Opus 5.5 Ultracode made a 5-minute animated music video with Godot. Every pixel is code, and the singer is a voice it built from scratch

you can see the final video* here: https://youtu.be/GlMAVqqcQZk A little robot can't hear music. It can only count: drips from a tap, stars through a window, a table shaking when the radio plays. That's the first two minutes of *Tell the Dog*, a song and music video I made with Claude Code. Nothing in it came from a song generator or a video model. Every character, set, light and camera move is Godot code running inside my own little godot game engine, and the robot's voice is a singer Claude wrote from scratch. Here's how it came together. The voice I wanted to see whether the model could sing. Qwen TTS can't. So Claude built a synthetic singer in Python: glottal pulses through a formant "vocal tract", with presets. One of them, `small_robot`, sounded like exactly what it is, and I loved it. Then I asked Claude to use that voice as *its own* and say whatever it wanted. It wrote a little story straight into phonemes (ARPAbet, no English in the file), about a machine that knew songs were real because every morning someone was listening. That story became the song's spoken intro. The song Claude wrote the lyrics by hand: a breakup told through the dog ("I keep his water bowl full, for the neighbor's cat. There is no cat."). It composed the melody and arrangement in my Godot music tool, then sang the verses as speech on the beat and the choruses as notes. The video Claude rebuilt the timing from the song's source files: every syllable, kick and story event to the millisecond. The robot's LED-bar mouth is driven by five bands of its own voice. -Nine agents in parallel for the foundations:the audio data, four cast builders (the robot, the dog, the maker, the relay animals) and four set builders (workshop, kitchen, street, fence tops and rooftops). Each one rendered contact sheets, *looked at them*, and iterated until they read well. - A shot list with a critic. One agent drafted 124 shots with every cut on the beat grid. A second tore it apart: the ending ran past the music, the render tool would drop the coda, and colour was being applied twice. A third revised it. All 170 sync points were checked against the audio. - Ten build batches turned the shot list into 37 scene scripts, with strict file ownership so parallel agents never collided. - Every scene is a pure function of time. Any frame renders on its own, so agents could check any moment, and the whole 5:44 film renders in about three minutes at draft quality. My input Most of the thrust of it was in the creation of the song. I did a bunch of A/B tests for the voice and fine tuning of instrument sounds. I experimented with different mixes and thought the robot matched the aesthetic best. And I tried to get Claude to open up by speaking in phenomes and got the sweet response that the video starts with. I checked in on the progress throughout the day, nudging it when I saw it was lost on dead ends like analyzing the mixed song for lyrics and beats when the source code was available (albeit on my other computer), making the music video plot based on the video game who's engine I was borrowing, etc. This video took ALL DAY, and I stopped it short at the refinement stage before I ran out of usage. Claude thinks it is still rough in places (the robot's mustard coat goes olive in one light), but it exists, and it made me laugh and tear up a little. *Splendor Bytes – "Tell the Dog" Happy to answer any q's tl;dr Opus made a funky pop song + music video all through code, no gen ai.