How one person produces thousands of lessons with AI

SuperEscuela is a free school that is played: short lessons with games, for Spanish-speaking children, on any phone, even offline. A small team cannot write a whole school by hand. So the content is written by AI, through a pipeline built to keep quality up as volume grows.

Four stages, AI in only two

1. Plan the tree (AI). From a topic or an official curriculum programme, the model proposes a tree of topics for one age level: what comes first, what depends on what, and what each topic should teach.
2. Walk the tree (plain code). A script goes through the tree and sends each topic to the next stage, in order. No AI, so nothing unpredictable.
3. Write the lessons (AI). The model writes every lesson of a topic in one go, with its practice questions, so a topic reads as one consistent piece rather than separate parts.
4. Publish (plain code). Each topic is exported as a small, self-contained file the app can download once and use offline.

Keeping the AI out of the mechanical stages is deliberate. Code that walks a list or writes a file should do exactly the same thing every time.

What keeps the quality up

• A strict format. Every topic is generated and checked against a fixed structure: the contract between the writing engine and the app.
• The curriculum as the source. Topics follow the official programmes of Costa Rica’s Ministry of Education (MEP) and SuperEscuela’s own learning path, not whatever the model happens to know.
• Written guidance by age. Instructions for the model differ by age band, based on a written framework of cognitive levels. A lesson for a seven-year-old and one for a fifteen-year-old are held to different rules.
• Small, independent topics. Each topic stands alone and levels never depend on each other, so a weak topic can be rewritten without touching the rest.
• Quality decides the model. The pipeline started on the smallest, cheapest model; measured against the bar children deserve, generation moved to a larger one, and a lesson still costs only a few cents to write.

People still own the result

The method and the content are open to review. Teachers and specialists are invited to go through topics and point out the error nobody saw, which is exactly the kind of check an AI cannot do for itself. The platform is in public alpha, and the content grows every week.