AI Prompts for Translation and Localization: The Words Are the Smallest Part

A page comes back from translation flawless and still does not work: the date reads as a different month, the price is in the wrong country's format, the keyword you built it around does not exist there. AI prompts for translation are mostly a checklist of what the words do not carry.

None of those failures is a translation error: they are decisions already taken in your own language, which is why they never show up in the text you sent.

A Translation Moves Words; the Market Re-Decides Four Things

What travels in a translation is real: the facts, the argument and the instructions. What does not travel are the decisions settled before you wrote a word - the formats, the query, the container and the register. Each is decided with the target market in front of you, and none is visible when you read the original.

The W3C splits the work in two: internationalization means building something so it can be adapted, localization means adapting it. Three of the four decisions are localization. The fourth, the room the message needs, can only be fixed while the layout is still being designed. A correct translation can therefore be locally wrong without containing a single error.

The Test

Read your own page looking for the four things that are not written on it. What date will a stranger read? What would they type to find this? How much room does the message need? How is the person addressed? If all four answers are "the same as always", you translated; you did not localize.

The Formats Are Invisible Until They Are Wrong

03/04/2026 is April 3 in some markets and March 4 in others, and the ambiguity survives every machine translation. Decimals behave the same way: 1.00 and 1,00 are one number in two places and different numbers in others. Currency adds a code or a symbol, before or after the figure, and a dollar sign shared by several countries - disambiguate with the code. Units are worse, because some are not conversions at all: a shoe size from one system is a different scale in another. And names and addresses come in a different order.

Every item is a field of your example rather than a rule of grammar, so the list can be reviewed with a pencil in five minutes. The standard behind it is Unicode CLDR, the data set that defines number patterns, separators and plural rules per locale.

The Query Is Not a Translation

The phrase that brings traffic in your language is a query, not a term. In the other market people type different words, sometimes shorter, sometimes a phrase that is not a translation of yours. Translate your keyword and you optimize a page for a query nobody writes.

That makes the title and the meta description new strings rather than moved strings, written with the local query in front - which is why a structure prompt asks for a keyword field instead of accepting a finished text, and why you run it once per market.

Search engines say the same. Google's documentation on localized versions of a page treats them as duplicates only when the main content remains untranslated, and gives three equivalent ways to declare the alternatives: HTML, HTTP headers and sitemap.

The Container Does Not Grow With the Language

The same message usually needs more room in the target language, and the expansion is relative: the shorter the string, the worse it behaves. A six-letter button can need three times the box; a long paragraph stretches a little. The W3C's article on text size in translation reproduces the average expansion rates IBM published in its guidelines for global solutions design: strings up to ten characters run 200% to 300% longer, falling to about 130% past seventy. Its own example is the word "views", which becomes "visualizzazioni" in Italian, three times the length.

Where it bites is predictable: buttons and labels in fixed boxes, table headers, menus, subtitles. Subtitle limits are counted in characters per second and set per language, which is why Netflix publishes a timed text style guide with language-specific variants. Dubbed voice-over runs longer and the edit gives way; slides and printed pieces change page count; and text baked into an image is not translated at all, it is regenerated - which is why the image prompts in the catalog keep text out of the frame.

The Pre-Check: Pad It Before You Translate

The standard check is pseudo-localization: replace the strings with longer simulated text before translating, so you find what breaks while it is cheap to fix. A manual version needs only the mockup - stretch every label and see whether the box holds. If the button survives double its size, the paragraph will be fine.

The Register Belongs to the Market, Not Your Brand Book

Some languages have two forms of "you", and the choice is not a style preference. Address people formally where they expect familiarity and the page sounds distant; do the opposite and it sounds disrespectful. The market makes that call, not your brand book.

A voice guide is an artifact of one language: its pillars, its tone scale, its ten words to use and ten to avoid are words in a specific language, so crossing a border means rebuilding it with local vocabulary and examples. The catalog covers this only halfway - the translation prompt carries a register field and the voice generator carries a language field, but the values come from the market. Capitalization is not universal either: title case in English is not title case everywhere.

What a Model Cannot Sign Off On

The translation prompt limits itself on purpose: it returns only the translation plus up to three adaptation notes. Those notes are a sample, not the inventory, and the rest of the decisions travel invisibly. Back-translation catches omissions and gross shifts of meaning; it does not catch a sentence that reads like a translation. It is a tool, not a signature.

The signature belongs to a native reader who goes through the page as a customer, checking the form of address, whether the joke lands and whether the numbers read correctly. Where text carries legal force - contracts, terms, safety information, regulated claims - a qualified human translator signs. A model drafts; it does not certify.

Then the checklist, five items you can count: every number read in the market's convention; every query researched instead of translated; every fixed box tested with the longest string; the register confirmed by a native reader; nothing assembled from fragments.

Three AI Prompts for Translation, One per Job

Three artifacts cover the three jobs, and none decides the fourth row on the sheet.

The Mover

The translation and localization prompt asks for source language, target language and register, and returns only the translation plus up to three notes. Its five rules are the minimum contract for moving text: native fluency before literalism; adapt idioms, humor, units and cultural references with a note when the adaptation is risky; preserve formatting and placeholders; keep the register; cap the notes. It still needs the market from you, the glossary of what must not be translated, and the units you want converted.

The Voice

The brand voice style guide produces three voice pillars, a tone scale by context, ten words to use and ten to avoid, and a five-point test. Its own rule is that it stays usable by a freelancer who has never met the brand, and it carries a language field: exactly the artifact you rebuild market by market.

The Page

The SEO article outline prompt takes a keyword, an audience and an intent field, and returns a title built around the keyword, a meta description, H2 sections with their H3s and a short FAQ block. The keyword field is where the local query goes, which is why you run it once per market.

Translation is one job and localization is three more. Start with the prompt that moves the text, rebuild the voice guide for the market you are entering, and run the outline prompt with the local query before anyone writes a word. Keep the four-row sheet open, and fill the last column when a native reader signs.