Clipify: An AI Highlight Clip Finder Built as a Claude Code Skill

Finding the good three seconds in a forty-minute recording is the part of clipping that eats the most time. An AI highlight clip finder called Clipify handles that search step directly: point it at a long video file and it transcribes the audio, scans for punchlines, reversals, and awkward pauses, and hands back three to five candidate moments with timestamps before any cutting starts.

The project is built by developer Louise de Sadeleer and lives on GitHub at clipify. It ships as a Claude Code skill, installed by cloning the repo into ~/.claude/skills/clipify, which makes /clipify available as a slash command. There is no app, no dashboard, and no account. It is a skill file plus a set of Python and ffmpeg scripts that an AI coding agent runs on the machine it's given.

What Clipify actually does

The workflow runs in a fixed sequence once a video path is provided:

1. Transcribe and propose: Whisper transcribes the video, and the skill scans the transcript for clip-worthy segments, surfacing 3 to 5 candidates with suggested titles.
2. Pick a moment and a format: the user chooses which candidate to cut, then picks 9:16, 16:9, or 1:1.
3. Reframe: for 16:9 sources going to 9:16 with two people on screen, it either pans between whoever is speaking or falls back to split-screen. There is no face-detection model involved. The skill eyeballs each face's mouth-and-chin area as a rectangle on one sample frame, then uses ffmpeg to measure per-frame motion energy in each rectangle. Whichever rectangle is moving more at a given moment is treated as the active speaker, and that timeline drives a hard-cut crop expression.
4. Caption: it burns word-by-word captions in opus-style (bold white text, yellow highlight on the active word), with karaoke and minimal styles also available, or a reference image can be pasted in to match a different caption look.

Finished clips land in a clipify_out/ folder next to the source video. The README states a rough benchmark of about 20 seconds of processing for a 20-second clip on Apple Silicon, and everything runs on-device: no cloud API calls, no OpenCV dependency.

How it differs from b-roll-finder

Clipify shares an author with b-roll-finder, an earlier Claude Code skill from the same maker, and the two get mixed up easily because both operate on video and both lean on Whisper transcription. They solve different problems. b-roll-finder starts from a script or narration track that already exists and sources accurate cutaway footage to match what's being said, useful for podcast intros, tutorials, or explainer video that needs illustrative b-roll. Clipify starts from raw, unedited long-form footage, usually an interview or a two-person conversation, and its job is to find and cut the highlight moments themselves, then reframe them to vertical. One finds supporting footage for a script. The other finds the script-worthy footage in the first place. A creator doing both long-form and shorts from the same recordings would plausibly use both.

Setup and what it costs to run

Requirements are: macOS (it uses VideoToolbox for hardware-accelerated decoding, though the README notes the videotoolbox flags can be stripped to run on Linux or Windows), Claude Code itself, ffmpeg with libx264, the whisper Python package, and numpy. All of that installs with brew and pip. No API keys, no signup, no per-clip cost. The repository is MIT licensed, so the scripts and the skill prompt are both fully readable and forkable.

That dependency list puts a floor on who this suits. It is not a drag-and-drop app for someone who has never opened a terminal. It assumes comfort running an agent skill and installing a couple of CLI tools, in exchange for skipping the subscription and upload queue of a hosted auto-clip service.

Who this is for

The README is explicit that Clipify is built for talking-head dialogue: interviews, podcasts, and two-person setups where a static camera holds both faces in frame. That's also where the motion-based speaker detection works best, since it depends on one person's mouth moving more than the other's at any given moment. A single speaker facing camera, a screen-recording, or a scene with more than two people in frame is outside what the pan logic was designed to track, and results there should be treated as unverified until tested.

It's aimed at solo creators, podcast editors, and small teams who already clip their own long-form recordings for LinkedIn, TikTok, or Reels and want the moment-finding and reframing done locally instead of paying for a hosted clipping tool or scrubbing footage by hand.

Current traction

As of today, the repository shows 519 stars and 56 forks on GitHub. It was created in May 2026, and the scan that first surfaced it recorded 502 stars with no velocity marker yet, so growth has been modest but steady since first appearance. It is licensed under MIT.

Source: github.com/louisedesadeleer/clipify