Video Use

We handed it a fourteen-minute panel film from 1968 and asked for a minute. This is what came back.

The source is a public-domain industrial film in which four men argue that their profession should take moving pictures seriously instead of treating film as a luxury. Every objection they raise still lands — only the barrier has moved from celluloid to compute.

What it is

Editing by conversation. You describe the video you want in plain English; it asks, confirms, executes, iterates — and it will not touch the cut until you have approved the plan in words. There are no presets and no timeline.

What makes it work is that audio is primary. It transcribes to word level, then reasons about the cut from the transcript rather than from thumbnails — cut points come from speech boundaries and silence gaps, and it only looks at pictures when a decision actually needs one. That is why the edit above lands where sentences end rather than mid-breath.

How this cut was made

Fourteen minutes in, 4,071 words out. We read the transcript, picked six passages that argue the same thing in 2026 as they did in 1968, and let it place every cut edge on a word boundary with 50–80ms of padding. The subtitles are generated from that same word-level data, so they cannot drift out of sync with the cut.

The whole edit is a JSON file listing six time ranges and a reason for each. Change a number, re-render — no re-transcription, because transcripts are cached per source.

What to know going in

It costs money to run. Transcription goes through ElevenLabs Scribe, so you need an API key and every source you transcribe is billed. The skill itself is MIT.

We only exercised one path. Transcript-driven cutting and subtitles are what you see above. It also does colour grading, generated overlay animations, and multi-take selection across many clips — we have not put those through their paces, so treat our recommendation as covering the editing core.

Its renderer assumes modern footage. The output scale is hardcoded to 1920px wide — the comment reads Scale to 1080p from 4K — so a standard-definition archival source gets upscaled rather than left alone. Our 640×480 scan came back as 1920×1440 and 43MB for 51 seconds. Render without subtitles, then do your own scale-and-burn pass if your material is old.

It needs speech. The whole method rests on a transcript. Silent footage — a montage, a product film with only sound design — falls back to a much weaker path.

Installing

Not ours, and we do not repackage it — install from the source repository. You will need an ElevenLabs API key, plus ffmpeg on your PATH.