Localization

Software Localization Skill

This skill creates high-quality localizations for software projects by using a dual-translation approach with verification.

Workflow Overview

1. Detect Project Type - Identify the localization framework and file format
2. Validate Target Language - Ensure a valid ISO 639 language code is provided
3. Extract Source Strings - Load English (en/en-us/en_us) as the base
4. Dual Translation - Two independent translator sub-agents produce translations
5. Verification - A third sub-agent compares results and selects the best translations
6. Output - Generate the localization file in the correct format

Step 1: Detect Project Type

Before any translation work, identify the localization system by examining the project structure.

Common Patterns to Check

Detection Commands

# Find common localization directories
find . -type d \( -name "locales" -o -name "locale" -o -name "i18n" -o -name "l10n" -o -name "translations" -o -name "lang" -o -name "*.lproj" -o -name "values-*" \) 2>/dev/null | head -20

# Find localization files
find . -type f \( -name "*.json" -o -name "*.yaml" -o -name "*.yml" -o -name "*.strings" -o -name "*.xml" -o -name "*.po" -o -name "*.pot" -o -name "*.resx" -o -name "*.arb" -o -name "*.xliff" \) 2>/dev/null | grep -E "(locale|i18n|l10n|lang|translation|messages|values)" | head -30

Important: Examine the English Source

Once you locate the localization directory, find the English source file. It may be named:

• en.json, en-US.json, en_US.json
• en.yaml, en-US.yaml
• en.lproj/Localizable.strings
• values/strings.xml (Android default)
• en.po, messages.pot

Read this file to understand the structure before proceeding.

Step 2: Validate Target Language

Required Information

You MUST have a valid ISO 639 language code before proceeding. If the user provides only a language name, look up the correct code.

Clarification Required For

ALWAYS ask for clarification when the user specifies:

• "Chinese" → Ask: Traditional (zh-TW, zh-HK), Simplified (zh-CN, zh-Hans), or another variant?
• "Spanish" → Ask: Spain (es-ES), Latin America (es-419), Mexico (es-MX), or general (es)?
• "Portuguese" → Ask: Brazil (pt-BR) or Portugal (pt-PT)?
• "Serbian" → Ask: Cyrillic (sr-Cyrl) or Latin (sr-Latn)?
• "Norwegian" → Ask: Bokmål (nb) or Nynorsk (nn)?
• "Malay" → Ask: Malaysia (ms-MY), Singapore (ms-SG), or Brunei (ms-BN)?

Common ISO 639-1/BCP 47 Codes

For reference, see ISO_CODES.md for the complete list.

Code Format

Match the format used in the existing project:

• If project uses en-US → use fr-FR, de-DE, etc.
• If project uses en_US → use fr_FR, de_DE, etc.
• If project uses en → use fr, de, etc.

Step 3: Extract Source Strings

Load the English source file and parse all translatable strings. Note:

• Preserve keys exactly - Do not modify key names
• Note placeholders - {name}, {{count}}, %s, %d, %@, etc.
• Note HTML/markup - <b>, <a href="...">, etc.
• Note pluralization - one, other, few, many forms
• Note context - ICU message format, gender variations, etc.

Step 4: Dual Translation with Sub-Agents

Delegate translation to two sub-agents: translator-alpha and translator-beta.

Instructions for Translators

Both translators receive the same instructions (see TRANSLATOR_INSTRUCTIONS.md):

1. You are a master linguist fluent in both English and {TARGET_LANGUAGE}
2. Translate each string maintaining natural, idiomatic expression
3. CRITICAL: Preserve all formatting exactly:
   • Placeholders: {variable}, {{variable}}, %s, %d, %1$s, %@
   • HTML tags: <b>, <i>, <a href="...">, <br/>
   • Special characters: \n, \t, \", '
   • Whitespace at start/end of strings
   • Markdown formatting if present
4. Adapt idioms appropriately - don't translate literally if unnatural
5. Consider formality level (formal/informal "you" where applicable)
6. Provide a confidence score (0.0-1.0) for each translation:
   • 1.0: Certain, standard translation
   • 0.8-0.9: High confidence, minor ambiguity
   • 0.6-0.7: Moderate confidence, context-dependent
   • Below 0.6: Low confidence, needs review

Output Format from Translators

Each translator returns:

{
  "translations": [
    {
      "key": "original.key",
      "source": "English text",
      "translation": "Translated text",
      "confidence": 0.95,
      "notes": "Optional notes about translation choices"
    }
  ]
}

Step 5: Verification with Adjudicator Sub-Agent

The translator-adjudicator sub-agent receives:

• Original English strings
• Translation A (from translator-alpha) with confidence scores
• Translation B (from translator-beta) with confidence scores

Adjudicator Instructions

See ADJUDICATOR_INSTRUCTIONS.md for full details.

The adjudicator evaluates each string pair and selects the better translation based on:

1. Accuracy - Does it convey the same meaning?
2. Idiomaticity - Does it sound natural in the target language?
3. Formatting preservation - Are all placeholders/tags intact?
4. Consistency - Does terminology match across strings?
5. Confidence scores - Factor in translator certainty

For each string, the adjudicator outputs:

{
  "key": "original.key",
  "selected": "A" | "B" | "merged",
  "final_translation": "The selected or merged translation",
  "reason": "Brief explanation of choice"
}

Step 6: Generate Output File

Create the localization file in the correct format for the project:

JSON Format

{
  "key1": "translation1",
  "key2": "translation2"
}

YAML Format

target_locale:
  key1: translation1
  key2: translation2

iOS Strings Format

/* Comment */
"key1" = "translation1";
"key2" = "translation2";

Android XML Format

<?xml version="1.0" encoding="utf-8"?>
<resources>
    <string name="key1">translation1</string>
    <string name="key2">translation2</string>
</resources>

PO Format

msgid "source text"
msgstr "translation"

Error Handling

• Missing source file: Ask user to specify the English source location
• Ambiguous language: Always ask for clarification (see Step 2)
• Unknown format: Ask user about the localization framework
• Large files: Process in batches if over 200 strings

Quality Checklist

Before finalizing, verify:

• All keys from source are present in output
• No extra keys were added
• All placeholders are preserved exactly
• File encoding matches source (usually UTF-8)
• File format matches project conventions
• Pluralization rules are correct for target language