School announcements in five minutes

01Why a generic AI falls short

Bayard sends a lot of email to families: event invitations, schedules, trip logistics, fee updates, sensitive notices. Each one has to sound like the school and no one else. Warm but formal, its own institutional 'we,' inclusive language written a specific way, the internal terms and course codes only they use. The team wrote every message by hand, from a blank page, and it ate whole afternoons.

A generic AI drafts fast, but it drafts like an AI, which is the one thing a school does not want in front of its families. Ask it for 'warm and professional' and you get a corporate email straining to sound close. A seventy-year-old institution's voice does not live in an adjective. It lives in what it has already written.

02How I encoded the school's voice

I stopped asking for descriptions and went to the source. Running code in Google Colab against the Mailchimp API, I pulled more than 3,500 of the school's real campaigns. Then I sorted them by hand, because they are not one voice but several, and grouped them into nine families: events and celebrations, operational logistics, trips and camps, fees and admissions, academics, House assignments, family workshops, institutional newsletters, and the sensitive crisis notices.

Each family writes differently, so I built a voice profile for each one. An event invitation is festive, with emojis and a clear call to confirm attendance. A fee update is short and dry, no emojis, signed by administration. A crisis notice is solemn and signed by the board. Then I encoded everything that makes an email sound like Bayard: the exact openings, the institutional 'we,' the inclusive-language rule, the school's own vocabulary (the Houses, course codes like P5 and S2), date formats, what each emoji means, and which signature each kind of message carries.

Those nine families now live inside the system as a role-filtered category taxonomy, and the voice rules sit in context files that assemble in a fixed order before every draft. The model does not invent the voice. It is encoded, based on real communications from 2024 to 2026.

03Everything Bayard Writer does

• Three profiles by role (Teachers, Leadership, CEI): same base voice, different signature and available categories.
• A chip-guided flow, not a free prompt: pick category, audience and length, then describe the event.
• Two turns: first three titles to choose with a click, then the full announcement.
• Voice by category: each type brings its own tone, emojis, signature and checklist.
• Length as a real ceiling: Short, Medium or Long injects the exact character count into the model.
• A never-invent-facts rule: if a date or amount is missing, it writes [COMPLETAR: ...] instead of hallucinating it.
• Sensitivity modulator: for delicate notices it hardens the tone, drops emojis and signs as the board.
• Review tips at the foot of each notice: what to attach, who to coordinate with, who reviews before sending.
• Copy with formatting: pastes with bold and bullets intact into Gmail, Word or Bayard Hub.
• It runs on the school's own Claude: the data stays in its account and you pay only for what you write.
• An admin panel: role-based access and usage metrics (tokens, cost, conversations) by role and by user.

04From an hour to five minutes

Drafts come back recognizably the school's, by category, in minutes. What used to take close to an hour per message now takes under five: the team edits instead of writing from scratch.

The safe, repetitive categories move on their own. The delicate ones still get a human read before they go out, by design: the assistant does the voice, the people keep the judgment.

And it left a pattern I use on every project: load an organization's full context into the AI so what comes out is unmistakably theirs, not every other AI's.