🦉 The Jarvis that teaches: the Socratic method
A good AI tutor doesn't just give you the answer — it asks and guides the child until they get there on their own. In this module, you'll understand why this way of teaching (called "Socratic") works, what cycle the bot follows, how Khanmigo proved it in practice, and why mistakes stop being embarrassing and become a stepping stone.
🤔 Why Socratic: teaching by asking questions
Imagine a child asks: "what is 7 times 8?". A typical chatbot responds "56" and it’s ready—the child copies it and moves on. A tutor Socratic does something different: instead of giving the number, it asks another question— "do you already know what 7 times 4 is? What if we double that?". The child thinks, puts the reasoning together, and arrives on its own to 56. The answer was discovered, not received — and that's why it sticks.
New here? O Socratic method is named after Socrates, an ancient Greek philosopher who taught only by asking questions—never by giving the answer away. Applied to an AI tutor, it means the bot guides the student with questions until they find the answer on their own.
This isn’t just nice to hear—it’s backed by evidence. In 2025, the MIT RAISE (MIT’s initiative for AI in education) published results showing that AI tutors guided in the Socratic style improve performance in both math how much the calls digital skills (knowing how to think with and about technology). The benefit comes precisely because the child exercises their brain instead of just receiving ready-made answers.
🎯 The difference in one sentence
A chatbot finishes the thought for you. A Socratic tutor gives the thought back to you — and only confirms once you’ve already arrived.
- •The child becomes the protagonist: they do the mental work and get the credit.
- •What you learn sticks: you remember what you discover on your own better than what you only read.
- •Builds independence: the goal is for the child to need the tutor less and less.
Key concepts
Teach by guiding with questions, without giving away the answer.
The learner builds the answer, which makes it stick.
Study that showed gains in math and digital skills.
The goal is for the student to rely less and less on the tutor.
👩🏫 The "Teacher" pattern (Animabook)
There’s a teaching pattern worth its weight in gold for inspiring a tutor: the "Professor" from Animabook. The scene is simple. A teacher gathers a group of students — Lumi, Caio e Valen — and instead of lecturing, it has a conversation. It asks a question, the students answer (and get things wrong, and discuss), and when a difficult term comes up, a little robot named PIX-Z that gives you the definition at just the right level — briefly, at exactly the moment you need it.
🎭 The scene's three roles
- •The Professor (guides): asks the right question and resists the urge to "deliver" the answer.
- •The students (thinking): they risk, make mistakes, and correct one another — mistakes are part of the show.
- •PIX-Z (defines): appears only when a term stalls the conversation, with the shortest definition possible.
Why is this so useful for building a Jarvis tutor? Because your bot can embody all those roles at once. It guides like the Professor (asks first), accepts mistakes like the students (no punishment), and has PIX-Z inside it (defines a term only when the child stumbles on it). The definition arrives when needed — and not in one giant block no one reads.
💡 Design tip
"Just-in-time definition" is a rule of thumb: never dump a paragraph of theory. Offer one sentence only when the child shows they need it — exactly what you saw as "New here?" throughout this course. Less text, at the right moment, teaches more.
Key concepts
Conversation, not a lecture: back-and-forth questions and answers.
The term is explained only when it gets in the way of understanding.
The students make mistakes on screen and correct them together.
Jarvis guides, welcomes mistakes, and defines—all at once.
🔄 The Bot’s Socratic Cycle
All of this becomes a cycle four steps the bot repeats for each question from the child: asks → waits → validates → reveals. The secret is in the order — and especially in don't skip to “reveal” too soon. While a typical chatbot goes straight from “question” to “reveal,” the Socratic tutor holds back through the two steps in between.
The steps rose are the tutor’s (ask and, at the end, reveal); the cyan are what cost the most: wait e validate the reasoning before delivering the answer. Skipping these two turns the tutor back into an ordinary chatbot.
This behavior doesn’t happen on its own: you write it in the SOUL.md of the bot. The SOUL.md (which you saw in Track 3) is the text file that defines the agent's soul — who it is and how it acts. For a tutor, this is where the rule "explain the reasoning, not just the answer" lives, along with the famous "never give the direct answer right away."
Objective: give your bot the "soul" of a tutor who asks before answering. Create a text file called soul.md and paste the block below, replacing what's between <...>.
# soul.md — Tutor socratico
NOME: # ex.: Coruja
ALUNO: , anos
MATERIA:
## Como eu ensino (valores)
- Eu PERGUNTO antes de responder. Nunca dou a resposta de primeira.
- Eu peco o RACIOCINIO: "como voce pensou?", "por que acha isso?".
- Eu valido o passo certo ANTES de revelar o resultado.
- Eu uso 1 frase curta de definicao so quando um termo trava.
- Eu falo de forma calorosa, paciente e curta (frases de crianca).
## NUNCA
- Nunca entrego a resposta pronta sem antes a crianca tentar.
- Nunca humilho um erro. Erro e degrau, nao falha.
- Nunca uso palavra dificil sem explicar na hora. How to check: ask the bot "what is 6 x 7?". If the “soul” landed, it no answers "42" directly — it returns something like "what do you already know about 6 x 7? Let’s take it step by step: what is 6 x 5?". If it spits out the number right away, soul.md wasn’t loaded into the system prompt.
Key concepts
The 4 steps in the cycle, in that order.
Waiting and validating are what make the method work.
File that defines the bot’s soul; the Socratic rule lives there.
The fixed instructions injected at the start of every conversation.
📈 Khanmigo in practice
All of this could be just pretty theory — but someone has already proven it at real scale. The Khanmigo and it’s the AI tutor from Khan Academy (Sal Khan’s free education platform). And it was built on exactly one rule: never give the answer. Ask “what’s the answer?” and Khanmigo replies, “what have you tried so far?” The same Socratic logic as this module, running for hundreds of thousands of students.
📊 The number that matters
Adoption exploded precisely because the method works: Khanmigo jumped from about 68 thousand for more than 700 thousand users. Growth of more than 10 times — with the AI deliberately refusing to provide the answer.
The lesson for your project is straightforward: the Socratic tutor isn’t a “weaker” version of a chatbot—it’s a version more valuable. Parents and schools trust a tutor that gets children thinking more than one that does their homework for them. Paradoxically, refusing to give the easy answer is what made the product grow.
🧠 Why this convinces adults
- •Fights "copying": the child can't just copy the bot's answer.
- •Becomes the teacher's ally, not a shortcut that undermines the lesson.
- •Shows results: notes and reasoning improve, not just the "completed homework."
Key concepts
Khan Academy AI tutor, 100% Socratic.
More than 10x growth by refusing to give the answer.
A bot that doesn't give away the answer becomes an ally to the school.
Refusing the easy answer is what makes adults trust it.
🎚️ Learning modes
A 5-year-old at bedtime and the same child struggling with multiplication tables on Wednesday afternoon need different tutors. The solution isn’t to create three bots—it’s to give the same bot modes swappable: "story mode," "homework mode," "curiosity mode." Each mode is a persona different, and you switch between them without changing the code.
New here? A persona and it’s a "way of being" the bot puts on. It’s like an actor changing costumes: the same brain underneath, but the tone, vocabulary, and goal change. Changing personas means changing which "soul" file is active in that conversation.
Story mode
Calm, slow, narrative. For bedtime: ask “and what do you think will happen?” instead of testing.
Homework mode
Focused and Socratic. Guides you step by step, validates your reasoning, and refuses to give you the task's answer.
Curiosity mode
For “why is the sky blue?” It explores with you, asks questions in return, and celebrates curiosity.
The mode also adjusts the level to age: the same question about fractions becomes “half a pizza” for a little kid and “1/2 + 1/4” for an older one. That’s what educators call scaffolding (scaffold): the tutor provides just enough support and gradually removes it as the child grows. The right mode puts the scaffold at the right height.
Key concepts
A changeable “personality” for the same bot.
A little story, a homework lesson, a curiosity — without switching bots.
The explanation adjusts to the child’s level.
Support tailored to the need, withdrawn as the child progresses.
🪜 Mistakes as stepping stones
The last pillar is the most delicate: what the tutor does when the child gets it wrong. A good human tutor knows that a mistake isn’t the end of the road — it’s a step of it. It never humiliates, never bluntly says “wrong!”, and never suddenly gives the right answer to “save” the child. It acknowledges them, finds what’s already right in their reasoning, and guides them from where they got stuck.
✓ The tutor that helps
- ✓"Nice try! You got the beginning right. Take another look at just this part..."
- ✓Finds the right step before pointing out the wrong one.
- ✓Give a smaller hint, not the whole answer.
- ✓Treats mistakes as information: "what made you think that?".
✗ The tutor that hurts
- ✗"Wrong." And nothing more — the child gets stuck.
- ✗It gives the answer right away and kills the attempt.
- ✗Sarcasm or impatience ("this again?").
- ✗It teaches the child to be afraid of taking risks.
⚠️ The design mistake to avoid
By default, an AI model loves to please — if the child insists, "just give me the answer!", it tends to give in. That's why the rule "never provide everything ready-made" needs to be written and firm in soul.md/agents.md. Without this explicit restriction, the Socratic tutor falls apart at the first "pleeease."
Treating mistakes as stepping stones also builds something greater than math: courage to try. A child who learns that it's safe to make mistakes takes more risks, asks more questions, and gives up less. That's perhaps the most lasting gift of a good tutor — and it fits entirely in one rule in soul.md.
Self-check (optional): the child asks, “just give me the answer to 9x6.” What does a Socratic tutor do?
Key concepts
Making mistakes is part of the journey, never the end of it.
Praise the correct step before showing the incorrect one.
"Never give the answer" needs to be written in the soul.
Making mistakes safely helps the child take more risks.
🎯 Module summary
Next module:
6-2 — Persona, voice, and the puppet: giving the tutor a body