Best AI Prompts for Translation (2026)

Prompts for translation#

Act as a localization lead. Translate and adapt the marketing copy below into Hebrew for readers in Israel. Keep an authoritative, professional tone, adapt idioms and examples instead of translating them literally, keep product names in Latin script, and use the glossary and approved sample I give you. After the translation, list every sentence where you changed meaning or were unsure. [PASTE YOUR COPY, GLOSSARY AND APPROVED SAMPLE]

Act as an internationalization engineer. Design the translation pipeline for a multi-language Next.js app that stores strings in JSON files: key naming, handling plurals and placeholders, which locales fall back to which, and how translated files are reviewed before release. My locales are [DESCRIBE YOUR LOCALES] and my current setup is [DESCRIBE YOUR STACK].

Act as a technical translator. Translate the documentation below into Latin American Spanish (es-419). Keep code, commands and identifiers untouched, use the term list I provide for technical terms, and mark any term you had to choose between with a bracketed note. [PASTE YOUR DOCUMENTATION AND TERM LIST]

Act as a multilingual SEO specialist. For the pages at [LIST YOUR URLS AND LANGUAGE/REGION VERSIONS], write the reciprocal hreflang annotations, including a self-reference on each page and an x-default, and list which pages need fully translated main content rather than a copy of the original.

Act as a bilingual linguistic reviewer. Compare the source text and the machine translation below. List mistranslations, dropped or added meaning, wrong register, and inconsistent terms, quote each problem, and propose a corrected line. Do not rewrite sentences that are already correct. [PASTE SOURCE TEXT AND TRANSLATION]

What to add to any translation prompt#

• The target locale as a language plus region (for example es-MX, pt-BR, zh-Hant), not just a language name. Google's hreflang documentation uses language and optional region codes, and W3C advises BCP 47 language tags.
• The reader and the purpose: who will read it and where it appears. Anthropic's guide tells you to explain what you want as you would to a new colleague who lacks context.
• Two to five examples of approved translations or terms. Anthropic recommends 3 to 5 relevant, varied examples in <example> tags, and Google's Gemini guide recommends few-shot examples with consistent format.
• Delimiters between instructions and the text to translate, such as XML tags, so the model does not treat your source text as instructions (Anthropic, Gemini).
• The output format, such as plain text, a JSON file with keys unchanged, or a table of source and target (Gemini: specify the shape of the response).
• For right-to-left languages like Hebrew or Arabic, say where the text will display. Direction is set by the dir attribute in HTML, not by the language tag (W3C).

A worked example#

Act as a localization lead. Translate and adapt the product page text inside <source> into Brazilian Portuguese (pt-BR) for small-business owners aged 30 to 50. Tone: friendly and direct, "voce" form. Keep the product name "ClearLedger" and all prices in USD unchanged. Use the term list in <glossary>. Match the style of the approved sample in <example>. Afterwards, list every sentence where you changed the meaning or were unsure.

<glossary>invoice = fatura; dashboard = painel</glossary>
<example>Envie sua primeira fatura em minutos.</example>
<source>Send your first invoice in minutes. Our dashboard keeps you on top of what's owed. Cancel anytime.</source>

A good result must:

• Use pt-BR spelling and vocabulary, not European Portuguese.
• Use "fatura" and "painel" as the glossary says and leave "ClearLedger" and prices unchanged.
• Rephrase "on top of what's owed" naturally instead of word for word.
• Match the register of the approved sample.
• End with a list of sentences it was unsure about.

The full prompt library is on our prompt library page.

Limits of this guide#

The vendor guides give general prompting advice, and none of them is specific to translation. The W3C and Google pages set rules for markup and indexing, not for translation quality. We did not test the prompts on any model, did not measure translation accuracy for any language pair, and could not read OpenAI's prompt guide through the archive route, so its claims here come from its live page on the date shown.