11 Skills to Improve Your YouTube Content The YouTube agent skill

Decide which part of the workflow you need help with

What problem does it solve?

You have a topic but no opening line. Your script is ready, yet the title and thumbnail repeat each other. A published video underperforms, and you do not know what to change first. This collection separates those tasks into 11 Claude skills: write from your source material, develop title and thumbnail pairings, and review a video using its transcript and retention data.

The main outputs are copy, briefs, and analysis. For a video about choosing a store location, you might first develop the explanation, mark where a map, comparison, or location shot belongs, and then record, edit, or animate those scenes with a production tool.

How does it compare with similar entries on this site?

Entry	Main purpose	Where this collection fits
Easel: Plan, Create, and Publish Social Content	Organize topics, image posts, videos, and publishing workflows in a local workbench	Focus on YouTube scripts, packaging, and performance reviews
16 Styles to Turn Ideas into Explainer Animations	Compare visual styles and build a short explainer animation	Establish the explanation and on-screen plan before making the animation
Turn a Script into a Hand-Drawn Explainer	Produce a narrated hand-drawn video with subtitles	Refine the script and title before handing the content to a video workflow

If your script is clear and you only need visuals, go straight to one of the animation skills. Use this collection when the unresolved question is what to say, how to open, or where viewers stop watching.

Choose from the 11 skills by task

Enter the /yt-… commands in Claude Code chat, not a terminal. Pick the task you need now; there is no requirement to run the entire set in order.

Skill	Output	What to prepare
yt-script	Opening options, a spoken script, and on-screen notes	Topic, audience, sources, and speaking style
yt-package	Title and thumbnail pairings, plus a thumbnail brief	The main promise or approved script
yt-edit	Suggested cuts for pauses, fillers, and retakes	A timestamped transcript
yt-comment	Comment triage, reply drafts, and a pin recommendation	Comments and your usual reply style
yt-plan	A schedule that fits your available time	Weekly working hours and existing material
yt-viral	Videos that outperform their own channel’s usual view levels	Public titles, view counts, and channel identifiers; at least 4 videos per channel
yt-retention	Retention drops and a prioritized revision	Retention CSV, with a transcript for content comparison
yt-shorts	Standalone short-video candidates, time ranges, and new openings	A timestamped long-video transcript
yt-seo	Description, search queries, and tag suggestions	Video content, audience, and relevant links
yt-chapters	Chapter boundaries and draft titles to revise	Timestamped subtitles
yt-audit	One change to prioritize for the channel	Recent titles, thumbnails, openings, and available analytics

yt-edit and yt-shorts recommend cuts and time ranges; you apply them in an editor. yt-package writes a thumbnail brief rather than generating an image. The collection does not upload videos or post comments.

Install the skills and set your speaking style

Copy the complete skill folders

This tutorial follows the official Claude Code copy installation. Prepare Git and Python 3, then run in a terminal:

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

# From the same terminal directory, copy the speaking-style template
mkdir -p ~/.claude/youtube
cp youtube-agent-skill/templates/voice.md ~/.claude/youtube/voice.md

You should have 11 yt- folders under ~/.claude/skills/, retaining each SKILL.md and its bundled files. Copying only the instructions loses dependencies: the chapter and retention tools read code from neighboring skills, and opening scores require hooks.json. The six Python tools use the standard library, with no additional Python packages.

Make the writing sound like you

Open ~/.claude/youtube/voice.md and describe your audience, usual phrases, words you avoid, claims you will not make, and video format. Fill in the existing headings in your working language. Include real transcripts if you already have videos. Most writing skills read this file first; when it is missing, they ask for your own videos to learn your style.

For this exercise, you could specify first-time store owners as the audience, direct language, concrete situations instead of jargon, and no invented business experience, footfall figures, or earnings. The file records writing and speaking preferences; this step needs no voice-generation service.

Open a new Claude Code session and enter /yt-script with your topic. Confirm it can read your profile, receive your material, and start proposing openings. If the command is missing, check that the folders are in the skills directory used by your Claude Code installation, then reopen the session.

Develop a script, then pair the title and thumbnail

Narrow the topic to one question

Choosing a store location is a large subject. Start with one question: why is foot traffic alone not enough? The learning goal is to distinguish people passing a shop from people likely to enter and buy. Prepare the store type, the method you want to teach, and supporting examples. If you have no case study, explicitly use a hypothetical teaching scenario.

The following requests were written for this tutorial; script creation and video production have not been run. Replace the bracketed information and send this in Claude Code:

/yt-script
Write an educational video script for someone preparing to open their first store.
Topic: why shouldn't you choose a location on foot traffic alone?
Store type: [type]. Sources: [file paths or text].
Explain only the difference between people passing and people likely to enter and buy.
Show opening options and the expectation each creates.
Wait for my choice before writing the full script.

The skill drafts 5 openings using a library of 21 structures, scores them, and shows the best two. The score checks English-language features such as specificity, direct address, stakes, curiosity, and brevity. It is a review aid, not a prediction of views. For a Chinese script, ask Claude to explain the choices in Chinese and judge their actual clarity; the bundled English word rules cannot reliably rank Chinese openings.

Choose an opening by asking what the rest of the video must deliver. “Is heavy foot traffic enough to justify the rent?” needs a decision method later in the script. A discussion of rent alone would not fulfill that promise. Once you choose, ask for a script that moves from situation to comparison, decision method, and conclusion, with on-screen notes for each section.

Read the first draft aloud. Does the explanation work without the visuals? Does each planned visual add meaning? Can you trace figures and cases to the supplied sources? Treat the skill’s 150-words-per-minute estimate as an English draft estimate, and check timing by reading; it is not a Chinese timing rule.

Choose the title and thumbnail together

After approving the main explanation, send:

/yt-package
Develop title and thumbnail pairings for the approved script.
The title should name the viewer's problem; the thumbnail should add a scene or contrast,
not repeat the same words.
Use only facts in the script. Add no invented personal results or unsupported numbers.
Follow the skill's candidate workflow, recommend three pairings,
and show which part of the script delivers each promise.
Give thumbnail composition, subject, and text instructions.
Wait for my choice before image production.

A title about checking who will enter a store might pair with a contrast between people outside two storefronts. That is a creative direction; adapt it to your store type and sources. At a reduced thumbnail size, the subject should remain recognizable and the text should add to the title. When working in Chinese, manually review duplication, wording, and readability rather than applying English word-count and vocabulary checks directly.

Changing the opening question can also require changing the body, title, and thumbnail. Check that the revised script answers the new question. If you only want to change thumbnail layout, explicitly preserve the approved script and facts. Once the writing is ready, use one of the animation workflows linked above to produce visuals.

Use transcripts and retention data to improve the next video

After recording, give yt-edit timestamped SRT, VTT, or transcription JSON and ask for suggested cuts. Check them against the footage: a quiet demonstration may need its pauses. Do not remove every period without speech. Filler and restart detection also use English-oriented rules; review Chinese transcripts manually.

Once a published video has data, prepare its audience-retention CSV and matching subtitles. The retention file needs at least 8 valid data points, arranged as position and retention percentage. If positions are percentages, provide the full video duration to convert them to seconds. Send the files to Claude Code:

/yt-retention
Analyze [retention CSV path] alongside [subtitle path].
Full video duration: [actual seconds]. Position units: [seconds or percent].
Check the file format and timeline first.
Find one drop worth investigating, quote the nearby transcript,
and recommend one concrete revision for the next version.
Separate observations from possible explanations; do not infer causation without evidence.

Check that the report’s positions align with the right subtitle passages, then watch those moments. A retention curve identifies places to investigate; it cannot by itself establish that a sentence, shot, or transition caused viewers to leave.

To repurpose a longer video, use yt-shorts to find passages that stand alone, then review the new opening, ending, and vertical crop. It supplies candidates and adaptation notes; you still edit the footage. For your own topic, reuse the sequence: define the audience’s question, write, pair title and thumbnail, produce, then revise using data. Replace the sources, speaking profile, and analytics with your own.

The collection is MIT licensed. See the repository README for the complete command list and updates.