The free YouTube agent skill for Claude, and the Facebook reel that pitched it

Get it without commenting for a DM

Skip the comment. The reel asks for "YOUTUBE" under the post so its page can message you an install guide, but you don't need one: the repository is public on GitHub.

# Paste this into Claude:
https://github.com/Jakeschincariol/youtube-agent-skill

Install this skill, then confirm /yt-script works.

# Or do it yourself, in Claude Code:
git clone https://github.com/Jakeschincariol/youtube-agent-skill.git
cp -r youtube-agent-skill/skills/yt-* ~/.claude/skills/

# Or as a plugin:
/plugin marketplace add Jakeschincariol/youtube-agent-skill
/plugin install youtube-agent

1. Use any one of the three routes above. The two plugin lines go inside Claude Code. Six skills run small Python 3 scripts, the reason the README calls them "not just prompts". The scripts need nothing beyond Python's standard library.
2. To keep the skills in one project, copy the same eleven yt- folders into that repository's .claude/skills/ folder instead.
3. Copy templates/voice.md to ~/.claude/youtube/voice.md and fill it in. The README budgets ten minutes for this, because "you have to say the words out loud." You can also send Claude three of your own videos and ask it to write the file.
4. Type /yt-script with the idea for your next video, and check that two scored hooks, the opening lines, come back. After the plugin route the command is /youtube-agent:yt-script, because Claude Code files a plugin's skills under the plugin's name.

How the virality engine finds a breakout

Raw views mostly measure how big a channel is. The engine's script divides each video's views by the median views of its own channel instead, so a small channel's surprise hit outranks a big channel's ordinary upload.

In the docstring's own case, 400,000 views on a channel whose median is 30,000 score about 13 times its usual. On a channel with two million subscribers, the same count is "a normal Tuesday".

One freak hit barely moves a median, where it drags an average up. The script also skips any channel with fewer than four videos, and says so.

Claude gathers the list the way you choose. The skill names yt-dlp --flat-playlist -J on a channel's address as the fastest way, and that needs yt-dlp, a separate program you install. The public page or a list typed by hand works too, but never your login: the skill sticks to public listings.

Each result also carries the name of the formula its title matches best, or Unclassified when none fits. That label "is a judgement about the title you can see, not a claim about why the video got its views," in the README's own words.

What the reel got right, and where it stretched

The reel comes from the Facebook creator Airi. GitHub lists Claude as co-author of the pack's one commit, dated 16 September 2026, so Claude helped write it. Nothing in the repository says Anthropic made or backs it, and Anthropic didn't release it.

Two frames of the reel, cropped to the video. We captured them from Facebook's page at about 15:10 UTC on 5 November 2026, when the reel had 274 reactions, 250 comments and 119 shares. Mindket translated the Chinese words on screen into English, and the frames as posted are on this post's Chinese page. The card in the first is drawn for the video: its folders differ from the GitHub page on this post's cover, and its one-line description differs from the repository's. Select a frame to open the reel.

Where the pack falls short

The hook score is the weakest part. The README admits as much: calibrated on 74 short-form hooks, it "separates a given creator's hits from their own misses barely at all." It also says "A low score is a reason to look again." Even the pack's own example for its Direct Address formula scores 41 by our hand count, in the WEAK band, and the classifier files it under another formula, The Superlative.

• Only six SKILL.md files tell Claude to read voice.md, though the README says every skill does: script, package, comment, plan, SEO and audit. The edit, viral, retention, Shorts and chapters skills skip it.
• "Nothing invents a number," the README promises, yet only those same six files tell Claude never to invent one. No code checks it.
• Nothing has been merged since the first and only commit, and there is no release. When we checked, GitHub showed 4.6k stars, 1.4k forks, 99 open issues and 13 open pull requests.

Last verified 5 November 2026, 15:35 UTC, against the GitHub page and the files under Sources. Nothing here is hands-on testing: no code from the repository was run, and the hook score is worked by hand from the published code.

Reference: what each of the eleven commands hands back

Besides its six Python scripts, the repository is plain text. It holds eleven SKILL.md files, one per command and each an Agent Skill, plus a voice template, two plugin manifests, a JSON file of hook formulas, the README and the licence.

• /yt-script turns one idea into five hooks written against 21 formulas, scores them with hookscore.py and keeps the best two. Then it writes the script beat by beat, with what is on screen and a runtime at 150 words a minute.
• /yt-package writes ten titles, runs them through title.py, checks them against the 60-character desktop cut and the 40-character mobile one, and shows the top three with their scores. Its thumbnail brief gives the words for the image, not the image itself.
• Give /yt-edit a timestamped transcript (SRT, VTT or Whisper JSON) and it lists cuts for dead air, filler and retakes, each with the start and end seconds you need to make it.
• /yt-comment sorts comments you paste into questions, corrections, praise and bait, drafts replies under 30 words to all but the bait, and names one comment to pin.
• Before it plans, /yt-plan asks how many hours you have and what is already half made. The plan holds one main video, one cheap one and three Shorts cut from the main one.
• /yt-viral ranks videos from channels in your niche by how far each beat its own channel. Its list holds one entry per video, with its channel, title, views, address and length. Run as python3 swipe.py collected.json --min 2.0, the script keeps videos at twice their channel's median or more. Without the flag it keeps anything from 1.5 times up.
• Download the audience-retention CSV from YouTube Studio, and /yt-retention names the single biggest leak: the first 30 seconds, a sudden drop or a slow slide.
• /yt-shorts finds stretches of 20 to 55 seconds in a long video that stand on their own, and writes a new first line for each.
• /yt-seo drafts the two description lines that show above "...more", tags for the spellings and tool names people type, and the three searches the video should win. It calls tags a weak signal, and fifteen plenty.
• /yt-chapters builds chapters from a transcript and checks YouTube's rules: the first at 00:00, at least three, each at least 10 seconds long.
• /yt-audit reads your last ten titles, your thumbnails at feed size, recent hooks and your upload rhythm, and ends on a single fix.

Reference: the pack's own example hook, scored by hand

hookscore.py rates a hook on specificity, address, stakes, curiosity and brevity, then counts the weakest of the five for 40% of the verdict. This is the example hooks.json gives for its Direct Address formula:

If you have under a thousand subscribers, this is the only video you need this week.

• It has no digit, no stake word such as lose or cost, and no curiosity word such as why or but.
• The Superlative wins because four formulas match once each, and the classifier takes the first one listed.

Frequently asked questions

Sources

1. Jake Schincariol: youtube-agent-skill on GitHub (the README and the repository page) (read November 5, 2026)https://github.com/Jakeschincariol/youtube-agent-skill
2. youtube-agent-skill: swipe.py, the virality engine (read November 5, 2026)https://github.com/Jakeschincariol/youtube-agent-skill/blob/main/skills/yt-viral/swipe.py
3. youtube-agent-skill: hookscore.py, the hook score (read November 5, 2026)https://github.com/Jakeschincariol/youtube-agent-skill/blob/main/skills/yt-script/hookscore.py
4. youtube-agent-skill: hooks.json, the 21 hook formulas (read November 5, 2026)https://github.com/Jakeschincariol/youtube-agent-skill/blob/main/skills/yt-script/hooks.json
5. youtube-agent-skill: LICENSE (MIT) (read November 5, 2026)https://github.com/Jakeschincariol/youtube-agent-skill/blob/main/LICENSE
6. Airi on Facebook: the reel on the YouTube agent skill (in Chinese) (read November 5, 2026)https://www.facebook.com/reel/1677865394050843
7. Prompt Astra on YouTube: Claude YouTube Agent setup and 11 skills (read November 5, 2026)https://www.youtube.com/watch?v=acdGnoy6Wl4
8. Jake Schincariol: chatgpt-youtube-agent-skill on GitHub (read November 5, 2026)https://github.com/Jakeschincariol/chatgpt-youtube-agent-skill
9. Jake Schincariol on GitHub: his public repositories (read November 5, 2026)https://github.com/Jakeschincariol
10. Python documentation: Using Python on Windows (read November 5, 2026)https://docs.python.org/3/using/windows.html
11. Claude Code docs: Extend Claude with skills (read November 5, 2026)https://code.claude.com/docs/en/skills