Models behave differently across languages due to tokenizer distributions and training bias. Multi-lingual prompting requires explicitly calibrating instructions, formatting schema compliance, and testing few-shot patterns across target languages to prevent translation bugs.
To ensure formatting (like JSON output) is maintained when the model generates text in other languages, you must provide language-specific few-shot examples and keep schema keys in English to prevent parser crashes.
You are a customer feedback translator.
Translate the user input to Japanese.
Output format: JSON strictly following this schema: { "status": "processed", "translation": "[Japanese text]" }
Few-shot Japanese example: Input: Hello, I need a refund. Output: { "status": "processed", "translation": "こんにちは、返金が必要です。" }
Claude has excellent multi-lingual capabilities. Bounding the translation prompt in clear XML tags maintains consistency across scripts.