Bayard Writer: an AI assistant that writes in a school's own voice

Bayard Writer is an AI writing assistant I built for Colegio Bayard, a bilingual school in Buenos Aires founded in 1956. It drafts the school’s communications to families in their exact voice. The hard part was never the model. It was building that voice from more than 3,500 of the school’s real emails, sorted by type, instead of a generic “school tone” prompt.

The problem: every school has a voice, and a generic AI doesn’t have it

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. Warm but formal, a specific institutional “we,” inclusive language written a particular way, the school’s own terms and course codes. The team wrote each message by hand. A generic AI could draft one fast, but it drafted like an AI, which is the exact thing a school does not want in front of its families.

The idea: a voice lives in real writing, not in a description of it

When I started, I did what most people would do. I asked the team to describe their voice. “Warm but professional.” Useful words, but you can’t hand them to a model and get the school back. Every draft built from a description came out sounding like a corporate email trying very hard to be warm.

So I stopped describing and started collecting. The school’s voice already existed, in years of emails they had actually sent.

How I built it

I pulled their real communications straight from the source. Running code in Google Colab against the Mailchimp API, I downloaded more than 3,500 of the school’s real campaigns.

Then I sorted them, because those emails are not one voice. They are several. I worked through them and grouped them into nine categories: events and celebrations, operational logistics, trips and camps, fees and admissions, academics, House assignments, family workshops, institutional newsletters, and the sensitive crisis communications.

And I built a voice profile for each category, because each one writes differently:

• An event invitation is festive, with semantic 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 measured, signed by the board, with no decorative anything.
• A House assignment is almost a fill-in template.

The audience changes the register too. How the school talks to families is not how the staff talk among themselves.

Then I encoded all of it: the exact openings, the institutional “we,” the school’s inclusive-language rule (the slash form, never the “e” or “x” forms), its own vocabulary (the Houses, the uniform supplier, the course codes like P5 and S2), date and time formats, which emoji means what, and which signature each kind of message carries.

The rule the school and I already shared

One of Bayard’s own rules: never invent a date, a place, a time, an amount, or a name. If the assistant doesn’t have a fact, it asks for it. That is exactly how I build everything. An AI that fills gaps with plausible-sounding fiction is worse than no AI. So Bayard Writer drafts the voice and the structure perfectly, and leaves the facts to a person who actually knows them.

What it runs on

Bayard Writer runs on the school’s own Claude (Anthropic’s Sonnet model), on free hosting (Vercel and GitHub), billed only for what it processes. The school’s data stays in the school’s account, and nothing trains a public model on it. It costs a low monthly amount, because you only pay for what you actually write.

What changed

The drafts come back recognizably the school’s, by category, in minutes. The team edits instead of writing from a blank page. The safe, repetitive categories move fastest. The delicate ones still get a human read before they go out, by design. The assistant does the voice; the people keep the judgment.

The takeaway

If you want an AI that sounds like your organization and not like every other AI, don’t describe your voice to it. Give it your real writing. The voice is already in there.

That is the core of how I work, on every project: loading the whole context of a business, or a school, into the AI, so what comes out is actually yours. If voice matters in what you send (a school, a clinic, a brand with a specific tone), I build AI chatbots and writing assistants trained on your own material. Tell me what you write most often and I’ll tell you what I’d try first. The first call is free, or ask Kyn right here.