Peter Yang releases free /video-short skill for agent-assisted video editing

• A free /video-short skill turns raw recordings into shorts with B-roll, captions, graphics and optional generated media, as Peter Yang demonstrated in his launch demo.
• Frame-pinned comments let an agent revise a draft, a workflow Yang compared to Google Docs comments.
• The intended input is a newly recorded short, as Yang clarified in a reply.

In the tutorial, the agent fixes shots that crop off Yang’s head. A single style.json file carries fonts, colors and caption styling across future shorts.

The MIT-licensed repository packages the editing workflow as a skill for an existing AI agent. Its README names Claude, ChatGPT and Gemini as compatible options.

An edit can switch between four layouts:

1. A-roll: The speaker fills the screen.
2. Punch: A tighter crop of the speaker for emphasis.
3. Split: A visual above, with the speaker below.
4. B-roll: Full-screen visuals while the speaker’s audio continues.

The skill also adds:

• B-roll footage: Sourced online, or from personal footage, which Yang confirmed the agent can use.
• Motion graphics: Written in HTML.
• Captions: Matched word for word to the speech.
• Generated video and music: Optional additions through Gemini APIs.

Automatic highlight extraction from long recordings falls outside the intended workflow, as Yang explained in a follow-up.

Yang lays out four stages in his written walkthrough:

1. Install. Add the skill once through the agent, or use the documented terminal command: npx skills add petergyang/video-short --skill video-short --global --yes.
2. Upload. Run /video-short edit my video at [path/to/video.mp4]. The skill transcribes the recording and trims retakes, false starts and dead air.
3. Plan. The agent researches the topic and proposes a scene-by-scene plan for approval, pairing spoken words with visuals and proposed generated assets. The user can request changes before approving.
4. Iterate. A draft opens in the review page, where submitted feedback feeds the next revision.

Yang calls the review page “Google Docs comments for video”. Notes attach to the paused frame, giving the agent a specific place to make the requested change.

The comment UI is the keeper here. Yang also says in the tutorial that several feedback rounds preceded the draft shown in the recorded review.

Each render saves a new version for comparison, such as v1 and v2. The repository also documents a smaller 720p sharing copy saved under renders/share/.

Yang built the simpler interface over HyperFrames, which he confirmed supplies the backend.

On first run, the skill checks dependencies and offers to install missing components. The setup documentation specifies:

• FFmpeg: Cuts and encodes video.
• yt-dlp: Retrieves B-roll from YouTube, X and other sites.
• Python 3.10 or newer: Runs the editing scripts.
• Node.js 22 or newer: Supports HyperFrames, which transcribes locally with the Parakeet speech model and renders the edit.

The setup needs around 1 GB of disk space, mostly for the speech model.

The skill is free, while optional media generation requires a Gemini API key with billing enabled. The README lists these generators:

• Nano Banana: Images.
• Omni Flash: Video clips.
• Lyria: Music.

The key is supplied through GEMINI_API_KEY. Creators can also supply their own video and music instead of generating those assets.

Yang’s launch write-up discloses a Google Gemini partnership; the tutorial is sponsored by Gemini.