I couldn't get him to sit through ten minutes in the car with a screen anymore. I searched, I found nothing, so I built the podcast I wanted. Here's how — and you can listen to it right here.
It's a Saturday morning. We're heading to my parents' place, an hour's drive. My son is four. He's sulking in his booster seat because he'd rather have stayed home playing in the living room.
Like always, I hand my phone back over my shoulder. YouTube Kids. I start the channel of the moment, the one that plays on loop at our place. Thirty seconds later, the algorithm has already decided what comes next: an AI-translated English voiceover, ultra-saturated colors, a character bouncing around, an ad. Another video. An ad. Another one.
Halfway there, I glance in the mirror. His eyes are glazed over. He's not listening to anything anymore, he's just watching colors move around. And in the meantime I'm listening to the most brutal sequence I've ever heard aimed at a four-year-old. Not brutal as in "there's blood." Brutal in its pace. In the cuts. In the fact that no story holds together for more than forty seconds before the next video takes over.
Getting out of the car, I tell myself: this is over. I'm going to find him something else.
I started hunting for kids' podcasts. Seriously, not just skimming. For two weeks. I wanted something specific: in French, narrative, recurring heroes, an episode that wraps up in ten minutes, no ads. Apparently I was asking for the moon.
Most French kids' podcasts are just read-aloud tales. One voice, one story, the end. No recurring characters. No through-line. A few gems exist — Bestioles, Octave et Mélo, Encore une histoire — but none of them ticked every box at once. Above all, none of them dealt head-on with the real fears of a four-year-old: the dark, the monster under the bed, the first day of school, the doctor.
On the morning of day fifteen, I sat down and said out loud: "OK, I'll do it myself."
YouTube Kids has no reason to help you stop. The longer your kid stays, the more the platform earns. A podcast that ends in ten minutes, on the other hand, is built to finish. It's the exact opposite philosophy.
Before I tell you the rest, let me introduce the three heroes. You'll meet them in every episode.
Alma has a red lantern. She's brave, a bit of a ringleader, but gently so. She's the one who reaches out a hand. Her lantern lights up when she needs to dig deep inside herself for courage.
Lewis has a blue lantern. He talks to animals, he daydreams a lot, he watches the clouds. He sometimes needs a nudge to step up to things, but once he goes for it, he goes all in.
James has a yellow lantern. He's the smallest, the cheekiest. He has the power to turn invisible when he's scared of something — but at some point, he has to reappear. And that's where his lantern helps.
In each episode, one of the three faces a specific fear. The lantern lights up at the right moment and gives them just the right power. And the episode ends once the fear is beaten, with a return to somewhere calm. No cliffhanger, no "to be continued next week."
I've already lined up more than thirty fears to work through: the dark, the monster under the bed, the first day without mom, the doctor, thunderstorms, moving house, losing the comfort blanket, a new little brother, the barking dog, the slide that's too high, the fear of falling, the fear of trying, the fear of daring. Enough to fill two or three years of weekly episodes.
Why no cliffhanger?
Before age six, your kid doesn't know the next part is coming. They hold onto the tension. They fall asleep on it, they're still thinking about it at lunch the next day. What we take for a narrative device with older kids is a little anxiety that lingers for the small ones. So: resolved ending, back to calm. Every single time.
Put your headphones on, or turn it up. You can play them right here, no need for Spotify or Apple Podcasts for now — that'll come in a few weeks.
If you're listening with a kid next to you, I'd suggest starting with the episode about a fear they're going through right now. That's how the podcast is meant to work.
Alma is sure there's something moving under her bed at night. Her red lantern is going to show her what grown-ups no longer see: the difference between a real threat and a shadow making noise inside your head.
Lewis is learning to skateboard at the skatepark with his dad. Except every time he stands up on the board, he already sees himself hitting the ground. His blue lantern is going to give him exactly what he needs to stop staring at the floor.
First treetop adventure course for Lewis and James, two four-year-old cousins. James wants to do everything, Lewis sees the height and freezes. The lantern is going to move from the brave one to the one who's doubting — and not in the direction you'd expect.
First real tournament. A crowd, referees, an opposing team. James wants to shine, Lewis just doesn't want anyone looking at him. They're going to have to learn something neither of them saw coming: playing as a pair.
These episodes are first drafts. If a scene feels weird, if a word is too harsh, if the music drowns out the voice, write to me. I need outside ears to dial in the sound and the pacing before the official release.
I'm going to walk you through what happens between the moment I have an idea for an episode and the moment you can listen to it. No jargon, I promise.
I start by laying out the topic of the episode in one sentence. For example: "Lewis is scared of falling at the skatepark, his lantern gives him balance." Then I ask an AI assistant (I use Claude) to write the full script, giving it all my constraints: short sentences, a four-year-old's vocabulary, no cruel villain, a resolved ending required, the lantern refrain repeated three times. I reread, I cut, I rewrite. A script fits on four pages.
This is where the most impressive tool in the whole chain comes in: an AI that generates ultra-natural French voices. It's called ElevenLabs. I give it my script with little cues in brackets — [warm], [gently], [curious] — and it hands me back the voice of Paul K, my narrator, or Philippine, my kid voice actor. I do three takes of each line and keep the best one.
For the music, I use another AI that composes short little pieces from a description. I type "little ukulele melody, cheerful, two notes repeated in an arpeggio" and I get a backing track. For sound effects, same thing: "footsteps crunching in gravel", "creaking door", "child laughing". I end up with fifty to seventy sound files per episode.
Now I have to glue all of this together. For that, I have a little program — written by Claude Code over a few sessions — that takes the voices, the music, and the sound effects, and hands me a clean final audio file. It's the step that took me the longest to get right, but once it's dialed in, it runs on its own.
Last step: I tune the sound so it comes through clearly in a car (more on that a bit further down). And the episode is ready.
For a twelve-minute episode: about six euros in AI tools, and four to six hours of my time, from idea to final listen. No studio, no gear to buy. My laptop, full stop.
When I pitched the project to a friend who works in audio, he said: "You're going to use three different kid voices, right?" No. Just one.
Why? Because that's what Brigitte Lecordier did. If you're not in the loop, she's the French voice of Goku in Dragon Ball. And of his son Gohan. And of his grandson Goten. One actress, three characters, three generations, on screen at the same time in some scenes. And it works.
It works because when it's the same voice playing the whole crew, you feel that the three of them really belong to the same world. They have a cousin-like timbre. They sound familiar with each other in the listener's ear. If I'd used three different voices, I'd have had three characters who happen to hang out together. With just one, I've got a trio.
In practice, I tune small differences in tone between Alma, Lewis and James. Alma is steady and determined. Lewis murmurs and daydreams. James has a smile in his voice, always cooking up something. You can tell them apart perfectly when you listen, I promise you.
For the narrator, it's the opposite: an adult voice, steady, low, calm. It's the uncle telling you a story at bedtime. He's called Paul K, the same deep voice I was already using on my podcast for grown-ups, AI Wars. Except here, I asked him to do everything two notches softer. No dramatic intensity, no threat, no suspense. Just a smile in the voice and some patience.
When you push the emotional intensity too far on an AI-generated voice, you get "overacting." It's awkward for an adult. For a four-year-old, it's downright scary. It took me three or four attempts to find the sweet spot — warm but never theatrical.
While preparing this series, I read and had summarized a lot about what works and what doesn't before age six. In the end, I came out with a list of eight rules that I stick at the top of every script. None of them is negotiable.
I've said it already, but it's rule number one. Each episode begins, builds, resolves. The car can stop at the end, the kid gets out calm, you can move on to something else.
Before age seven, irony isn't understood. Everything is taken literally. If I have a character say "Oh great, more rain" in a jaded tone, the kid understands that the character is happy. No irony, no sarcasm. Straight down the line.
Low, sustained sounds — the "something shady is going on" kind from the movies — are anxiety-inducing and impossible to decode before age five or six. All my music is in a major key, with light instruments: ukulele, marimba, glockenspiel, flute. Never a low drone, never heavy brass.
The threat exists — it's what moves the story forward — but it stays gentle. At most a "grumpy one" who talks loud and gets won over. Never a cruel, calculating or ambiguous character. At this age, evil has to be beatable with a bit of clumsy kindness.
No screaming, no menacing crescendo, no moment where you clench your hands on the wheel. The emotional peak of each episode is joyful, not tense. It's hard to pull off when you're used to telling stories for adults, but it's non-negotiable.
You know those cartoons that slip in references to cult shows to make the parents watching alongside laugh? I won't do that. It breaks the pace, it distracts the kid, and deep down, it's a bit about flattering the parent. My job is to speak to the kid. If I want a format for myself, I make a podcast for grown-ups — and in fact, I've got one.
When two characters talk without the narrator stepping back in, the kid loses track of who's speaking after three lines. So I cut it off: three lines max, then the narrator recaps what's going on and we're off again.
Beyond a second and a half of total silence, attention drops off. Either we breathe with a light, discreet sound effect, or the narrator keeps going.
Eight rules, held across every episode. The hardest one to follow is number five. My whole way of writing for adults rests on tension. For kids, you have to unlearn it.
If you take away one technical point from this article, make it this one.
A podcast for adults, you listen to on headphones, or in the living room. A kids' podcast on a car ride is a completely different thing. There's constant background noise: the road, the engine, the wind. And the kid is listening on the car's speakers, which are rarely great.
If you tune the sound like a regular podcast, two things happen. Either the narrator disappears into the road noise, and the kid gives up. Or you push everything loud, and the music blows your ears out during a peak.
I found the fix by trial and error. I tune the sound like a radio show, and I drop the music twice as much as usual whenever the narrator is speaking. That way the voice always comes through, clear, over the ambient noise. And the music never jumps out at you.
It's a small detail. It's also the difference between a podcast you listen to closely and one you switch off after three minutes.
No project goes off without a hitch. Here are the three things that really gave me a hard time, told plainly.
On episode 2, I hadn't noticed that generating three takes of each line burned through three times more credit than I'd expected. After sixty-five lines, red screen, no more credit, half an episode stuck. The fix: since then, I keep an eye on the counter. And I only regenerate the failed takes, not the whole episode.
On the first render of episode 2, I felt like I was listening to a stretching class. The same little melody looping for two minutes under one scene. An adult tunes out. So does a four-year-old, except they can't say why. New rule now: one piece of music = its own length, no looping. I compose more pieces but shorter ones, and it breathes.
A sneaky trap: my assembly program had a little flaw that made it skip certain music tracks without saying a word. It took me three hours to figure out why my master sounded "empty" in places. Lesson: on an automated process, always look at what didn't get picked up. Not just what worked.
The first four episodes are right here, ready to play, free, no sign-up. I'd be thrilled if you played them for your kid and told me what they think. Write to me, I read everything. I need your feedback to fine-tune before the official release on Spotify and Apple Podcasts.
The method I describe here works for pretty much any narrative podcast: kids, teens, brand content, audio fiction, whatever. The tools are the same. All the code I use to assemble the episodes will be released open source once I've really got it dialed in — I'd say by the summer. Sign up for the newsletter so you don't miss it.
The fears I work through in the catalog (separation, the dark, the doctor, the monster, and so on) come from several sources of research in child psychology. If you see something missing, a word that isn't right, an approach that's debatable, tell me. I'd rather fix it now.
The podcast is made for kids aged 4 to 6, and tuned to the ear of a 4-year-old. If a 4-year-old gets it, a 6-year-old always follows along — the reverse isn't true. Episodes run between 10 and 16 minutes, exactly the length of a school run.
YouTube Kids chains videos together to keep the child watching as long as possible. Ad breaks hit at the worst possible moment. On a car ride, you can't cut it off or keep an eye on it. A podcast that ends in 10 minutes with a beginning, a middle, and a calm ending is built to stop when the engine does.
Yes, one hundred percent. The narrator is called Paul K, a French voice I picked in a tool called ElevenLabs. The three kids — Alma, Lewis and James — are all played by a single voice called Philippine. I set the tone, the emotion, and the pacing with little cues in the script. Then I listen to every take and keep the best one.
Between five and eight euros for a 12-to-15-minute episode. On top of that, my time: four to six hours, from the brief to the final master. The only gear involved is my computer. No studio, no sound engineer, no actors to pay.
No, and I'll say it without blushing: I'm not a developer. I use Claude Code to write the little scripts that stitch together the voices, the music, and the sound effects. What really takes human work is writing the script, choosing the voices, and listening carefully so I only keep the good takes.
The first four episodes can be played right here on this page, further up. They'll also land on Spotify, Apple Podcasts and all the major platforms within a few weeks. The easiest way to get a heads-up is to sign up for the newsletter at the bottom of the page.
I test AI for real and I share what works, no jargon and no hype. If this article helped, the easiest way not to miss anything is my Friday letter. And if you have a question or a doubt: reply to me, I read everything.