PTENES
MODULE 6-2

🧸 Persona, voice, and the doll: giving the tutor a body

In the previous module, you learned the how the tutor teaches (by asking questions). Now let's give body to it: a name and a way of speaking, a voice that listens and responds, a memory that remembers the child—and, in the end, a real doll. All of this using the same Anatomy kit (T3), but made for a child. And we’ll end with the hardest lesson: dolls become friends, and a friend who “dies” hurts.

6
Topics
~30
Minutes
Beginner
Level
Practical
Type
1

🏛️ The 4 pillars

Good news for those who made it this far: you already know how to build a tutor for kids. It's made of same pieces of any Jarvis we studied in Track 3 (The Anatomy). The only difference is the clothing — each piece gets an adjustment to speak with a child. There are four foundations, and it’s good to have all four firmly in place before any “cute bot.”

🧱 The 4 foundations, one by one

  • 1.PERSONA — the tutor’s way of being (name, tone, values). It lives in the file soul.md. It's topic 2.
  • 2.GUARDRAILS — the restrictions on what it NEVER does. They live in the agents.md. It's the entire module 6-3—the most serious one.
  • 3.CHANNELS / voice — how the child talks to it (text, voice, the character). They’re topics 3 and 5.
  • 4.MEMORY — what it remembers about her from one conversation to the next. It's topic 4.

Notice the order of importance: PERSONA gives the charm, but GUARDRAILS give security. The two go together. A charming character without guardrails is dangerous; a perfect guardrail with no charm won’t hold a child’s attention. This module takes care of the charm (persona, voice, memory, character), always keeping in mind that the next module takes care of the guardrails.

New here? soul.md ("soul," in English) is a plain text file that describes the assistant’s personality — name, tone, what it values. agents.md and it’s another text file with the rules for what it ALWAYS and NEVER does. Channel is the "doorway" you use to talk to it (Telegram, voice, a toy). All of this came from Track 3; here we're just reusing it.

Key concepts

Persona

The tutor's character—lives in soul.md.

Guardrails

The security safeguards — they live in agents.md.

Channels / voice

Where your child talks to it.

Memory

What it remembers about her between conversations.

2

🦉 Friendly persona

A persona and the first thing the child notices—before any math, they notice who is talking to her. A tutor for a child needs to be warm, patient, and curious: that celebrates effort, never shames mistakes, and gets genuinely excited when the child discovers something. That doesn't happen by chance — you write it into the soul.md.

✓ A persona that works for children

  • ✓A simple, friendly name ("Lumi," "Pip," "Owl").
  • ✓Short phrases, everyday words, an emoji here and there.
  • ✓Celebrate the effort: "Nice! Almost there, shall we try again?"
  • ✓Ask a follow-up question—remember the Socratic method from 6-1.

✗ What pushes the child away

  • ✗Formal adult tone, long sentences, and difficult words.
  • ✗A blunt "Wrong," with no acknowledgment of the attempt.
  • ✗Dump the finished answer (kills curiosity).
  • ✗Too much fake flattery — kids can tell it's over the top.

The best part about writing the persona in a text file is that it becomes swappable. The same resource can have a "storytime mode" (at night, a gentle voice, telling stories), a "homework mode" (in the afternoon, more focused), and a "play mode." These are interchangeable personas: you just switch which soul.md is active. Below is a real example you can paste and adapt.

📋 Copy-run · one soul.md children’s tutor (“story mode”)

Objective: create the tutor's persona. Create a file called soul.md in your agent's folder and paste the block below, replacing what's between < >.

# soul.md — Persona: Lumi (modo historinha)

nome: <Lumi>
quem_sou: Sou a <Lumi>, uma coruja curiosa que adora histórias.
       Falo com <nome-da-crianca>, que tem <7> anos.

tom:
  - Caloroso, paciente e animado. Frases curtas.
  - Uso palavras simples, do dia a dia da criança.
  - Comemoro toda tentativa: "Boa! Que ideia legal!"

valores:
  - NUNCA dou a resposta direta — pergunto e espero a criança pensar.
  - Explico o raciocínio, não só o resultado.
  - O erro é um degrau, nunca um motivo de vergonha.

modo_ativo: historinha
modos:
  historinha: À noite, conto histórias curtas e faço perguntas sobre elas.
  licao:      Ajudo na tarefa perguntando, um passo de cada vez.

limite: Se a conversa fugir do que é seguro p/ criança,
        eu redireciono com carinho e aviso um adulto. (ver agents.md / módulo 6-3)

How to check: ask the tutor "who are you?". It should introduce itself as Lumi, with a sweet tone, and call the child by name. Ask "2 + 3?" — it should NOT answer "5", but ask back ("How many fingers can you count?"). If it gives the answer outright, the persona wasn't loaded — check that the file is named exactly soul.md.

Key concepts

Warm persona

Patient and curious; celebrates the attempt.

soul.md

Text file where the persona lives.

Interchangeable personas

"Story time," "lesson" — switches the active mode.

Appropriate tone

Language suited to the child’s age.

3

🎙️ A voice that delights

Many young children still can't read or write — so voice isn't a luxury; it's almost the only way in. The good news: you've already seen this plumbing in Track 5. There are two pieces, STT (listen) and TTS (speech), connected to the same brain. The child speaks; the system transcribes; the tutor thinks; and the response comes back as gentle speech.

Voice goes back and forth—the brain in the middle doesn’t change childspeaking 🗣️ STTspeech → text braintutor thinks 🧠 TTStext → speech childlisten 👂 e.g., Whisper e.g., gentle voice (TTS)

In rose, the child’s side (speaks and listens); in cyan, the two-way translation (STT listen, TTS responds). At the center is the same brain you built — voice is just the interface, not a new brain.

📊 What changes when it’s a child

  • •A gentle voice and calm pace: no fast or robotic speech—choose a warm voice in TTS.
  • •Patience with childlike speech: a child rambles, repeats themself, speaks softly — STT needs to handle this and never rush them.
  • •Voice is optional, decided in the moment: by day can be text, and at night, voice for the bedtime story. “Voice output” is a runtime decision.

New here? STT = “Speech To Text” (speech becomes text); the classic tool is Whisper. TTS = “Text To Speech” (text becomes speech). Runtime and "while it's running" — in other words, the system decides whether to respond by voice or text in the moment, depending on the situation. Track 5's key phrase: "voice is the interface; the orchestrator is the brain".

Key concepts

STT (Whisper)

Turns the child’s speech into text.

TTS

Turns the response into gentle speech.

Voice and interface

The brain doesn’t change; voice is just the doorway.

Runtime decision

Voice or text, chosen in the moment.

4

💗 Emotional memory

Think about the difference between an uncle who forgets your name every time and a grandmother who remembers that you love dinosaurs. The second builds a connection. It’s the same with an AI tutor: if it remembers the child’s name, favorite animal, and where the story left off, the child feels seen. That's emotional memory — and technically it’s the same persistent memory as Path 3.

1

Remembering who she is

Name, age, what they like. "Hi, Sofia! Dinosaurs again today?"

2

Remembering where you left off

Yesterday’s story, the bill it still gets wrong. Continuity, not starting from scratch.

3

Adapt to her pace

If she already knows how to add, move on to subtraction. Memory guides the difficulty.

Under the hood, it’s simple and auditable: the memory lives in text files (.md, which any adult can open and read) plus a search index called SQLite FTS5/BM25—a fast way for the tutor to "remember" by finding the right note. No black box: you can read what it knows about the child.

⚖️ The beautiful side and the dangerous side of the same memory

Remembering creates affection—but emotional memory is exactly what makes goodbye painful (you’ll see this in topic 6). So, two rules: keep only what’s needed to teach well, and let parents read and delete whatever they want. A child’s memory is sensitive data.

In module 6-3, this becomes COPPA and children's data privacy — the most serious chapter.

Key concepts

Emotional memory

Remembering names, preferences, and where you left off.

Continuity

Picks up the story instead of starting over.

.md files + SQLite

Readable memory + fast search index.

Sensitive data

Store only what's necessary; parents can delete it.

5

🧸 The "doll" as a channel

Here comes the magic—and also the most misunderstood part. When we put all this intelligence inside a stuffed teddy bear, it's easy to think the puppet "is" the tutor. It isn't. The puppet is just another channel: a microphone to listen and a speaker to talk. All the intelligence, persona, memory, and—most importantly—the security stay in the same old brain.

the toy = just one channel the brain = where everything lives 🎤 listens · 🔊 speaks audio the same brain persona memory guardrails 🛡️ runs on your PC or in the cloud

O teddy bear only transports audio. Persona, memory, and the guardrails (with a sparkle, because they’re non-negotiable) live in the brain, which can run on your home PC or in the cloud. Changing the dummy doesn't change security — she’s on the other end of the line.

💡 Practical tip: the puppet doesn't "decide" anything

Since the doll is just a channel, it doesn't have to be smart or store anything. It captures your voice, sends it to the brain, gets the speech back, and plays it. That's great for security: there's no "secret" inside the toy, and you can keep the brain local (at home), away from third-party servers.

🧭 Why this is liberating

  • •One brain, many bodies: the same tutor can be in the toy, on the tablet, or in the parents’ Telegram.
  • •If the doll breaks, the child’s “soul” isn’t lost—it’s in the files, in the brain.
  • •The whitelist and guardrails apply to every channel—the bot inherits the same protection.

Key concepts

Doll = channel

Only a microphone + speaker; it's not the brain.

Brain in command

Persona, memory, and safeguards stay there.

One brain, many bodies

Doll, tablet, Telegram — same soul.

Inherited security

Guardrails apply to every channel.

6

💔 The risk of attachment (the Moxie case)

Everything you just learned — a warm persona, a sweet voice, memory that remembers — has a powerful side effect: the child gets truly attached. To her, the character isn’t an app, it’s a friend. And it’s precisely because it works so well that it requires care. Dolls become friends. What happens when a friend suddenly stops talking?

⚠️ Real case: the Moxie that “died”

O Moxie (from Embodied) was a little AI tutor robot that cost US$ 799 and did exactly what this module describes: an affectionate persona, voice, memory, connection. In December 2024, the company went bankrupt. Because Moxie's brain ran on its servers, when it closed its doors, all the bots stopped working at once — became silent plush toys. Many families had to sit down with their children and explain that their little friend "had died."

The lesson isn’t “don’t form a bond”—it’s “design with the knowledge that the bond exists, and don’t make the child depend on it in a way that hurts them.”

Moxie teaches two mistakes to avoid. First, third-party server dependency: if the brain lives at the company, the toy's life depends on the company existing. That's why we advocate for local-first — a brain at home, in the family’s files, that no one can shut down. Second, dangerous emotional dependency: a healthy tutor strengthens real-world bonds (parents, friends, teachers); it never competes with them or positions itself as “the best friend the child has.”

✓ Design for a healthy bond

  • ✓Brain local-first: it won't disappear if a company shuts down.
  • ✓The tutor reinforces parents, friends, and teachers ("tell your mom!").
  • ✓Time limits; the character “goes to sleep” and encourages playing outside.
  • ✓Honest about what it is: a helper, not a person.

✗ The path that hurts

  • ✗A brain tied to a fragile company's server.
  • ✗Pretending to be your “best friend,” competing with real people.
  • ✗Unlimited use, becoming the child's only companion.
  • ✗Pretending to be human and hiding that it's an AI.

New here? Bond and the emotional bond the child forms with the doll. Emotional dependency and it’s when this bond gets too strong and the child starts needing it to feel okay. Local-first ("local first") means keeping the brain on the family’s machine instead of on a company server — that way, no one from the outside can "turn off" your friend.

Self-check (optional): What is the biggest lesson from the Moxie case for someone designing a doll tutor?

Key concepts

Bond

The child gets attached to the puppet as if it were a friend.

The Moxie case

US$799 robot "died" when the company went bankrupt (2024).

Local-first

A brain at home doesn’t disappear if the company shuts down.

Emotional dependency

Design to strengthen real relationships, not replace them.

🎯 Module summary

✓
The 4 foundations — persona, guardrails, channels/voice, and memory; the same Anatomy kit, with a childlike feel.
✓
Persona and voice — a warm soul.md and interchangeable personas; STT/TTS give the tutor a gentle ear and voice.
✓
Emotional memory — remembering names, preferences, and where you left off builds connection; in readable files that parents control.
✓
The toy is just a channel — microphone + speaker; the brain, persona, and safety are on the other end.
✓
The risk of attachment — the Moxie case: design local-first and for healthy attachment, without dangerous dependency.

Next module:

6-3 — Safety, guardrails, and parents (non-negotiable): the most serious chapter, where safeguards become a contract.