Young Tries AI

Turn one long video into five Shorts with Clipper inside Claude Code

This guide is for creators comfortable with the terminal who want vertical Shorts from a long video. You end up with captioned 1080x1920 clips and ready-to-paste posting text.

Video and imageFor creators

Before you start

  • Install Claude Code.
  • Install uv.
  • Install ffmpeg.
  • Free up about 3 GB of disk for the speech model and packages.
  • Have a video you own or have permission to clip: a YouTube link or a local file ffmpeg can read.
  • Tested on Windows; macOS and Linux should work but are untested. An NVIDIA GPU is optional; on CPU a smaller model is used, slower.
  1. Pick the line for your system and run it in a terminal: winget on Windows, brew on macOS, apt on Ubuntu or Debian. The macOS line is shown here.

    Run in the terminal
    brew install ffmpeg
  2. Open Claude Code and add the Clipper marketplace by pasting the line below.

    Paste into Claude Code
    /plugin marketplace add gouthamk16/clipper
  3. Still inside Claude Code, install Clipper. The first clip takes a few extra minutes while the engine installs its Python packages and downloads the speech model. After that it starts straight away.

    Paste into Claude Code
    /plugin install clipper@clipper
  4. Run the doctor command below, or simply ask Claude to check your setup.

    Run the setup check
    uv run --project clipper doctor
  5. Run the clip command with a local file. For a YouTube link, put the link in its place. Add a number at the end for how many clips you want, for example 3.

    Paste into Claude Code
    /clipper:clip my-podcast-episode-12.mp4
  6. Talk to it like you would talk to an editor. Claude changes the pick and re-renders, and nothing is downloaded or transcribed twice.

    Paste into Claude Code
    clip the three funniest moments from this stream: make #2 shorter and give #4 a punchier hook
  7. Open the clipper-output folder. Each clip is a vertical 1080x1920 MP4 with captions burned in, plus a .txt file with title, description and hashtags to paste into YouTube, Instagram or TikTok. The clips.json file lists every clip's timing and text.

  8. Ask Clipper for this month's fastest-growing long videos on your topics, ranked by views per day.

    Paste into Claude Code
    /clipper:find gaming streams, tech podcasts

Watch out for

  • Clip only what you own or have permission to clip. Reposting someone else's video can bring copyright claims or takedowns.
  • Since 1 October 2026, YouTube gives less reach to channels that mostly re-upload other creators' videos.
  • The authors saw clips of 21 and 34 seconds reach the Shorts feed and nothing over 38 seconds get reach. Clipper aims for 15 to 35 seconds.
  • Expect the first run to be slow, because it installs Python packages and downloads the speech model (about 3 GB).
  • Without an NVIDIA GPU, a smaller speech model is used and transcription takes longer.

Questions

Does it upload my video to a clipping service?
No. It runs on your machine, with no uploads to a clipping service, no watermark and no per-minute credits. Claude reads the transcript to pick moments and write hooks.
What does each clip include?
Each clip has a clean cut of the moment, a hook line on screen in the first three seconds, word-by-word captions, and a crop that follows whoever is talking. Pauses longer than 0.7 seconds are removed, and a posting file holds title, description and hashtags.
How does it choose moments?
Claude reads every sentence of the transcript. It uses YouTube's most replayed parts, viewer timestamps in comments and audio energy spikes as leads. It picks clips that hook in the first second, stand on their own and end on something.
What is the TYPESAFE_API_KEY for?
It is optional. With a TypeSafe API key set as TYPESAFE_API_KEY, Clipper can pick and render clips without Claude, scoring candidates with Jev. Without a key, Claude does the picking.

Sources

Based on