Bilingual Newsroom Prompt Library, French and English

Twelve journalists using AI tools on their own produce twelve house styles. The fix is not a policy document, it is a working library: 30 to 40 prompts, each tested against your own published articles, each with a named owner, living where your editors already work rather than in a shared drive nobody opens. We build that library for French and English newsrooms, which is a different job from translating one set of prompts twice. French press writing has rules a general-purpose model breaks by default: sentence-case headlines rather than title case, guillemets with non-breaking spaces, accents kept on capitals, the twenty-four hour clock, and a register that reads flat and impersonal next to a British or American desk. English prompts written by a French team drift the other way, toward long subordinate clauses that sound translated. Both failures reach published copy within a month if nobody sets the guardrails. Why a library rather than a chatbot policy: A policy tells people what not to do. A library gives them a faster path than the thing you are trying to stop. Adoption follows convenience, so every prompt here has to beat the alternative of typing something improvised, or your editors will keep improvising. What is inside: Headline and deck sets split by surface, because a headline written to earn a Discover click and a headline written for a Search result are not the same object and should not come out of the same prompt. Standfirst and summary prompts tuned to your own average length, taken from your archive rather than from a general style guide. Structured data drafting: NewsArticle fields, author, section, and image alt text written to be useful to someone using a screen reader instead of stuffed with keywords. Translation prompts in both directions with a terminology file, so proper nouns, job titles, institution names and legal terms stay fixed instead of being reinvented article by article. Fact and sourcing checklists that force the model to mark what it cannot verify, plus refusal rules for anything touching health, courts, finance, or named private individuals. Rewriting prompts for updates, where the job is to change a live article without destroying the parts already earning traffic. How we test them: Each prompt runs against a set of your published pieces and is scored blind by your own editors against the human original. We keep what wins or ties, we cut the rest, and we tell you which tasks the model lost. Usually two or three do, and knowing which ones is worth more than the prompts that work. The four weeks: Week one, we read your style guide and 40 recent articles per language, and interview four editors. Week two, first draft library and the terminology file. Week three, blind testing and revision. Week four, delivery, two training sessions of 90 minutes each, and installation into your workflow, whether that is CMS custom fields, a Slack shortcut, or documents in Notion. What you keep: The prompts, in plain text, versioned, yours to change without us. A one page rule sheet per desk. The scoring results, failures included. A short maintenance note on when a prompt needs rewriting, which is usually when your section structure changes or the model version behind it moves. What this does not do: It does not write your articles, and it will not let you publish unread model output. It does not connect to your CMS through an integration we build; we work with what you already run. It does not include model or API costs. And it does not fix a style guide that does not exist. If yours is three pages written in 2016, that is a separate conversation and we will raise it before we take the work. Say no to us if: Your newsroom has decided AI has no place in it, in which case a library sits unused and you should keep your money. Your editors have no time for the blind test, because untested prompts are guesses with formatting. Or you want the one prompt that writes finished articles. That prompt does not exist, and anyone selling it to you is selling you a correction cycle.