58 Minutes of Footage, One Sentence: Claude Edits a 10-Minute Vlog

Indie developer Pieter Levels (@levelsio) tweeted that his vlog today was edited entirely by Claude Code with Opus 5.5. The starting point was ordinary: copy the video clips from the DJI into a folder, open Claude Code in that folder, and give it the requirement.

The only thing he said was:

Claude can you edit these clips into a 10 minute vlog for X and YouTube

He made a few small adjustments afterward, but not many.

The final video is here:

What is really worth looking at is the intermediate workflow. When replying to Sam Parr, Pieter broke it down:

• Used mlx-whisper to transcribe all speech. The model was Whisper large-v3-turbo, running on his MacBook Pro's GPU via Apple's MLX.
• Extracted frames from the video every few seconds and assembled them into contact sheets. This let Claude "see" what was in each clip.
• Used SpeechBrain's ECAPA model for voice matching to distinguish which parts were his voice and which were someone else's.
• Used OpenCV and the YuNet face detector to make thumbnails.
• The edit itself was a JSON cut list: which clip, start time, end time. Claude read the transcript, looked at the frame grids, and selected clips itself.
• Used Pillow to add on-screen captions to the first minute of the X and YouTube versions.

He did not ask for these steps one by one. Pieter said he did not know these had run behind the scenes until Sam Parr asked him how he edited it. Initially, he gave only that one sentence: "edit into a 10-minute vlog."

The amount of footage was not small: 58 minutes, 14 clips, finally cut to 10 minutes. A netizen asked whether Claude concatenated them in order or actually selected content, and Pieter replied that it cut out a lot. Another joked that going from 58 minutes to 10 minutes was equivalent to deleting 48 minutes of "so yeah, anyway."

People also asked about the cost. Pieter said he pays $100 a month for Claude, but that is for all his sites; this edit cost about $0.10.

The clever part of this workflow lies elsewhere. It first turns the video into an intermediate layer that the model can read: transcript text, extracted frame grids, voice labels, and a JSON edit table. The model makes decisions on text and image indexes, and finally outputs an inspectable edit list. A seemingly dumb approach like contact sheets actually solves the problem that "the model cannot see the footage."

Daniel Morgan is doing something similar. His description gets closer to the core of the problem: making AI cut video is already not hard; what is hard is making it cut your way. His approach is to have Claude Code edit the video into a Final Cut timeline and turn every note he gives into a rule.

Max Rohowsky showed another automated editing and motion graphics pipeline: screen recording and camera footage go into Tella or Cap, Parakeet does timestamped transcription, and then Remotion and Higgsfield add motion graphics in the right places. He explicitly said this is not necessarily 1:1 with Pieter's workflow, but it does something close: transcribe, cut, add graphics.

Someone suggested Pieter use Remotion, and he replied that he did not need to, because Remotion would make everything look the same and too flashy. This judgment is quite important. Automated editing tools easily flatten style; what Pieter wanted this time was a vlog, not templated motion graphics.

For developers, the most copyable part of this workflow may be a JSON editing state. Clips, in-points, and out-points can all be audited, modified, and rolled back. For users without a technical background, the signal is also clear: the barrier to video editing is dropping, although the first pass will not be perfect. Craig said in a reply that editing once made him stop after publishing 10 YouTube videos, because a 10-minute video took several hours outside of work, raising kids, and housework. Pieter put it more directly: without AI editing, he would not make vlogs, because editing is too painful and slow, and hiring someone is even more troublesome.

Right now this setup is still rough, needs a human to take a look, and needs a small amount of adjustment. But it has already compressed work that originally took several hours into one requirement plus a few revisions.