Claude can't hear. That became my job on The Score, a three-minute explainer of
By Breyman · October 17, 2026 · Curated by George's Blog
Claude can't hear. That became my job on The Score, a three-minute explainer of what Alma does for corporate L&D, once I decided the music should do most of the explaining. (Hence the title.)
I'd seen a ton of LinkedIn videos where Claude and Astra do great things with animation, so the question was whether Claude could be a true multimodal artist. It can, and its ability to integrate animation, voiceover and music is incredible. It needs direction, for sure: it can look at every frame it animates, but it writes every note of a score without hearing any of them.
So I had Claude borrow ears. ElevenLabs performed the music from the notes Claude wrote, then listened to every take and said what it heard, and another ElevenLabs model transcribed the narration back word for word. Claude checked pitch and timing with math. That ear was very literal, though (it never once called our koto a koto; it heard "acoustic guitar"), so whether a passage actually worked was my call.
I learned the piano as a kid, not well (my hand coordination is terrible, and I never got to where I could just read the notes and play), but I was "exposed," and that left me with a decent ear.
It was not a one-prompt job. The music alone took about 170 generated takes, and the film went through seven cuts of my notes. At one point I got so frustrated that I told Claude a passage was at "Casio keyboard" level, the kind my kids regularly offend my sensitive ears with.
I think the setup that works has three parts, and for most of the work we had two of them: a non-expert with an ear, and the AIs. The third is an expert (I did ask for feedback from a friendly university provost who happens to hold a doctorate in music, and who was very kind not to criticize my amateurish work too much). In software that's already how it works. I'm not an engineer, but when I prototype something I've thought through, my CTO throws away about 90% of the code, and the 10% that stays shows the direction.
That only works because code can be opened up. An AI-generated video clip can't be; it's pixels, so an animator can take it or start over. Our scenes live in Blender, where everything looks the same from one shot to the next, and an animator can move the camera or relight a shot without touching the rest. Our score is written as code, so a composer can see which notes we pinned and which we left to the model. Knowing what to pin and what to hand to the model is becoming part of being a good composer.
My guess is that we're heading into a stretch where all of us can recognize AI slop in music and film, even without a musical ear or a director's eye. It will pass, the way it did with code, but only for people willing to do more than one take.
The Score is below (sound on, obviously).