How AI Tutoring Works: A Technical Guide

AI tutoring is the use of artificial intelligence to deliver personalized, interactive instruction to a student. Unlike a static textbook or a pre-recorded video, an AI tutor responds to what the student says, adapts its explanations based on comprehension, and adjusts difficulty in real time. The technology behind this is a large language model — specifically, in Evenfield's case, the Anthropic Claude API.

The System Prompt

Every AI tutoring session begins with a system prompt — a set of instructions that tells the AI who the student is, what subject is being taught, what the learning objective is, and how to behave. Evenfield builds this prompt dynamically for every session. It includes the student's age, grade level, the specific lesson content, instructor notes, and the student's recent performance history from prior sessions. This means the AI tutor is never starting from scratch — it knows where the student left off.

Age Differentiation

A nine-year-old and a six-year-old cannot be taught the same way. Evenfield enforces this at the prompt level. The nine-year-old receives longer explanations, harder questions, and more abstract reasoning challenges. The six-year-old receives shorter sentences, simpler vocabulary, more encouragement, and questions framed as play. The AI tutor for a six-year-old is limited to two sentences per response. The nine-year-old tutor is limited to three. Both are restricted to one question per response — short, focused exchanges that keep a child engaged.

Comprehension Signals

After every response, the AI tags its output with a comprehension signal: on track, needs attention, stuck, or off topic. These signals are stripped from the student's view but visible to the instructor in a live activity panel. If a student receives three consecutive "stuck" or "needs attention" signals, the instructor is alerted — they can then pause the AI tutor and take over direct instruction.

Session Memory

Every session is analyzed after completion. The full conversation transcript is sent to the Claude API with an analysis prompt that extracts concepts covered, concepts mastered, concepts the student struggled with, and observations about learning patterns. This analysis is stored in a persistent memory system — H.U.N.I.E., the central memory engine of the Jonomor ecosystem — and loaded into the next session's system prompt. The AI tutor remembers.

This is how AI tutoring should work: not as a chatbot answering questions, but as a structured lesson delivery system that knows the student, adapts to their pace, and gets smarter with every interaction. Evenfield is built on this architecture because the alternative — generic, one-size-fits-all instruction — is what the traditional school system already provides. There is no point in building something if it is not better.