An MIT YouTube agent skill with 401 stars gives Claude eleven commands for a creator's weekly workflow

Running a YouTube channel alone is a stack of small, repeating jobs: write a hook, name the thumbnail, cut the long video into Shorts, read the retention curve. A new agent skill by Jake Schincariol, released on GitHub on 16 September 2026 under MIT and now at 401 stars, turns each one of those into a Claude Code command.

The repository is Jakeschincariol/youtube-agent-skill. It carries no outside measurement: the author's own description is that the pack "generates scripts, titles, thumbnails and more for YouTube videos." It does not publish, and it does not have a credential for the YouTube API. It is a prompt kit that assumes the creator still presses the buttons.

The eleven commands, as the README names them

The commands are plain slash entries in Claude Code: /yt-script writes a video script with hook options drawn from a formula set. /yt-package pairs a title and a thumbnail concept. /yt-edit turns a transcript into an edit decision list. /yt-comment triages comments and drafts replies. /yt-plan schedules a week of content. /yt-viral reviews what viral videos in a niche have in common. /yt-retention reads an audience retention curve and names where viewers leave. /yt-shorts finds the Shorts inside a long video. /yt-seo writes descriptions and tags. /yt-chapters produces YouTube-compliant chapters. /yt-audit runs a channel-level review.

That list covers the week of a creator who posts twice: a script, a package, an edit, a shorts pass, and a retention review on the previous upload.

Where the hook scorer is honest

One command sits apart from the rest. /yt-script produces five candidate hooks per idea, scored by a Python tool called hookscore.py. The README's own account of calibration is the sentence to quote: "The panel was built and calibrated against 74 real short-form hooks (the first 15 seconds of auto-captions, top-8 and bottom-8 by views across five channels)."

The author then states its own limit. The tool "separates deliberately bad hooks from real ones well" but is weak at telling a creator's own hits from misses. In other words, 74 hooks across five channels is enough to catch something that would not land at all, and not enough to pick the winner from a batch of plausible ones. A creator should treat the score as a filter, not a predictor.

For a solo team

The pack's value sits in the shape of the commands rather than in a cleverer model. Each one is prompt plus structure: the retention command expects a CSV, the chapters command returns compliant timestamps, the shorts command returns a list of cuts rather than prose about them. That is the saving. The repetitive work on a channel is turning formless content into a format, and the agent cannot do the YouTube-side work either way. A product marketer whose company already uses Claude Code gets eleven ready-made jobs; a solo creator who was going to put the whole month into a chat window gets a framework for doing it.

Source

Primary source: Jakeschincariol/youtube-agent-skill on GitHub.