Edit client footage dropped in a folder into captioned, graded cuts

The author, who posts as Asafsanany and edits video for clients, first tried having Claude edit clips interactively and found each small fix slow and rarely good enough. Over 13 days they built a local studio instead, with Claude co-writing almost all of its 1,173 commits. Each job is a folder per client and job. A watcher waits until the folder has stopped changing for a minute, then queues it. Every edit is a single non-interactive claude -p process run with a permission mode that refuses anything not pre-allowed, so it never waits on a person. A skill built on the open video-use method tells Claude Code to drive ffmpeg: cut from a word-level transcript, grade colour, then check its own render. They rewrote the skill to make its own judgment calls instead of asking questions, and to end with a line starting "Blocked:" when something is truly missing. An ElevenLabs key is part of the setup. Two jobs run at once, and a launchd agent keeps the studio up.

On a typical job, about 7 minutes of footage comes back around 23 minutes later as a cut with captions and a colour grade, and nobody opens a timeline. The result is a browser preview, not a render. The author types or speaks notes, a revise run applies them, and their median job takes three rounds before approval. Notes they mark as style become per-client rules for later jobs. Only after approval does a short run render the final MP4. The author still watches every result and says what is wrong; that step stays manual by design.

The headless setup only works because the skill forbids questions and the permission mode refuses anything unlisted; without both, an unattended run can hang. The engine checks the output files itself before calling a run a success rather than trusting the agent. Claude and ElevenLabs are paid. The code is published as Studio Engine under MIT, so this is also a launch post for the author's own project.