translation

name	translation
description	Translate text and documents between languages with cultural and contextual adaptation

Translation

Translate text, documents, and content between languages with attention to context, tone, and cultural nuance.

When to Use

• User wants text translated between languages
• User needs a document, email, or webpage localized
• User wants to understand foreign-language content
• User needs culturally adapted content (not just word-for-word translation)
• User wants multilingual versions of their content

When NOT to Use

• Code localization / i18n implementation (use create-artifact skill)
• Writing original content in a specific language (use content-machine skill)

Methodology

Step 1: Understand the Context

Before translating, clarify:

• Source and target language(s)
• Content type: Casual text, business document, marketing copy, technical/legal, creative/literary
• Audience: Who will read the translation? (formal vs. informal register)
• Purpose: Information transfer, marketing impact, legal accuracy, or cultural adaptation?
• Tone: Should the translation match the original tone or adapt for the target culture?

Step 2: Translation Approach

Choose the right approach for the content type:

Content Type	Approach	Priority
Technical/Legal	Precise, terminology-consistent	Accuracy over fluency
Business	Professional, clear	Balance accuracy and fluency
Marketing/Creative	Transcreation — adapt the message	Impact over literal accuracy
Casual/Chat	Natural, colloquial	Fluency over formality
UI/UX	Concise, action-oriented	Brevity and clarity

Step 3: Translate

Translation principles:

• Translate meaning, not words — convey the intent, not a word-for-word mapping
• Preserve tone and register — formal stays formal, casual stays casual
• Handle idioms properly — find equivalent expressions in the target language, don't translate literally
• Maintain formatting — preserve bullet points, headers, bold/italic, links
• Keep proper nouns as-is unless there's a standard localized form
• Preserve technical terms — use established translations from the field, don't invent new ones

Formality — The T-V Distinction

Many languages force you to pick a formality register for "you." Get this wrong and the whole translation reads off. Ask the user or default to formal for business.

Language	Informal (T)	Formal (V)	Notes
Spanish	tú / vos	usted	LatAm vs. Spain differ; Argentina uses vos
French	tu	vous	vous is also plural — context disambiguates
German	du	Sie	Sie is capitalized; verbs conjugate differently
Portuguese	tu / você	o senhor / a senhora	Brazil: você is default-neutral
Russian	ты	вы
Japanese	—	—	No T-V; uses keigo — three honorific systems (teineigo polite, sonkeigo respectful, kenjōgo humble). In-group vs out-group matters more than hierarchy: you humble your own boss when talking to a client.
Korean	반말	존댓말	Six speech levels; verb endings change completely

Locale Data — Don't Guess, Use CLDR

Dates, numbers, and currencies follow Unicode CLDR (Common Locale Data Repository — what every OS and browser uses). Don't hardcode formats; look them up per locale:

Locale	Date	Number	Notes
en-US	3/14/2026	1,234.56	12-hour clock default
en-GB	14/03/2026	1,234.56	24-hour clock
de-DE	14.03.2026	1.234,56	Comma is decimal separator
fr-FR	14/03/2026	1 234,56	Thin-space thousands separator
ja-JP	2026年3月14日	1,234.56	Year-month-day, era names possible (令和8年)
ar-SA	١٤/٠٣/٢٠٢٦	١٬٢٣٤٫٥٦	RTL, Arabic-Indic digits, Hijri calendar option

In Python: babel.dates.format_date(d, locale='de_DE') handles this correctly.

Text Expansion — Plan for It

English → other languages changes length. Critical for UI strings, button labels, and anything with character limits:

Target	Expansion vs. English	Watch out for
German	+35%	Compound words don't wrap: Eingabeverarbeitungsfunktionen
French, Spanish	+20–25%
Russian	+15%	Plus Cyrillic glyphs are often wider
Finnish, Dutch	+30%	Agglutinative compounds
Chinese, Japanese, Korean	−10% to −30% (shorter)	But need larger font sizes for legibility

UI rule of thumb: design with 35% horizontal slack. Tell the user if their 12-char button label becomes 22 chars in German.

Tooling — When to Use What

Tool	Use when	Don't use when
LLM (you)	Context matters, tone adaptation, idioms, marketing copy, anything needing judgment	You need 10,000 strings translated identically
deep-translator (Python)	Bulk throughput, wrapping Google/DeepL/Microsoft APIs, simple string-in-string-out	You need glossary enforcement (it doesn't expose DeepL's glossary API)
deepl official SDK	Need glossary/termbase support — it actually exposes this
translate-toolkit	Parsing XLIFF/TMX/TBX/PO files — the standard localization formats

QA — Back-Translation

For high-stakes content (legal, medical, safety warnings): translate source→target, then independently translate target→source, then diff semantically against the original. Divergences flag ambiguity or drift. Don't use the same model/prompt for both directions — you'll get confirmation bias. This catches errors but is NOT proof of correctness; it's a smell test.

Glossary Injection

For multi-document projects, build a termbase (JSON of {source_term: target_term}) and inject it into your system prompt. Forces consistency across documents — "dashboard" shouldn't be tablero on page 1 and panel de control on page 5.

Best Practices

1. Specify the locale, not just the language — es-MX ≠ es-ES ≠ es-AR (vocabulary, formality, vos vs tú)
2. Flag transcreation — when you significantly adapt rather than translate (slogans, jokes), tell the user what you changed and why
3. Preserve placeholders exactly — {name}, %s, {{count}} must survive unchanged and in grammatically valid positions
4. Pluralization is hard — English has 2 forms (1 item / 2 items); Russian has 3; Arabic has 6. CLDR plural rules define these.
5. RTL is layout, not translation — Arabic/Hebrew/Persian text is RTL but numbers and embedded Latin stay LTR. Flag it; don't try to "fix" it in the text.

Limitations

• AI translation may miss subtle cultural nuances a native speaker would catch
• Legal and medical translations should be reviewed by a certified professional
• Poetry, humor, and wordplay may lose their effect — note this when it happens
• Cannot verify pronunciation for audio/speech contexts
• Some specialized terminology may not have established translations