How it works

Stream VODs into Shorts, without watching them back.

Point a workflow at a stream recording. It transcribes the audio, scores candidate moments using both what was said and what was visibly happening, keeps the best few, cuts them at exact timestamps, crops to 9:16 and burns in captions. You get ready-to-post clips instead of a six-hour file you were never going to scrub through.

The part most clipping tools get wrong is the selection. Slicing every sixty seconds produces thirty clips of which two are good. Picking by transcript alone finds the loud parts, not the interesting ones.

The workflow

1. File. The VOD stays on your disk. Rendering happens locally with ffmpeg — nothing is uploaded to cut a clip.
2. Transcribe. Speech becomes text with timestamps. This runs locally, so a three-hour recording does not cost three hours of API time. You get transcript and segments, and the segments matter — they are what the captions are built from later, so the words land on the right frames.
3. Analyse the video visually. This is the step that separates a good clip from a loud one. The analysis looks at frames as well as the transcript, so a moment is chosen because something happened, not just because someone shouted. It returns a list of candidates with start, duration and a title.
4. Split the list, then sort and keep the first. Three good clips beat thirty mediocre ones — and every clip you do not render is money you do not spend.
5. Cut a clip. Each candidate is cut at its own timestamps. Cutting is re-encoded rather than copied, because a copied cut snaps to the nearest keyframe and your clip starts a second and a half late.
6. Crop to 9:16. Vertical, centred, 1080×1920 — the format TikTok, Shorts and Reels all want.
7. Burn in captions. Using the original transcript segments plus the clip's own start and duration, so the timing is right even though the clip is a slice out of the middle.
8. Check the video before anything leaves the machine: aspect ratio, duration, and that there is audio at all.

Why the captions need the original segments

A common failure: captions are generated from the clip after cutting, which means transcribing the same audio twice and paying for it twice. Worse, the second transcription drifts, and the words no longer sit on the frames where they were spoken.

Passing the original segments plus the clip's start offset costs nothing and is exact. That is why the caption node takes clipStart and clipDuration as well as the segments.

What it costs

Transcription runs on your machine and costs no credits. The visual analysis pass is the one paid step, and it runs once per recording — not once per clip. Cutting, cropping, captioning and checking are local ffmpeg work.

Variations that work with the same shape

• Podcast episodes — same workflow, usually with a static background instead of a crop.
• Gameplay recordings — the visual analysis is what finds the moment; the transcript may be nearly empty.
• Webinars and talks — the transcript carries most of the weight here.
• New uploads from a channel — a YouTube feed trigger can start this on every new VOD, without an API key.