Claude Code edits 58 minutes of vlog footage into a 10-minute cut, Levelsio says

• Claude Code cut about 58 minutes across 14 clips into a roughly 10-minute vlog, according to levelsio's follow-up.
• The agent chose its own transcription, frame-sampling and voice-matching tools after a one-sentence brief, levelsio says.
• The edit lived in a JSON cut list, and levelsio reported making only a few adjustments in his account.

Transcription ran on his MacBook Pro's GPU, while contact sheets let Claude inspect the footage as grids of still images. A speaker-verification model helped distinguish his voice from other people's.

Levelsio says he copied his DJI clips into a folder and opened Claude Code there. His entire brief was: “Claude can you edit these clips into a 10 minute vlog for X and YouTube”.

He identifies the model as Opus 5.5. The source material amounted to about 58 minutes across 14 clips, reduced to roughly 10 minutes.

His filming kit, listed beneath an earlier vlog, was a DJI Action 6 and Mic 3. That earlier vlog asked what's left to do when AI can build everything.

He linked the finished edit in a follow-up:

Claude assembled three ways to interpret the raw footage, according to levelsio:

1. Transcription: Whisper large-v3-turbo ran through mlx-whisper, using Apple's MLX on the MacBook Pro's GPU.
2. Visual inspection: Frames extracted every few seconds were combined into contact sheets, giving Claude a view of what each clip contained.
3. Speaker matching: SpeechBrain ECAPA identified levelsio's voice among other speakers. SpeechBrain's model card describes speaker verification through comparisons between voice embeddings.

He says he neither requested these individual tools nor knew the implementation details until someone asked how the vlog had been edited.

Claude stored the edit in a refreshingly plain format: a JSON list of cuts. Each entry identified:

• The source clip.
• The start time.
• The end time.

Claude selected those segments by reading the transcripts and inspecting the frame grids. The cut list recorded its editing decisions as references to the original footage.

Levelsio described two finishing steps alongside the edit:

• Thumbnails: OpenCV with YuNet's face detector.
• Captions: Levelsio says he added on-screen text using Pillow, Python's imaging library, including captions for the first minute on YouTube.

He described simple vlog edits like this as already doable in a reply, and predicted more complex edits would be possible “within a year”.

His cost estimate was “like $0.10”. He put that guess alongside a $100/month Claude subscription that he also uses for all his sites.

Without AI editing, he says he probably wouldn't be vlogging, citing slow, painful editing and the additional hassle of hiring someone.