Making videos as programs with Claude: frame rendering plus ffmpeg stitching
The author makes every video with Claude and no editing software, rendering frames in code and combining them with ffmpeg, and describes the 'harness' process of fixes and checks around the model.
Every “made with one prompt” video hides the part that actually did the work. I make all my videos with Claude, and none of them goes through a video editor. Each one is a program. A function draws one frame, a browser calls it 960 times, and ffmpeg glues the frames together. The first still of my last big one showed the player walking on thin air. The bridge was in the script and in the HUD counter. Nobody had written the code that builds it. That was one of six bugs. Particles from a crystal burst flew straight into the camera. A helper named top passed the syntax check and still crashed the browser. A caption said “too narrow” over a number that had just gotten smaller. A label sat on top of the headline. My wait loop searched for the render process, found itself every time and waited forever. A better prompt didn’t fix any of them. The fixes came from everything around the model: a rules file it reads first, checks that have to read zero, a person looking at the stills, and a loop with a stop rule a machine can verify. That layer has a name now: the harness. Anthropic says an early tester ran a 680,000-line migration on Opus 5.5 in under a day. That wasn’t a clever prompt either. It was hours of work with nobody typing the next message. I wrote up all of it. The article has the six bugs, the seven parts of a harness and a version you can build tonight, starting from a five-line rules file. Everyone screenshots the prompt. Nobody screenshots the checks.
This post quotes another post: x.com/dzerb1/status/2108880829929636231