your documents: building the glossary, a library of translation prompts, a review grid. For a programme focused on prompting techniques, see our page on prompt engineering training. Depending on your entitlements, these courses can be funded by your OPCO (French training fund).
For the full set of prompt writing techniques, go back to the complete prompt engineering guide for SMEs, and to the glossary for the basics.
Need to translate a website, a catalog or customer documents? Book a free 30-minute audit: we review one of your documents with you and define the right glossary and level of review.