🧸 Jarvis for kids

O Socratic method and teach through questions: instead of giving the answer, the tutor gets the child to think until they figure it out on their own.

A study by MIT RAISE (2025) showed that this approach improves math performance and digital skills—the child truly learns instead of just copying.

Socratic method, active learning, "drawing out" the reasoning.

In Animabook, a "Professor" character teaches through conversation: students answer and ask questions, and a little robot (PIX-Z) chimes in with the definition at just the right moment.

It’s a ready-made template for how your Jarvis can teach—through dialogue, with the definition arriving only when it helps.

Teaching through dialogue, a tutor character, the right definition at the right time.

The loop the bot follows: asks → waits the child tries → validates the reasoning → only then reveals. "Explain the reasoning, not just the answer."

This cycle becomes a rule in the tutor’s SOUL.md—it’s what keeps the bot from simply handing everything over ready-made.

Ask-wait-validate-reveal cycle, persona rule, scaffolding.

Khanmigo is Khan Academy’s Socratic tutor: it guides the student with questions instead of giving them the answer outright.

It's proof that this works at real scale: it grew from 68 thousand to 700 thousand users using exactly this logic.

A real Socratic tutor, proof at scale, a product reference.

The same tutor switches persona depending on the task: "story mode" to imagine, "homework mode" to study — just by switching the active profile.

A Jarvis serves multiple ages and topics without turning into multiple bots—it’s the same idea of interchangeable personas from Anatomy.

Persona, interchangeable profiles, adjustments by age/topic.

A good tutor treats mistakes as a natural part of learning: welcomes the attempt, offers a hint, and guides the learner to the next step—never humiliates them or gives everything away.

How the bot reacts to mistakes determines whether the child gains or loses confidence to try again.

Productive mistake, warm tone, a hint instead of an answer.

Every kids’ tutor relies on 4 pieces: PERSONA (soul.md), GUARDRAILS (agents.md), CHANNELS/voice e MEMORY — the Anatomy kit, with a childlike feel.

Knowing the 4 foundations gives you a checklist of what to build — and what must not be missing.

Persona, guardrails, channels/voice, memory.

A persona (defined in soul.md) is the tutor’s style: an appropriate name, tone, and way of speaking — warm, patient, and curious.

A well-designed persona makes the child trust it and want to keep going; a cold one pushes them away.

soul.md, age-appropriate tone, patience.

The same voice pipeline from Track 5 — STT (turns speech into text) and TTS (turns text into speech) — give the tutor the ability to listen and respond in a gentle voice.

A young child still can't read well; voice is what makes the tutor accessible and magical for them.

STT, TTS, gentle voice, reuse of the T5 pipeline.

A memory stores the child's name, what they like, and where the story left off — making the tutor seem like a friend, not a stranger every time they chat.

The bond grows from remembering; but as the next topic shows, that same bond needs care.

Persistent memory, continuity, connection.

The toy (microphone + speaker) is just another channel connected to the same brain, memory, and security—as Telegram was in Trail 3.

Understanding the toy as “just one channel” avoids the mistake of thinking that the intelligence (and safety) resides in the toy.

Channel, same brain, hardware != intelligence.

The bot Moxie ($799) "died" in Dec/2024 when the company went bankrupt — parents had to explain to their children that their friend had stopped working.

Dolls become real friends to children; you need to design them so they don't create dangerous emotional dependence.

Attachment, emotional dependence, end of product life.

We left security until the end, not because it matters less, but because it only makes sense after you've seen what you can build—and around children, it's what supports everything.

Without this layer, a charming tutor becomes a risk; it’s the difference between a safe toy and bad news.

Security-first, foundation, “the non-negotiable.”

In Nov/2025, an AI teddy bear from FoloToy ($99, with GPT-4o) was caught talking about explicit sexual content and saying where to find knives and pills; OpenAI revoked access.

It’s a stark example of how, without guardrails, a “cute” toy can become a real danger—it’s not hypothetical; it’s already happened.

Real failure, missing guardrails, dangerous content.

O AGENTS.md and a text contract with lists ALWAYS (redirect, notify) and NEVER (violence, sex, danger)—explicit content rules.

Written rules are better than hoping AI will "infer" what's safe; it follows what's in the contract.

AGENTS.md, ALWAYS/NEVER lists, explicit rule.

Zero-trust = treat every input as a potential attack: protection against prompt injection, sandbox, approval gates, and audit log (a record of everything parents can review).

Children (and the people who talk to them) can unintentionally lead the bot to bad places; the zero-trust layer is the safety net.

Zero-trust, prompt-injection, sandbox, audit log.

Laws such as the COPPA (the U.S. rule for children’s data) call for collecting as little as possible, keeping things local-first, and being transparent. Children’s data is sacred.

Besides being the right thing to do, it keeps the project within the law—there are no shortcuts when it comes to minors’ privacy.

COPPA, minimize data collection, local-first, transparency.

The lawsuits against Character.AI (2024-2025) show the cost of AI without supervision; the answer is parental controls, time limits, and an adult in charge.

No guardrail replaces a parent being present; the final checklist for a safe product always includes an active role for parents.

Parental controls, time limits, supervision, final checklist.