Toss the footage into one folder, and Claude finishes the rough cut in a café
Adam pours tea in a café while the timeline on his laptop screen grows by itself. The workflow has only four steps. What's really worth seeing is that it extracts frames to watch the footage, provides an edit list with timestamps, and, when he's not satisfied, changes only cut 3.
Adam posted a demo on X: a person is pouring tea in a café while a laptop edits a video in the background. All the footage is tossed into one folder, Claude Code is opened and switched to Claude Fable 5.2, one sentence describes the desired mood and length, and then he goes to drink tea. When he comes back, he checks the draft on his phone.
His four steps:
- Put the raw clips in the same folder
- Open Claude Code in that folder and switch to Fable 5.2
- Send one message describing the mood and length
- Go get a drink and check the draft on your phone
The key isn't that the timeline starts moving. Norri asked in the replies whether it actually looks at the footage or just makes random cuts at random timestamps. Adam said it first uses ffmpeg to extract frames from every clip, watches them, and then decides how to cut; it also relies on this to decide which shot to put in the first 2 seconds. This detail matters more than the demo visuals. Extracting frames is not the same as understanding narrative, but at least it isn't cutting blind.
He posted the prompt in the replies, roughly as follows:
Re Edit the raw clips in this folder into one vertical video for X, 9:16, 15 to 30 seconds long. Look through every clip first by pulling frames with ffmpeg and keep only the strongest moments. Cut pauses and shaky parts, and put the best shot in the first 2 seconds. Lay music.mp3 under the whole video. Add short English captions at the top. Render draft.mp4 with ffmpeg and list every cut with timestamps so I can change any of them
The most practical part of this workflow may be the last sentence—listing the timestamp of every cut. Aidan asked what to do if he didn't like a particular cut; Adam replied, 'swap cut 3,' and a few minutes later a new draft was rendered. Jarvi asked about captions; Adam said captions are also written and burned in at the top, and if any line is wrong, just point it out and re-render. Big V added in the replies: even as models get stronger, an edit list with timestamps is still worth keeping; you can also mark some pauses that are intentionally kept, because the breath before a reveal may be exactly what makes a short video work. This reminder is closer to real work than full automation.
It's also worth throwing some cold water on this. Ammar said bluntly, show me the export not the timeline filling up. A full timeline doesn't mean it's well edited. TheEasySale pointed out that the most valuable part of editing is choosing shots; Fable's criteria for choosing the best takes are what really matter—even if the emotion is right, if the pacing is wrong it's still unwatchable. Mohith asked how much manual fixing would still be needed; Adam didn't give a proportion in the material. Juan's experience was similar: using Claude and ffmpeg to stitch still frames into an Instagram reel—the draft is automatic, but the final cut and approval are still in his own hands.
So what's worth watching in this demo isn't the slogan of one person replacing a small studio. More concretely, it turns rough cutting into a draft that can be rolled back: extracting frames to watch the footage, producing a first version according to mood and length, using timestamps to locate problems, and re-rendering only the affected part. The human is still there, judgment is still there; it's just that the evening is no longer entirely handed over to the timeline. As for whether the exported final video is watchable, you have to look at the export.
Original post: Adam's demo
发布时间: 2026-11-15 15:26