Alexsandro Assis
← Build Journal
Applied AI 6 min

AI for technical translation: keeping context across PT, EN and ES

How to use AI in technical translation without losing tone, domain terms, canonicalId, and editorial consistency across languages.

Using AI to translate technical posts looks simple: send the text, ask for English and Spanish, save the result. In practice, that is the fastest way to lose context, tone, and consistency.

Technical translation is not just swapping words. It is preserving intent, structure, domain terms, metadata, and the author’s voice.

The problem with literal translation

A literal translation can look correct, but fail in important details:

  • Technical terms change inconsistently.
  • The personal tone becomes generic.
  • Titles lose strength.
  • Slugs become awkward.
  • The toc stops matching the headings.

In a technical blog, these small errors accumulate and break trust. The reader may not know how to name the issue, but they can feel when a text is too translated and not thought through enough.

Context before the prompt

Before asking for translation, the AI needs context. Not only the text.

A good input package includes:

  • source and target language;
  • author persona;
  • site audience;
  • terms that should remain in English;
  • terms that should be translated;
  • expected frontmatter structure;
  • rule to keep the same canonicalId across versions.

The clearer the contract, the lower the chance that AI improvises where it should not.

The technical translation contract

In a trilingual Astro site, translation needs to respect technical fields.

Some fields change by language:

  • title
  • description
  • language
  • canonical
  • displayDate
  • backLabel
  • tocLabel
  • sourcesHeading

Others must remain the same or equivalent:

  • canonicalId
  • section structure;
  • sources;
  • editorial intent;
  • related internal links.

This contract turns translation into a process, not a guess. AI helps accelerate the work, but the system needs to define what can vary and what must not break.

Review as a required step

AI accelerates, but it does not remove responsibility. Every technical translation needs a review step to verify clarity, naturalness, and precision.

In my ideal flow, AI generates the first version, n8n organizes the files, and human review approves before deploy.

The goal is not publishing in three languages at any cost. It is publishing in three languages while keeping the same trust.

References