Multilingual Content Management

Instructions

Design and manage content pipelines that serve users in multiple languages with cultural sensitivity, not just word-for-word translation. Multilingual products require parallel content management, locale-aware AI behavior, and layout systems that accommodate diverse scripts.

Translation Workflow Architecture

Structure the translation pipeline as a multi-stage process:

1. Source authoring: All content authored in the primary language with translation-ready conventions (no idioms, no culturally embedded references, short sentences, consistent terminology)
2. Translation memory: Maintain a translation memory (TM) database that stores previously translated segments. New content is first matched against TM before sending to translation
3. Machine translation pass: Use MT (DeepL, Google Translate API, or Claude) as a first pass for new segments not found in TM
4. Human review: All machine-translated content reviewed by a native speaker for accuracy, tone, and cultural fit
5. In-context review: Translated content reviewed in the actual product UI to verify layout, truncation, and visual coherence
6. Publication: Approved translations published to the locale-specific content store

Content Structure for Multilingual

Design content storage to support parallel language versions:

content/
├── en/                    # English (source)
│   ├── ui-strings.json    # Interface strings
│   ├── help/              # Help articles
│   └── knowledge/         # AI knowledgebase
├── es/                    # Spanish
│   ├── ui-strings.json
│   ├── help/
│   └── knowledge/
└── ar/                    # Arabic
    ├── ui-strings.json
    ├── help/
    └── knowledge/

• Keyed strings: All UI text referenced by key, never hardcoded. Keys are descriptive (e.g., dashboard.equity.title, not str_042)
• Pluralization: Use ICU MessageFormat for pluralization rules, which vary dramatically across languages (Arabic has 6 plural forms; English has 2)
• Variables: Externalize all dynamic content with named placeholders: "Welcome, {userName}" — never concatenate translated strings
• Gender: Support grammatical gender where the target language requires it. Use gender-neutral defaults where possible

Locale-Aware AI Responses

When the product includes AI-generated content (chatbots, recommendations, summaries):

• Language detection: Detect user language from locale settings, not from input text (a Spanish-speaking user may type some English)
• Response language: AI responses in the user’s configured locale, regardless of input language
• System prompts per locale: Maintain locale-specific system prompts that account for cultural communication norms (directness, formality, honorifics)
• Knowledgebase per locale: Parallel knowledgebases for RAG — do not simply translate English KB documents. Local resources, regulations, and organizations differ by locale
• Fallback chain: If content is unavailable in the user’s locale, fall back to the closest related locale, then to the source language, with a visible indicator (“This content is not yet available in your language”)

RTL Layout Support

For Arabic, Hebrew, Farsi, and other RTL scripts:

• Logical properties: Use CSS logical properties (margin-inline-start not margin-left, padding-block-end not padding-bottom)
• Bidirectional text: Set dir="auto" on user-generated content containers. Use dir="rtl" on known RTL locale pages
• Mirrored layouts: Navigation, reading order, and icon direction (arrows, progress bars) mirror in RTL mode
• Numerals: Arabic-Indic numerals (٠١٢٣) vs. Western Arabic numerals (0123) — follow locale convention
• Mixed content: Handle bidirectional text correctly when RTL and LTR content appear in the same paragraph (e.g., brand names, URLs, code snippets in Arabic text)

Cultural Adaptation

Translation is necessary but insufficient. Cultural adaptation addresses:

• Date and time formats: MM/DD/YYYY (US), DD/MM/YYYY (Europe), YYYY年MM月DD日 (Japan). Use locale-aware formatters, never hardcoded patterns
• Number formats: Decimal separators (. vs ,), thousands separators, currency positioning
• Color meaning: Red means danger in Western contexts but luck/prosperity in Chinese culture. Audit color semantics per locale
• Imagery: Icons, illustrations, and photos may carry different cultural meanings. Review all visual assets per locale
• Humor and tone: What is friendly in American English may be too casual in German or too indirect in Japanese. Adapt tone, not just words
• Local resources: Replace US-centric resources (phone numbers, organizations, laws) with local equivalents

Content Staleness Management

• Translation status tracking: Every content item has a translation status per locale (current, stale, missing, in-review)
• Source change detection: When source content changes, automatically mark all locale versions as stale
• Priority queue: Rank stale translations by user impact (high-traffic pages first, error messages before marketing copy)
• Coverage dashboard: Visualize translation completeness per locale, per content type

Inputs Required

• Target locales and their priority ranking
• Content types to localize (UI strings, help articles, AI knowledgebase, marketing, legal)
• AI integration points that need locale-aware behavior
• RTL language requirements
• Cultural adaptation requirements beyond translation
• Translation resource availability (human translators, MT budget)

Output Format

• Content structure specification with locale directory layout
• Translation workflow diagram with stage gates
• Locale-aware AI configuration (system prompts, KB structure, fallback chain)
• RTL layout audit checklist
• Cultural adaptation guide per target locale
• Translation coverage dashboard specification
• ICU MessageFormat pattern library for common UI patterns

Anti-Patterns

• Translate-and-forget: Translating once and never updating when source content changes
• String concatenation: Building sentences by concatenating translated fragments — word order varies across languages
• English-centric layouts: Designing fixed-width layouts that break when German text is 30% longer or Chinese text is 50% shorter
• Machine translation without review: Publishing raw MT output without human review, especially for safety-critical content
• Monolingual knowledgebase: Using a single English knowledgebase for AI features and expecting it to serve all locales
• Flag icons for languages: Using country flags to represent languages (Spanish is spoken in 20+ countries; which flag?)
• Assuming Latin script: Hardcoding font stacks, text measurement, or input validation that assumes Latin characters
• Locale as language: Conflating locale (cultural conventions) with language (text translation) — Brazilian Portuguese and European Portuguese are the same language but different locales