RU
Get in touch
Chapter 18

Translator: Choose the Mode First

Translation seems simple until the mode is chosen. Pure translation, localization, editing, and live voice interpretation are different systems. Mix them, and the model will start explaining, improving, or shortening where the meaning had to be preserved.

So the mode is chosen first, and each has its own boundaries.

  • Faithful translation - convey the meaning, tone, and structure. You can't add facts or comments.
  • Localization - adapt units, forms of address, and idioms per policy. You can't change the legal meaning.
  • Editorial rewrite - improve clarity within the brief. You can't pass an edit off as a literal translation.
  • Live interpretation - low latency and continuous speech. You can't drag in agent tools and long explanations if the pipeline doesn't support them.
SQL
# Task
Translate <source> from Russian to English for developer documentation.

# Preserve
- meaning, degree of certainty, headings, and order;
- code, URLs, placeholders of the form {{name}}, and ICU message syntax;
- API names and glossary values without translation.

# Terminology
{{approved_glossary}}

# Ambiguity
- Don't pick a domain meaning if the context doesn't confirm it.
- Return ambiguous_terms[] with short options.
- Don't add notes inside the translation.

# Output
{ translation, ambiguous_terms[], glossary_violations[] }

Translation examples should also show not easy phrases but exactly the hard boundaries: a placeholder inside a sentence, a glossary term, a markdown table, an ambiguous form of address, units of measure, and text that must not be softened.

Live translation may not be an agent. In the Live API documentation Google explicitly separates Live Agent and Live Translation: the translation pipeline runs continuously, without tools or ordinary system instructions. Don't solve with a prompt what is better expressed as a separate model mode.

Links