What they did
The author wanted a short vertical video on why cola tastes different from glass, cans and plastic bottles. They wrote notes first: the script, a shot-by-shot order, the voice style and the look they wanted. Then they sent one short paragraph to a single Claude Code session, with every agent on Fable 5.2, and told it to run to the end without stopping. The session split the shots with one agent per scene, fact-checked the script, generated the voice, wrote word-by-word captions, reviewed the plan, built each scene in 3D as code and rendered it to MP4 with Remotion. It all ran inside a folder of format instructions the author had written ahead of time.
How it went
The run took 2 hours 25 minutes from the first log line and produced a finished 55-second video with 3D bottles, condensation, foam, molecule animations, narration and captions. The fact check found no false claims, the render check found no empty or frozen frames, and the run ended with a list of what it still doubted. It used 167.8M tokens, 95 percent of them cache reads, which would be about $158 at API prices but was covered by a Max plan. The author still wants a watch-through with sound, sound effects and tighter timing in one scene.
Worth knowing
The time spent writing the notes and the instruction folder is not in the 2 h 25 min, and the author sells that folder as a product. Since most tokens are the project re-read each turn, lock the script before scenes are built. All checks were made on stills, not on the moving video. The voice was billed separately by an unnamed provider.