Image: GitHub
Why it mattersThe review page is the piece most agent workflows do not have, because it lets a non-technical user point at a specific frame instead of describing the problem in chat.
Giving feedback to an AI agent in a chat window is painful when the problem is on screen. Peter Yang's new /video-short skill, open-sourced on his GitHub, tries to fix that for one specific task: turning a raw recording of a person talking into a published short with captions, B-roll, motion graphics and music, through an agent the user already pays for.
The repository has gathered 99 stars in the four days since Yang pushed it. The install line, from the README, is a single paste into Claude Code, ChatGPT Desktop or another agent: "Install the /video-short skill globally from https://github.com/petergyang/video-short."
What the skill actually does
The skill supports four short layouts the README lists: A-roll (the speaker full screen), Punch (a tighter crop of the speaker), Split (a visual on top, speaker below), and B-roll (a visual full screen while the speaker keeps talking). It can add captions matched word-for-word to the speech, HTML motion graphics, B-roll pulled from the web, and generated music through Google's Lyria. A Gemini API key with billing is required only for the generated video and music; everything else runs on the user's own machine.
The piece worth pointing at is the review page. After the agent produces a draft edit, the skill opens a page that plays the short on a scene timeline. The user pauses on any frame and leaves comments pinned to it. The agent applies those comments and renders a new version. Yang's framing of the problem the review page solves, from the README: "making a great one by hand still takes hours. And when you ask AI to help, giving feedback in chat is painful because you can't point at the exact frame you want to fix."
The dependencies, kept on the user's machine
The skill installs ffmpeg to cut and encode video, yt-dlp to pull B-roll from YouTube and other sites, Python 3.10 or newer, and Node.js 22 or newer. It also installs HyperFrames, a Heygen-released open-source tool that transcribes the speech locally with NVIDIA's Parakeet model and does the render. Plan for about 1 GB of disk, mostly for the speech model.
Each render saves a new version to the project folder, so the user can compare v1 and v2 by playing them back. A smaller 720p copy appears in renders/share/ for sending from a phone.
For a content team
For anyone running a small content operation, this is the pattern worth copying rather than the tool to adopt as-is. Pick one specific workflow, write the plan the agent should follow as a Markdown skill, and build one review surface the agent and the human share: a page with the artifact and comments pinned to it, not a chat thread. The skill is MIT-licensed, so the code of the review page is available to read and lift the shape from.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
