Skill 09

Turn a long video into Shorts

Turn a long video (a gaming stream, a livestream, a podcast, a film, a lecture) into several vertical Shorts, Reels or TikToks with cuttingfloor - one import, the moments the user wants cut to 15 to 60 seconds with the hook first, each panned and scanned to full-screen 9:16, captioned, and exported once. Use when someone asks for Shorts, Reels, TikToks or clips from a long video, or to repurpose, clip or cut down a stream, episode or film.

LiveWatch it live now Kazi gives Big Buck Bunny a happier ending.Kazi is one way in: ChatGPT, Claude or any MCP app connects the same way.Free · a few minutes · no sign-up
01 / Copy it

How do I give this to my agent?

  • Kazi Paste it into a chat and ask Kazi to save it as a playbook. It is already in the format Kazi uses.
  • Claude Paste it into a chat, or into a project’s instructions so every chat in that project follows it.
  • ChatGPT Paste it into a chat, or into a project’s instructions so every chat in that project follows it.
  • Connected to cuttingfloor? Your agent can find this skill itself. Just ask for the job in your own words.
The whole skill, 40 lines

The techniques are their own skills: pan-and-scan (with subject-tracking) for the vertical picture, shorts-style-captions for the words. This adds what is particular to Shorts.

When to use

  • A long video and a request for several short vertical clips from it.

Prerequisites

  • The cuttingfloor MCP server is connected.
  • The video is reachable as a file: a direct public link, an upload, or the user's upload page. A video site's page (YouTube, Twitch) is not a file; ask the user for the file itself.

Steps

  1. The assistant should import the video once and start, together, the analyses every Short will share: transcript (model base), scenes {threshold: 10} and the subject positions pan-and-scan needs (faces, or track). One job.get {job_ids} waits for all of them.
  2. The assistant should choose the moments the user asked for, using the transcript and media.seek in plain words for what they describe ("the clutch play", "when she laughs"), and check each on a 16-frame render.frames sheet of the media (free). Each Short is 15 to 60 seconds and opens on its hook: the line or action that makes someone stop scrolling, in the first two seconds.
  3. The assistant should build each Short on its own 1080×1920 timeline at the video's frame rate, cut with clip.add {source: {in, out}}, and pan and scan it to full-screen vertical (pan-and-scan).
  4. The assistant should caption it from the transcript with caption.add {in_clip} in transcript seconds, 3 to 6 words a segment with word timings, leaving out bracketed sound tags such as [Music playing]; for a styled band or box, shorts-style-captions.
  5. The assistant should look at one frame per shot, then start every Short's export.video (mp4_h264) together, and hand back each download link and page link with where in the long video it came from.

Decision table

SituationDo
A gaming stream with a facecam in a cornerA stack: the facecam cropped (crop) across the top third, the gameplay panned below
Music drowns the speech in the transcriptRetime or retype those captions; caption no sound tags
The user names the momentsUse their times; search only what they didn't name
Many ShortsOne import and one set of analyses; one timeline and one export each

Required behavior

  • One import and one set of analyses for all Shorts.
  • Every Short opens on its hook and says where it came from.