Music videos I made with Claude (Opus 5.5) in Claude Code, with everything they need to render. Every video here is code: each frame is rendered from the song time by programs Claude wrote, then encoded with ffmpeg. There is no editing software in the loop. I directed in conversation, and Claude agents wrote the plans, the engines, the scenes and the review passes and ran the renders.
| Video | Song | Look | Made with | Watch |
|---|---|---|---|---|
pdoom-tabletop |
"I'm Upping My P(doom)" (Claude-Pop version) | Stop-motion-style miniature on a researcher's desk at night: Clawd as a vinyl toy, a peg-doll researcher, a brass P(doom) gauge | Blender 5.2 (EEVEE) from Python scripts; 4K 60 fps | YouTube |
paranoia |
HEARTSTEEL, "PARANOIA" (last minute) | League of Legends montage of my own plays: CCTV intro, pentakills synced to the drums, impacts, split-screens, kill callouts | Python/PyTorch footage engine (RIFE slow-mo, time remaps, GPU effects, synthesized SFX); 1440p 60 fps | YouTube |
Each folder is self-contained: its README covers the concept, layout, requirements, render commands and credits,
and its docs/ hold the planning doc (TREATMENT.md), the brief the agents worked from (ENGINE.md) and notes on
how it was made.
The same rules run through every video:
- Every frame is a pure function of song time
t. No randomness, clocks or state carried between frames; seeded hashes instead. Frames can then render in parallel, out of order and resumably, and the previews match the final render exactly. - Plan first. A treatment (concept, look, palette, typography, one section per scene) comes before any code.
- Timing drives the edit. The song is analysed first (beats, downbeats, drum hits, word-level lyric times), and cuts land on downbeats, hits on beats.
- One file per scene, written by parallel agents. A lead agent writes the shared engine and a brief, then hands each agent one scene. The scene agents never edit shared code; they report engine bugs or needs to the lead.
- Look at the frames. Contact sheets, frame strips and stills are rendered and read back by the agents, and each shot is revised until it reads well: first and last frames, motion across consecutive frames, transitions.
- Something happens in every shot.
- Render offline, on rented GPUs when it's heavy. Trials at 1080p; finals at full resolution and 60 fps on Vast.ai RTX 5090s, encoded with NVENC.
The code is released under the MIT License. Songs, gameplay footage, fonts and other third-party pieces keep their own licenses; each video's README lists them.
0 comments
log in to comment.