Building a school platform with an AI pair programmer

SuperEscuela is a free learning platform for Spanish-speaking children, built in Costa Rica: a student app that works offline on old phones, and a records system for teachers. It is our own project, and we built it the way we build for clients: experienced engineers deciding, an AI coding assistant doing much of the typing.

Write the rules down, for the AI and for people

An AI assistant is only as good as the context it gets. The project has one short file of rules that every session reads first. A few of them:

• Discuss before doing. The assistant only changes code on an explicit instruction (“fix”, “build”, “update”). A question gets an answer, not a pull request.
• Hard constraints, stated once. Plain HTML and JavaScript with no frameworks, offline-first, light enough for the oldest phone. The assistant never has to guess the stack.
• Non-negotiables in writing. “Every teacher owns their own data. No leaks between accounts, ever.” Every read is scoped to the signed-in teacher; every write checks ownership first.
• One explanation per folder. Each part of the system documents itself, so the assistant reads the part it is changing instead of guessing from the whole.

A separate status file records what is ready, what is blocked and what was decided, so each working session starts where the last one stopped.

A sprint, start to launch

In early September the teacher records system went from a prototype to production in a few days. It moved from files to a real database, became multi-user with strict separation between teachers, and launched with attendance by day, grades by period and reports for families.

Before launch, the same AI that helped write the code was turned against it:

• An AI code review went through the week’s changes and listed problems in order of severity. Several were fixed on the spot; the rest went on a written list, in order.
• An AI security review tested whether one teacher could reach another’s data. The separation design held, and the sign-in issues it found were fixed and verified before launch.
• A 53-check test battery ran against the real system: sign-in, a teacher’s full workflow, grade and attendance rules, 28 attempts to cross between two accounts, all rejected, and 20 simultaneous writes. It found one real bug, which was fixed before launch.

What the AI changes, and what it does not

The assistant makes an engineer much faster at writing, reading and checking code. It does not decide what to build, what is safe to ship or what the data rules are. Those stay with people, in writing, where the assistant can read them every time.