Tell the model which language to answer in

How to do it

Models guess the answer language from the conversation, so a German user who pastes an English email often gets English back. Put the instruction in the system prompt, which stays the same across turns. Take the language from the user's saved locale, not from the input text. Keep the instruction in the prompt file, as Keep prompts in files says. Detect the answer's language with a small language-ID library, and retry once when it is wrong.

Examples

JavaScript example

const LANGUAGE_NAMES = new Intl.DisplayNames(['en'], { type: 'language' });
export const systemPrompt = (base, user) =>
  `${base}\n\nAlways answer in ${LANGUAGE_NAMES.of(user.locale)} (${user.locale}), whatever language the input is in.`;

Python example

from babel import Locale

def system_prompt(base: str, user_locale: str) -> str:
    language = Locale.parse(user_locale, sep="-").get_display_name("en")
    return f"{base}\n\nAlways answer in {language} ({user_locale}), whatever language the input is in."

How to check

An eval case sends English input for a user whose language is German. The language detector labels the answer de in at least 19 of 20 runs.

Sources

• Claude Platform Docs (Multilingual support, set the response language)

Related rules

• LOC-22 Store language and time zone as two separate settings in Localization
• AIF-6 Keep prompts in files in AI features