YouTube Agent: free Claude skills that prepare your YouTube video

YouTube Agent is a free, open source pack of 11 Claude skills that helps you prepare a YouTube video: writing the script, the title and thumbnail idea, the cut list, replying to comments, and reading your channel's numbers.

Three things to know before you start:

• The pack writes and prepares, and you do the cutting, designing, and publishing. It doesn't open the video file, design an image, or upload anything to YouTube.
• Claude does the writing, so scripts in other languages work too, including Arabic. But most of the six check tools are built around English words, so outside English some of their checks fail. We cover this in its own section below.
• The pack is free and doesn't ask for an API key, but Claude usage still depends on your subscription.

We wrote this article from the project page and its files on November 4, 2026 (version 1.0). We read the code of all six tools, but we did not run them, and we did not try the pack on a real channel. Commands change between versions, so check the project page before you run anything.

github.com/Jakeschincariol/youtube-agent-skill

How to install it

In Claude Code, install it as a plugin:

/plugin marketplace add Jakeschincariol/youtube-agent-skill
/plugin install youtube-agent

Or copy the folders yourself:

git clone https://github.com/Jakeschincariol/youtube-agent-skill.git
cp -r youtube-agent-skill/skills/yt-* ~/.claude/skills/

Then try the first command and write your video idea after it:

/yt-script

The tools only need Python 3, with no libraries to install.

Start with your voice file

Most of the skills first read a file called voice.md. It covers who you're talking to, how you actually talk, the words you use and the ones you don't, and what you won't claim. If a skill doesn't find the file, it asks you for three of your videos to write it from. The project page says this file is worth ten minutes of your time, because you'll be reading the script in your own voice on camera.

Here are the transcripts of three of my videos. Use them to write the voice file following the templates/voice.md template in the YouTube Agent pack, and save it to ~/.claude/youtube/voice.md. Capture the way I talk and my dialect exactly as they are, and don't improve them.

The 11 skills

You run each skill with a command that starts with a slash (/) before its name, or you ask Claude for it by name.

• yt-script: turns the idea into a script. It writes five hooks from 21 formats, shows you the best two, then writes the full script with what appears on screen in each section.
• yt-package: writes ten titles and shows you the best three, then describes the thumbnail: the expression, the composition, and the three words. It doesn't design the image itself.
• yt-edit: takes the timestamped transcript and gives you a list of what to cut: silences, filler, and repeated sentences. You apply it in your editing software.
• yt-comment: you paste in the comments, and it sorts them into questions, corrections, praise, and provocation, writes replies in your voice, and suggests one comment to pin.
• yt-plan: a weekly plan that fits your available hours: one main video, one lighter video, and three Shorts.
• yt-viral: ranks videos in your niche by how far each one beats its channel's median views, rather than by channel size. It needs at least four videos from each channel, and it reads public lists only.
• yt-retention: reads the audience retention file from YouTube Studio and identifies where viewers left.
• yt-shorts: picks clips from a long video's transcript that work as Shorts, and writes a new first line for each one. It doesn't cut the video.
• yt-seo: writes the description, the tags worth using, and three search phrases the video should show up for.
• yt-chapters: writes the video's chapters from the transcript and checks YouTube's rules: they start at 00:00, there are at least three chapters, and each chapter is at least 10 seconds.
• yt-audit: reviews the whole channel and ends with one fix to start with, instead of a long list.

What works outside English?

Skills are instructions for Claude, so you can ask them to write in Arabic or another language. The six tools are different. They're fixed Python programs the skills call, and most of them only recognize English letters. If you write in English, the limits below don't apply. The project page doesn't mention any of this. We first wrote this article for creators who work in Arabic, and this is what we found in the code for Arabic text:

• hookscore.py (hook scoring): it doesn't read Arabic words. Worked out from the code, any Arabic hook without numbers scores between 9 and 25 out of 100, depending on punctuation, however good it is. Ignore that number.
• title.py (title check): the title length measurement works. The checks for vague words and for repetition between the title and the thumbnail don't, and it may tell you the words are different when they're identical.
• deadair.py (cut list): silences work because they depend on timing. It only looks for filler and repeated sentences in English.
• chapters.py (chapters): it sets the breaks from pauses alone and doesn't suggest an Arabic chapter title, so write the titles yourself.
• swipe.py (video ranking): the ranking works because it's numbers. The title format classification is based on English words.
• retention.py (audience retention): it works, because it reads numbers.

The fix is simple: ask Claude to do this check itself instead of relying on the tool's number.

Use the yt-script skill. The idea: how to choose your channel's first microphone. Write the script in my voice, following voice.md. hookscore.py only reads English words, so if the script isn't in English, don't rely on its number. Rate the five hooks yourself and explain the reason for each rating.

Use the yt-edit skill on the transcript file transcript.srt. If the transcript isn't in English, deadair.py only catches silences, so read the transcript yourself and add the filler words and repeated sentences with their timestamps. Don't edit the video file.

What to watch out for

The score isn't a prediction, even in English. The project's owner says the hook score tells deliberately bad hooks apart from real ones, but it barely separates a hit from a flop by the same creator. A low score is a reason to rethink, and a high one isn't a promise.

The format only describes the title. When yt-viral says a title used a certain format, that judges the visible words only. It doesn't explain why the video succeeded. The list of videos has to be collected first, with a tool like yt-dlp, from the public page, or by hand.

Nothing gets published automatically. Each skill ends with text you copy, and you do the uploading.

Review before you publish. Every skill says not to make up a number or a source, and to ask you if it needs one. Those are only instructions and they guarantee nothing, so review the script before you read it on camera.

Follow us at @nexcodekw.