Long video into vertical clips

The content already exists. The job is finding the parts that stand on their own and reframing them for a phone — not writing anything new.

The flow

1. Get the file

If it is already on disk, skip this.

python scripts/download.py <url>
python scripts/download.py <url> --browser chrome   # signed-in content

Only download what there is a right to use. Someone else's video is someone else's video, whatever the platform makes technically easy. If the person asking does not own it and is not licensed to use it, say so instead of downloading it — the one exception being reference you study privately and do not republish, which is what the reference-research skill is for.

2. Find the moments worth cutting

python scripts/find_moments.py talk.mp4
python scripts/find_moments.py talk.mp4 --min 20 --max 55 --top 8

Transcribes the whole thing and proposes self-contained windows, ranked by how well they open and how densely they are spoken.

Treat the ranking as a shortlist, not a verdict. It is reliable at discarding the 90% that is throat-clearing and mediocre at picking the single best clip. Read the transcript of the top few and choose yourself.

A clip is worth cutting when:

• It opens on a complete thought, not on "and so"
• It makes sense to someone who has not seen the rest
• It has a payoff inside the window, not after it

3. Cut and reframe

python scripts/cut.py talk.mp4 --start 412 --end 458 --captions

Use --focus left or right with crop when the person is not centered — which is most interview footage.

--captions transcribes the cut clip and burns captions into it. Always use it. A clip from a talk is nothing but speech, and most of the feed is muted.

4. Check before publishing

Watch the first two seconds. That is where a repurposed clip usually fails: the speaker is mid-breath, or the first word is "and". Nudge --start a second later and cut again — it takes seconds.

When to build instead of cut

If the source has no usable hook, a cut clip will not gain one. Cases where the right move is the vertical-video skill instead:

• The good line is 40 minutes in and needs setup to make sense
• It is a slide deck being read aloud
• The audio is unusable

You can also combine them: cut the clip here, and use it as background media in a vertical-video project with your own narrated hook in front.

Mistakes that cost the most

Cutting on a timestamp instead of a thought. A clip that starts mid- sentence is over before it starts.

Cropping blind. Check that the speaker is actually in frame after cropping — crop with default centering throws away whoever sits on the left of a two-person interview.

Clips over 60 seconds by default. If the idea needs 90 seconds, fine, but it should be a decision.

Skipping captions because the original had good audio. The original was not watched on mute in a supermarket queue.

Scripts

scripts/download.py · scripts/find_moments.py · scripts/cut.py — all take --help.

Needs FFmpeg, yt-dlp and faster-whisper. The vertical-video skill's scripts/doctor.py --install sets all of it up.