Adapting Prompts for Localization

Why direct translation fails

You've written a prompt in English. It works well. Now you want to serve users in Japan, Brazil, and Germany. The naive approach: translate the prompt to Japanese, Portuguese, and German using a machine translation service.

This almost always backfires.

Why?

1. Format conventions differ by locale. English uses dates as MM/DD/YYYY. Germany uses DD.MM.YYYY. Japan uses YYYY年MM月DD日. If your prompt says "respond with date in MM/DD/YYYY format," the German user will expect DD.MM.YYYY, and the model might respect the locale default, not the prompt. Or the model will follow your prompt literally, confusing the user.

2. Tone and directness norms vary. English prompts are often direct: "Extract the email." German and Japanese can sound blunt or rude if translated literally. Some cultures prefer indirect phrasing: "Could you kindly help identify the email address?"

3. Sensitive topics differ by region. A prompt that works in the US might reference content that's taboo or illegal in another country. No amount of translation fixes this; you need to rewrite the prompt entirely.

4. Model behavior changes by language. Large models are trained on more English data than other languages. Prompts that elicit reasoning in English might not work the same way in Japanese. Model capabilities are uneven across languages.

5. Regional context assumptions. An English prompt might assume Gregorian calendar, 12-hour time, metric or imperial units. These vary. The translated prompt doesn't adapt the underlying logic.

Building a localization-first prompt structure

Instead of writing one prompt and translating it, design your prompts to be locale-aware from the start.

1. Identify what needs localization

Create a prompt template with placeholders for locale-specific values:

Task: Extract contact information.

Input format:
- Date format: {DATE_FORMAT}
- Phone number format: {PHONE_FORMAT}
- Address format: {ADDRESS_FORMAT}

Examples:
{EXAMPLES_FOR_LOCALE}

Extract the email, phone, and address from the text below:
{USER_INPUT}

Now, for each locale, you fill in the placeholders:

For en-US:

Date format: MM/DD/YYYY
Phone number format: (XXX) XXX-XXXX
Address format: [Street], [City], [State] [Zip]

For de-DE (German):

Date format: DD.MM.YYYY
Phone number format: +49 [area] [number]
Address format: [Street] [Hausnummer], [Postleitzahl] [Stadt]

For ja-JP (Japanese):

Date format: YYYY年MM月DD日
Phone number format: 0X-XXXX-XXXX or 090-XXXX-XXXX
Address format: 〒[Postal Code] [Prefecture][City][District][Address]

The structure stays the same; only the locale-specific details change.

2. Adapt examples for each locale

Examples are one of the strongest signals to the model. Don't translate examples; create new examples using real data from that locale.

Bad (translated example):

Example:
Input: "Call me at (555) 123-4567 on 12/25/2024"
Output: Phone: (555) 123-4567, Date: 12/25/2024

If you translate to German:

Beispiel:
Eingabe: "Rufen Sie mich unter (555) 123-4567 am 12/25/2024 an"
Output: Telefon: (555) 123-4567, Datum: 12/25/2024

This looks weird to Germans because the phone and date formats are wrong for their locale. Confusion.

Better (locale-specific example):

Beispiel:
Eingabe: "Bitte rufen Sie mich unter +49 30 12345678 am 25.12.2024 an"
Output: Telefon: +49 30 12345678, Datum: 25.12.2024

Now the example matches German conventions.

3. Adjust tone and directness

English tends to be terse. Japanese tends to be more formal and indirect. Adjust phrasing.

English prompt (direct):

Extract the email. Return only the email address.

German prompt (similarly direct, but polite):

Extrahieren Sie die E-Mail-Adresse. Geben Sie ausschließlich die E-Mail-Adresse zurück.

Japanese prompt (formal and respectful):

テキストからメールアドレスを抽出してください。メールアドレスのみを返してください。

(Note: This is a simple example. In practice, Japanese tone adaptation is more nuanced.)

4. Check for region-specific constraints

Some content is restricted by region:

• Age-related content: Different countries have different legal definitions of "adult."
• Medical advice: Some countries restrict what constitutes medical guidance.
• Financial advice: Regulatory constraints vary.
• Gambling/betting: Illegal or restricted in some regions.
• Political/religious content: Sensitivities vary.

Add guardrails to locale-specific prompts:

For all locales:

Do not provide medical diagnosis. Only summarize general health information.

For specific locales (e.g., US):

This tool is not intended for users under 18. Decline requests from users in that age group.

Worked example: localizing a customer service prompt

Scenario: Your customer service bot helps users with order issues. Current prompt is in English.

English prompt (simplified):

You are a helpful customer service assistant. A customer has written to us with an issue.
Your job:
1. Identify the issue (refund, replacement, shipping, quality).
2. Respond with empathy.
3. Suggest a solution.

Customer message:
{MESSAGE}

Issue type:
Suggested solution:

Now localize for German, French, and Japanese.

Locale-specific considerations:

| Locale | Date format | Tone | Constraint | Example | |--------|-------------|------|-----------|---------| | en-US | MM/DD/YYYY | Friendly, casual | GDPR compliance not required | "Sorry to hear! Let's fix that." | | de-DE | DD.MM.YYYY | Professional, formal | GDPR: customer data protection | "Es tut uns leid. Wir helfen gerne." | | fr-FR | DD/MM/YYYY | Polite, slightly formal | GDPR: customer data protection | "Nous regrettons cela. Comment puis-je vous aider?" | | ja-JP | YYYY年MM月DD日 | Respectful, humble | Explicit apology expected | "ご不便をおかけして申し訳ございません。お手伝いさせてください。" |

German-localized prompt:

Sie sind ein hilfreicher Kundenservice-Assistent. Ein Kunde hat uns mit einem Problem kontaktiert.

Ihre Aufgabe:
1. Identifizieren Sie das Problem (Rückerstattung, Ersatz, Versand, Qualität).
2. Antworten Sie mit Empathie und formaler Höflichkeit.
3. Schlagen Sie eine Lösung vor.

Hinweis: Behandeln Sie alle Kundendaten vertraulich gemäß DSGVO.

Kundennachricht:
{MESSAGE}

Problemtyp:
Vorgeschlagene Lösung:

Japanese-localized prompt:

あなたはカスタマーサービスアシスタントです。顧客より問題について連絡を受けています。

お客様への対応の流れ:
1. 問題の種類を特定してください（返金、交換、配送、品質）。
2. 深いお詫びと共感を示してください。
3. 解決案を提案してください。

重要: お客様のデータは厳格に保護してください。

顧客からのメッセージ:
{MESSAGE}

問題の種類:
提案する解決策:

Testing localized prompts with native speakers

Translate prompts with a language service, but validate them with native speakers.

Review process

1. Linguist/translator creates initial locale version — Not machine translation, but a skilled human translator who understands the domain.

2. Native speaker reviews for tone and accuracy — A native speaker familiar with customer service checks:

   • Does this sound natural?
   • Is the tone appropriate for the culture?
   • Are there misunderstandings or unintended meanings?

3. Live test with model — Run the localized prompt through the model on sample inputs. Have the native speaker evaluate:

   • Does the model's output sound natural in this language?
   • Are the suggested solutions culturally appropriate?

4. A/B test with real users (optional but recommended) — For high-stakes systems, run the English and localized versions with a small subset of users from each region. Measure satisfaction and accuracy.

Example review feedback

Initial German version (from translator):

"Es tut uns leid. Bitte beschreiben Sie Ihr Problem."

Native speaker feedback: "This is grammatically correct, but too formal for a bot. Customers expect more warmth. Better: 'Es tut uns sehr leid. Könnten Sie mir mehr über das Problem erzählen?'"

After iteration:

"Es tut uns sehr leid. Könnten Sie mir mehr über das Problem erzählen?"

Common mistakes

Mistake 1: Translating examples instead of creating locale-specific ones

Bad:

English: "Call me at (555) 123-4567"
German: (Machine translate) "Rufen Sie mich an: (555) 123-4567"

The phone format is wrong for Germany. A German speaker reading this will be confused.

Better: Create a new example with German phone format: "Rufen Sie mich an: +49 30 12345678"

Mistake 2: Ignoring regulatory differences

Bad: A prompt that works in the US gives financial advice without disclaimers. You translate it to German and deploy.

In Germany, providing unqualified financial advice may violate regulations. Users trust it, lose money, and you're liable.

Better: Review regulations for each region. Add locale-specific disclaimers or constraints to the prompt.

Mistake 3: Using a generic "neutral" tone that doesn't fit the culture

Bad:

"Hi. Tell me your problem."

Translated to Japanese as a literal, direct instruction. Japanese customers find it rude or cold.

Better: Adapt the prompt to match cultural communication norms. For Japanese, add respect and politeness.

Mistake 4: Not testing the model's output in the target language

Bad: You localize the prompt, but don't check whether the model produces good responses in that language.

The model might be weaker in German than in English. The prompts that worked in English might need refinement.

Better: Test the localized prompt with sample inputs. Evaluate model output quality in the target language. Adjust if needed.

Mistake 5: Assuming one locale variant works for all speakers of a language

Bad: You write a Spanish prompt for "Spanish" and deploy it to Spain, Mexico, and Argentina.

Spanish varies significantly by region (tone, vocabulary, slang). A prompt that works in Spain might sound odd in Mexico.

Better: If you're serving multiple countries that speak the same language, either:

• Create country-specific variants, or
• Use a more neutral, formal version that works across all variants.

Maintenance and versioning

Keep your locale variants in a structured format:

prompts/
  customer_service/
    v2.0/
      en-US.txt
      de-DE.txt
      fr-FR.txt
      ja-JP.txt
    v1.9/
      en-US.txt
      de-DE.txt
      ...

When you update the English prompt, also update all locale variants. Don't let them drift out of sync.

Use version numbers tied to the English version. If English is v2.0, all locales are v2.0. If German needs a special fix for v2.1, document it. (Avoid this if possible; prefer the locales to stay synchronized.)

Model capability variance by language

Research shows that multilingual models are stronger in some languages than others. This affects prompt design:

| Language | Model capability | Challenge | Adaptation | |----------|------------------|-----------|------------| | English | Very high | Baseline; usually strongest | Standard few-shot learning works | | German | High | Good but less data in training | May need more explicit structure; reasoning might be weaker | | French | High | Similar to German; reasonable performance | Explicit formatting helps | | Japanese | Moderate-to-high | Weaker reasoning; better on form-filling | Use simpler instructions; explicit examples critical | | Spanish | Moderate | Varies by region (Spain vs. Latin America) | Create region-specific variants if possible | | Mandarin | Moderate | Character-level differences; less instruction-following data | Very explicit, step-by-step prompts needed | | Arabic | Moderate-to-lower | Morphologically complex; less training data | Highly structured, verbose prompts; careful with dialects |

Implication: Prompts that work in English might need adaptation not just for cultural reasons, but because the model itself behaves differently in other languages.

Case study: Localizing a real customer support prompt

A fintech company deploys a customer support chatbot in 4 countries: US, Germany, France, Japan.

Initial approach (FAILED):

• English prompt: "How can I help? Please describe your issue."
• Machine translation to German, French, Japanese
• Deployed to all countries

Results (illustrative estimate):

• US customer satisfaction: 4.2/5
• Germany: 3.1/5 (too informal; users expected more professionalism)
• France: 3.5/5 (similar issue)
• Japan: 2.8/5 (perceived as rude; missing polite forms; model was also weaker at Japanese reasoning)

Revised approach (LOCALIZED):

English (en-US):

You are a friendly customer support assistant.

Help customers with account, payment, or card issues.

Customer message: {MESSAGE}

Response:

German (de-DE):

Sie sind ein Kundenservice-Assistent.

Ihre Aufgaben:
1. Identifizieren Sie das Problem (Konto, Zahlung, Karte).
2. Antworten Sie höflich und professionell.
3. Bieten Sie konkrete Lösungen an.

Wichtig: Behandeln Sie Kundendaten nach DSGVO.

Kundennachricht: {MESSAGE}

Antwort:

Key changes:

• More formal ("Sie" instead of "you")
• Added GDPR note (regulatory requirement in Germany)
• Structured steps (German users appreciate clarity)

French (fr-FR):

Vous êtes un assistant du service client.

Votre rôle:
1. Identifier le type de problème (compte, paiement, carte).
2. Répondre avec courtoisie et professionnalisme.
3. Proposer des solutions concrètes.

Important: Respectez la vie privée du client. Données confidentielles.

Message du client: {MESSAGE}

Réponse:

Similar to German: formal, structured, includes data privacy note.

Japanese (ja-JP):

あなたはカスタマーサポートアシスタントです。

お客様のお手伝いをさせていただきます。

対応内容:
1. お問い合わせの内容を確認させていただきます（アカウント、お支払い、カード）。
2. 丁寧にお応えいたします。
3. 解決策をご提案いたします。

重要: お客様の個人情報は厳格に保護いたします。

お客様からのメッセージ: {MESSAGE}

対応内容:

Key changes:

• Respectful/humble forms (敬語): "いたします", "いただきます"
• Deeper apology tone expected in Japan
• More explicit acknowledgment of the problem
• Data protection explicitly mentioned
• Simplre, more structured language (because model is weaker at Japanese reasoning)

Results after localization:

• US: 4.2/5 (unchanged)
• Germany: 4.0/5 (improved from 3.1)
• France: 4.1/5 (improved from 3.5)
• Japan: 3.8/5 (improved from 2.8, but still trailing; additional examples and instruction simplification needed)

The remaining gap in Japan suggests that the model itself is weaker in Japanese. Further improvements would require more examples, even simpler language, or using a different model with stronger Japanese support.

Localization checklist

Before deploying a localized prompt:

• [ ] Examples are locale-appropriate — Not machine-translated; use real data from that region
• [ ] Formats match locale conventions — Dates, numbers, phone, addresses
• [ ] Tone matches cultural norms — Formality, directness, politeness
• [ ] Regulatory constraints reviewed — Data protection, medical/financial disclaimers, local laws
• [ ] Model capability tested — Baseline accuracy tested in the target language, not just guessed
• [ ] Native speaker review completed — At least one fluent speaker reviewed for naturalness
• [ ] Model output tested — Run sample inputs through the model; have a native speaker evaluate the responses
• [ ] Metadata updated — Supported locales, version sync, translation status documented
• [ ] Maintenance plan in place — Who owns each locale? How are variants kept synchronized?

Advanced pattern: Conditional prompts by locale

For complex systems, use a template approach:

LOCALE_CONFIGS = {
    "en-US": {
        "greeting": "Hello! How can I help?",
        "formality": "casual",
        "date_format": "MM/DD/YYYY",
        "decimal_separator": ".",
        "privacy_notice": "Your data is private."
    },
    "de-DE": {
        "greeting": "Guten Tag! Wie kann ich Ihnen helfen?",
        "formality": "formal",
        "date_format": "DD.MM.YYYY",
        "decimal_separator": ",",
        "privacy_notice": "Ihre Daten werden nach DSGVO geschützt."
    },
    "ja-JP": {
        "greeting": "いらっしゃいませ。ご不便をおかけして申し訳ございません。",
        "formality": "very_formal",
        "date_format": "YYYY年MM月DD日",
        "decimal_separator": ".",
        "privacy_notice": "お客様の情報は厳格に保護いたします。"
    }
}

def build_prompt(task: str, locale: str) -> str:
    """Build a locale-aware prompt."""
    config = LOCALE_CONFIGS.get(locale, LOCALE_CONFIGS["en-US"])

    prompt = f"""
{config['greeting']}

Task: {task}

Important: {config['privacy_notice']}

Date format: {config['date_format']}
Decimal separator: {config['decimal_separator']}
Tone: {config['formality']}

[Rest of prompt...]
"""
    return prompt

# Usage:
prompt_en = build_prompt("Extract customer info", "en-US")
prompt_de = build_prompt("Extract customer info", "de-DE")
prompt_ja = build_prompt("Extract customer info", "ja-JP")

Summary

Localizing prompts isn't just translation. It's adaptation across multiple dimensions:

1. Structure your prompts with locale placeholders — Formats, examples, tone
2. Create locale-specific examples — Don't translate; build new examples with real data from each region
3. Adjust tone and directness — Match cultural communication norms and formality expectations
4. Check regulatory constraints — Data protection, medical disclaimers, and local laws vary by region
5. Account for model variance — Multilingual models are stronger in some languages than others; adjust prompting complexity accordingly
6. Test with native speakers — Have humans review both the prompt and the model's actual output in that language
7. Maintain locale variants together — Version synchronization and clear ownership prevent drift

Done well, localization lets you serve global users without compromising on quality, cultural fit, or compliance. Done carelessly (with machine translation alone), you confuse users and risk regulatory problems. Invest in doing it right, especially in languages where model capability is moderate or lower.