Bilingual AI Content Workflows

Letting a model write directly to the public page creates three predictable problems: invented claims that nobody verified, a voice that drifts between authors, and German text that reads like a translation instead of Swiss business language. The fix is not a better prompt - it is a pipeline in which the model only ever sees a draft role.

1. Draft

   The model produces a first version from a content brief: audience, claims allowed, forbidden topics, source pages to reuse. Write the brief in English; it anchors terminology for both languages.

2. Review

   A person checks every factual claim against something verifiable - a live page, a repository, a contract. Anything the model asserts that is not in the brief gets deleted, not softened.

3. Publish

   The reviewed English version is translated and edited natively for de-CH - not machine-post-edited - then both versions ship together or not at all.

A gate is a question a reviewer must answer before the text moves on. Four are enough:

• Facts: can every claim be pointed at a verifiable source - a live system, a document, a released artifact?
• Voice: does it sound like the rest of the site - direct, technical, no marketing inflation?
• Scope: does it avoid forbidden territory - unreleased features, client names without consent, aspirational roadmap presented as fact?
• Locale: is the German written for Switzerland - ss, never the German Eszett, formal Sie, local business register terms?

Do not translate the prompt - translate the approved output. Maintain a small bilingual glossary of terms that must stay consistent (service names, product names, legal entities) and a short style sheet: sentence length, how to address the reader, what adjectives are banned. Our own rule of thumb: if a sentence would embarrass us in a contract, it does not go on the website.

• Fully automated publication - every public word still passes a human gate.
• AI-translated legal text - Impressum, privacy and contract language is written natively in each language.
• Voice-cloning individual people - content speaks as the company, not as a simulated person.

• Content brief exists before any draft (audience, allowed claims, forbidden topics)
• Every factual claim verified against a source before publish
• DE version written natively for de-CH, not machine post-edited
• Both locales shipped together
• Glossary and style sheet updated with any new term