I Built My Daughter a Kids' Song Studio with Claude Code
Lyria composes, Gemini judges, Whisper times every line, Remotion animates. Then the song sang Grandma and the screen showed Grandpa. What that taught me.
My daughter is a year and a half, and her favorite song is Baby Shark.
Somewhere around the hundredth play I asked myself: why only sharks? Why isn't there one like it about foxes, pigs, bears?
So I decided to build it.
Not one song. A whole studio that makes kids' songs about any animal I pick, including an animated video ready for YouTube.
How it works
With Claude Code and Google's models. One command:
- Lyria 3 composes and sings 6 versions
- Gemini listens to each one and scores it
- Whisper times every line
- Remotion animates a video that moves to the beat
Grandma, Grandpa, and a happy judge
Then we sat down to watch the song about the fox family.
The song sang "Grandma Fox".
On screen appeared Grandpa Fox, with a mustache and a cap.
Lyria liked Grandma's verse so much it sang it twice. Grandpa it simply skipped. And the video, fully obedient, showed Grandpa on Grandma's line.
And the judge? It gave the song an excellent score. It checked whether the song was good, not whether it was what I asked for.
So we built a little detective: it listens to the song, works out who is really being sung in each line, using two separate methods (because Whisper gets confused by short names), and compares that to the plan. Any mismatch, and everything stops. Nothing goes up to YouTube on its own.
15 animal songs later, she has a channel of her own. The pig got 76, the bear 59.
What I learned
A high score is not the same as the right result.
The hard part of AI isn't generating. It's knowing when not to publish.
Which check like that is missing in your work?
The channel: Hoppy Zoo on YouTube