An agent that writes articles in a brand’s voice

A publication, a brand, a team that writes every week: reviews, profiles, news. Every author has their own pen, and the brand’s voice gets diluted. The same empty phrases come back, an invented detail slips through, the English version ends up saying something other than the French one.

Review turns into rewriting. And nobody has the time to check every style rule by hand.

It is a Claude Code skill, used by the team of a food publication. Someone drops an interview, a transcript or a brief into a topic’s folder. It identifies the type of content (review, profile, news, recipe, sponsored content), writes the article, has a script check it, then files it next to the interview, ready for review.

It publishes nothing. Publishing stays a gesture of the team.

It has a second mode: translating an article that is already published, from French to English or the other way round, as a strict mirror. Same paragraphs, same order, same facts.

Four ingredients, merged: the facts of the brief, the brand’s voice (charter, vocabulary, list of forbidden terms), the structure of a template for each type of content, and a published article of the same kind to calibrate the tone. Five rules:

• Facts come from the brief, never from elsewhere. Nothing invented, nothing imported from other publications, nothing described as observed that was not. What is missing is listed for the person who reviews, not filled in.
• The charter is written down. Forbidden terms, empty phrases, invented comparisons, presumptuous claims: each safeguard is a rule, not a matter of taste.
• One template and one example per type. Among the reference articles, it takes the one closest to the subject, of the same type of business when possible.
• Rules are calibrated on what is published. The list of closing headings that are allowed comes from a survey of the reviews already online: most end without a closing section, so the template does not invite adding one.
• A translation is a mirror. Nothing added, nothing cut, nothing reordered, nothing “improved”.

A script checks what can be checked: forbidden terms, generic or stale headings, closing headings, dashes. A red result is fixed before delivery.

For a translation, a second script compares the structure of the source and of the translation: same blocks, in the same order, with comparable lengths. A defect already present in the original is passed on to the person who reviews, not quietly fixed.

Three situations from an imagined week (illustrative example, fictional content):

• An interview with a pastry chef. The file is called “call Tuesday”. The agent recognises an interview by its content, picks the profile template and writes the article. The script flags an empty phrase: it is fixed before delivery.
• A brief without the price of the dishes. The article does not mention it. The list handed to the person who reviews says what is missing.
• A French article to translate. The English version is filed next to the source. The script flags an added paragraph: the translation is redone.

• A rule that nothing checks ends up forgotten. The ones that matter are in a script, not only in the charter.
• Forbidding weighs more than recommending. A closed list of terms and headings changes the text more than “write well”.
• Translating is not writing. Loading the writing templates for a translation pushed it towards a structure the original did not have: the translation mode has its own template, and nothing else.
• The team names its files as it likes. The agent recognises them by their content, and only offers to move or rename them with the person’s agreement.

The same pattern works for any team that produces brand text: product sheets, press releases, descriptions, blog posts. A written charter, published examples, a check script, a person who reviews.