Translate Content

Translate Content

Translate English source content into a target locale, preserving technical accuracy and structural integrity.

When to Use

• Localizing a skill, agent, team, or guide into a supported language
• Updating a translation that has become stale after source changes
• Batch-translating multiple items within a domain or content type
• Creating initial translations for a new locale

Inputs

• Required: Content type — skills, agents, teams, or guides
• Required: Item ID — the name/identifier of the content (e.g., create-r-package)
• Required: Target locale — IETF BCP 47 code (e.g., de, zh-CN, ja, es)
• Optional: Batch list — multiple IDs to translate in sequence

Procedure

Step 1: Read the English source

1.1. Determine the source file path:

• Skills: skills/<id>/SKILL.md
• Agents: agents/<id>.md
• Teams: teams/<id>.md
• Guides: guides/<id>.md

1.2. Read the entire source file to understand context, structure, and content.

1.3. Identify sections that must stay in English:

• All code blocks (fenced with triple backticks)
• Inline code (backtick-wrapped)
• YAML frontmatter field names and technical values (name, tools, model, priority, skills list entries, allowed-tools, tags, domain, language)
• File paths, URLs, command examples
• <!-- CONFIG:START --> / <!-- CONFIG:END --> blocks in teams

Expected: Full understanding of source content with clear mental separation of translatable prose vs preserved technical content.

On failure: If source file is not found, verify the ID exists in the registry. Check for typos in the content type or ID.

Step 2: Scaffold the translation file

2.1. Run the scaffolding script:

npm run translate:scaffold -- <content-type> <id> <locale>

2.2. If the file already exists, read it to check whether it needs updating (stale) or is already current.

2.3. Verify the scaffolded file has translation frontmatter fields:

• locale — matches target locale
• source_locale — en
• source_commit — current git short hash; the English revision a human translated against
• fence_basis_commit — the revision this file's frozen fences were verified against (#552). Equal to source_commit on a fresh scaffold, because a scaffold is a byte copy, and they diverge from the first edit of either kind. Never move this one by hand.
• translator — attribution string
• translation_date — today's date

Expected: Scaffolded file at i18n/<locale>/<content-type>/<id>/SKILL.md (or .md for other types) with correct frontmatter.

On failure: If the scaffold script fails, create the directory manually with mkdir -p and copy the source file. Add frontmatter fields manually.

Step 3: Translate the description

3.1. Translate the description field in the YAML frontmatter into the target locale.

3.2. For skills, the description is inside the top-level frontmatter. For agents/teams/guides, it is also in the top-level frontmatter.

3.3. Keep the translation concise — match the length and style of the original.

Expected: Description field contains an idiomatic translation that accurately conveys the original meaning.

On failure: If the description is ambiguous, keep it closer to literal translation rather than risk misinterpretation.

Step 4: Translate prose sections

4.1. Translate all prose content section by section:

• Section headings (e.g., "## When to Use" → "## Wann verwenden" in German)
• Paragraph text
• List item text (but not list item code/paths)
• Table cell text (but not table cell code/values)

4.2. Preserve these elements exactly as-is:

• Code blocks (``` fenced and indented)
• Inline code (backtick-wrapped)
• File paths and URLs
• Skill/agent/team IDs in cross-references
• YAML/JSON configuration examples
• Command-line examples
• **Expected:** and **On failure:** markers (translate the label, keep the structure)

4.3. For skills, translate the standardized section names:

• "When to Use" → locale equivalent
• "Inputs" → locale equivalent
• "Procedure" → locale equivalent
• "Validation" → locale equivalent
• "Common Pitfalls" → locale equivalent
• "Related Skills" → locale equivalent

4.4. For agents, translate:

• Purpose, Capabilities, Available Skills (section name only — skill IDs stay English), Usage Scenarios, Best Practices, Examples, Limitations, See Also

4.5. For teams, translate:

• Purpose, Team Composition (prose only — IDs stay English), Coordination Pattern, Task Decomposition, Usage Scenarios, Limitations

4.6. For guides, translate:

• All prose sections, troubleshooting text, table descriptions
• Keep command examples, code blocks, and configuration snippets in English

Expected: All prose sections translated idiomatically. Code blocks identical to English source. Cross-references use English IDs.

On failure: If uncertain about a technical term, keep the English term with a parenthetical translation. Example: "Staging-Bereich (Staging Area)" in German.

Step 5: Verify structural integrity

5.1. Confirm the translated file has the same number of sections as the source.

5.2. For skills, verify all required sections are present:

• YAML frontmatter with name, description, allowed-tools, metadata
• When to Use, Inputs, Procedure, Validation, Common Pitfalls, Related Skills

5.3. Verify code blocks are identical to the English source by running the checker, not by reading:

node scripts/check-i18n-fence-parity.js --locale <locale> --id <content-id>

Scope it to the file you just wrote. npm run validate:i18n-fences runs the whole corpus, which still carries a backlog (#477) and would fail for reasons unrelated to your work — training exactly the ignore-the-red-check reflex this gate exists to prevent.

Every fence whose tag is not text, markdown or md is frozen: its body must match the English source byte-for-byte, comments and string literals included. Repair with npm run normalize:i18n-fences, which previews; add -- --write to apply once the preview looks right. -- --tag yaml,json scopes the run to given fence tags, which is how the #477 backlog is cleared in reviewable batches.

The normalizer covers all four content trees — skills, agents, teams, guides — so it repairs exactly what the checker flags. -- --tree guides,agents scopes a run by tree the way --tag scopes it by fence tag, which is how the mirror slice was cleared as its own batch (#518). A --tree naming a tree your --locale carries no translations for exits 2 rather than reporting a clean-looking zero.

This step said "diff the fenced blocks" from the day the i18n tree was created, and the corpus still accumulated 1,307 violations, because reading a fence and diffing it are not the same act (#472). Run the command.

5.4. Check line count: skills must be ≤ 500 lines.

5.5. Verify name field matches the English source exactly (it is the ID, never translated).

Expected: Structurally valid translated file that passes validation.

On failure: Compare section-by-section with the English source. Restore any missing sections.

Step 5.5: Verify prose is translated

5.5.1. Sample 3 prose paragraphs from the body of the translated file. Choose paragraphs from different sections — not headings, not code blocks, not frontmatter.

5.5.2. Confirm each sampled paragraph is written in the target language, not English.

5.5.3. If any sampled paragraph is still in English, the translation is incomplete. Return to Step 4 and translate the remaining English prose before proceeding.

Expected: All 3 sampled prose paragraphs are in the target language, confirming the body text has been translated — not just headings and frontmatter.

On failure: Identify which sections still contain English prose. Translate them before continuing to Step 6.

Step 6: Write the translated file

6.1. Write the complete translated content to the target path using the Write or Edit tool.

6.2. Verify the file exists at the expected path:

• Skills: i18n/<locale>/skills/<id>/SKILL.md
• Agents: i18n/<locale>/agents/<id>.md
• Teams: i18n/<locale>/teams/<id>.md
• Guides: i18n/<locale>/guides/<id>.md

Expected: Translated file written to disk at the correct path.

On failure: Check directory exists. Create with mkdir -p if needed.

Validation

• Translated file exists at i18n/<locale>/<type>/<id>
• name field matches English source exactly
• locale field matches target locale
• source_commit field is set to a valid git short hash
• check-i18n-fence-parity.js --locale <locale> --id <id> passes for this file
• All cross-referenced IDs (skills, agents, teams) are in English
• File is under 500 lines (for skills)
• npm run validate:translations reports no issues for this file
• Prose reads idiomatically in the target language

Common Pitfalls

• Translating code blocks: Code, commands, and configuration must stay in English. Only translate surrounding prose.
• Translating the name field: The name field is the canonical ID. Never translate it.
• Translating tag values: Tags in metadata.tags stay in English for cross-locale consistency.
• Inconsistent terminology: Use the same translation for a technical term throughout the file and across files in the same locale.
• Literal translation of idioms: Translate the meaning, not the words. "Common Pitfalls" should become the locale's natural equivalent, not a word-for-word translation.
• Missing source_commit: Without this field, freshness tracking breaks. Always include it.
• Batch throughput over quality: Scaffolding-only output — where headings are translated but body text remains in English — is not a valid translation. Prefer fewer complete translations over many partial ones.
• Exceeding 500 lines: Translations may expand ~10-20% vs English. If near the limit, tighten prose rather than removing content.

Related Skills

• create-skill — understand the SKILL.md structure being translated
• review-skill-format — validate translated skill structure
• evolve-skill — update skills that have changed since translation