How Our AI Learning Coach Teaches — Without Giving Away the Answer

The AI Learning Coach is one of the few places in the app where a child talks directly to an AI system, not just receives AI-generated feedback on an exercise. That's a meaningfully different responsibility, so we built it around two deliberate constraints: a guided teaching style, and explicit safety rules baked into every prompt.

On teaching style: the coach is instructed never to hand over a complete answer immediately. Instead, it uses a Socratic approach — asking a guiding question, offering a hint, and only escalating toward a fuller explanation across a conversation if the child is still stuck. It's designed to behave like a patient tutor sitting next to your child, not a search engine.

On safety: every prompt sent to the underlying language model includes explicit rules — no violent, scary, romantic, or drug-related content, no requests for personal information, a warm redirect if the conversation goes off-topic, and encouragement to talk to a trusted adult if a child seems upset. There's also a keyword pre-filter as a first line of defence before a message ever reaches the model.

We want to be honest about what this is and isn't. This is prompt-based safety on a relatively small, efficient language model — not a purpose-built, moderation-tuned system, and not a guarantee. That's exactly why every AI Coach conversation is fully logged and visible to the guardian in the app, and why we keep iterating on the safety rules rather than treating them as finished.

If you ever see something in an AI Coach conversation that concerns you, we want to know — please reach out to us directly. A big part of doing this responsibly is treating it as an ongoing responsibility, not a box we checked once.