YouTube SEO Skills

YouTube SEO Skills is a suite of eight Claude Code skills — one orchestrator and seven specialists — that audit a channel, diagnose why a video underperforms, rewrite metadata, cluster keywords, score thumbnails with computer vision, and extract patterns from competitors. It models YouTube's modern recommender — session watch time, persona matching, surface-specific CTR — rather than treating YouTube as a keyword-match search engine.

What is YouTube SEO?

YouTube SEO is the practice of getting videos distributed by YouTube's recommendation and search systems. On modern YouTube, keyword relevance is a minor factor: the system optimizes for session watch time, so APV (Average Percentage Viewed), retention shape, and surface-specific CTR — Browse, Search, and Suggested each behave differently — dominate. Metadata still matters for retrieval and entity association, but it cannot rescue a video people stop watching at 0:30.

The eight skills

• youtube-seo — the orchestrator. Routes intent to the right sub-skill and holds the shared ranking model, the data-source matrix, and niche benchmarks.
• youtube-seo-audit — full channel audit with parallel delegation to the other skills, topical authority scoring, traffic-source mix, upload-velocity consistency, and a prioritized action plan.
• youtube-seo-video — single-video deep dive: retention-curve diagnosis, intro hook strength, APV against the niche benchmark, entity coverage, Key Moments chapter schema, and audio loudness.
• youtube-seo-optimize — paste-ready output: Browse-primary and Search-primary title variants, entity-rich descriptions, tags, chapters, a 15-second hook script, translated metadata for top markets, pinned comment, and end-screen plan.
• youtube-seo-keywords — search-intent classification, Knowledge Graph entity mapping, seasonality curves, a refined opportunity formula, and topic clusters built for topical authority.
• youtube-seo-thumbnail — OpenCV face and emotion detection, dominant-color k-means, contrast and text-region checks, 120px mobile legibility, CLIP-embedding similarity against the live SERP grid, and a native Test & Compare plan.
• youtube-seo-channel — identity and About copy, channel keywords, playlist session chains, bingeable series structure, trailer script, community-tab strategy, and translations.
• youtube-seo-competitor — competitor detection, hook and title pattern extraction from transcripts, retention strategy inferred from pacing, format-market-fit gaps, and an entity gap matrix.

The six-tier ranking model

The shared model in youtube-seo/SKILL.md orders signals by real influence: Tier 1 watch-time and session signals (APV, AVD, 0:30 retention, retention cliffs, returning-viewer rate, per-surface CTR), Tier 2 metadata and semantic signals, Tier 3 thumbnail pre-click signals, Tier 4 engagement and velocity, Tier 5 channel-level authority and cadence, Tier 6 technical and safety signals. Plus a separate Shorts ruleset for the 0.5–1.0 second hook, loopability, caption overlay, and trending audio.

It carries concrete thresholds rather than adjectives: APV at or above 55% is excellent and below 35% means suppressed distribution, 0:30 retention should hold 70% of viewers, any drop over 10% in under 5 seconds is a structural cliff, audio should land near −14 LUFS, and a benchmark table gives CTR, APV, and like-ratio medians across eight niches from tech review to Shorts.

First-party data only

The suite deliberately refuses third-party analytics scrapers — VidIQ, TubeBuddy, Ahrefs, SocialBlade — because they rate-limit, change their HTML, and break runs. It works down a ten-step priority chain instead: YouTube Data API v3, your Studio CSV exports, yt-dlp, the YouTube suggest API, WebFetch, Google Trends, Whisper, FFmpeg, and OpenCV with CLIP. Each step degrades gracefully, and when a Tier 1 signal such as real CTR or a retention curve is genuinely missing, the skills ask you for it rather than inventing a number.

Who it is for

Creators, channel managers, and video marketers who already use Claude Code and want diagnosis backed by their own Studio data — not a score from a browser extension.

What you can build

A scored channel audit with a ranked fix list, a diagnosis of why one video stalled, paste-ready titles and descriptions for a launch, a keyword cluster that builds topical authority, a thumbnail scored against the actual search results it will sit in, and a competitor teardown that names the format gap you can take.

Why it matters

Most YouTube SEO advice optimizes the parts that are easy to see — tags, keyword density, description length — and ignores the retention curve that actually decides distribution. This suite inverts that: it puts watch time first, refuses to guess when the data is missing, and hands back changes you can paste into Studio yourself.

1. Install the skills

The repository ships an installer that fetches the release tarball and copies the eight skill folders into your Claude Code skills directory.

Global install, available in every session:

curl -fsSL https://raw.githubusercontent.com/deeployCO/youtube-seo-skills/main/install.sh | bash

Project-local install:

curl -fsSL https://raw.githubusercontent.com/deeployCO/youtube-seo-skills/main/install.sh | bash -s -- --local

Pin to a tagged release:

curl -fsSL https://raw.githubusercontent.com/deeployCO/youtube-seo-skills/main/install.sh | bash -s -- --ref v0.1.0

Prefer not to pipe a script into your shell? Download it from this page and copy the folders by hand:

git clone https://github.com/deeployCO/youtube-seo-skills.git
cd youtube-seo-skills
cp -r youtube-seo* ~/.claude/skills/
cp -r scripts ~/.claude/skills/youtube-seo/

The helper scripts must live inside the orchestrator skill at youtube-seo/scripts/ — the installer does this for you.

2. Install the helper dependencies (optional but recommended)

pip install -r ~/.claude/skills/youtube-seo/scripts/requirements.txt
brew install ffmpeg yt-dlp     # macOS

Core analysis needs yt-dlp. Thumbnail computer vision adds opencv-python, numpy, and pillow. CLIP SERP-similarity scoring additionally pulls open_clip_torch and torch, which is a large download — skip it if you do not need that one feature.

3. Add a YouTube Data API key (optional)

export YOUTUBE_API_KEY="your-key-here"

Without a key the skills fall back to yt-dlp. With one you also get tags, statistics, topicDetails, and captions metadata.

4. Export your Studio CSVs for real retention data

Tier 1 signals — true CTR by surface, APV, and retention curves — exist only in YouTube Studio and only for channels you own. Export them from Studio → Analytics → Advanced mode and hand the CSV to the skill when asked.

5. Run a skill

Each skill is user-invokable as a slash command:

/youtube-seo-audit       https://youtube.com/@yourhandle
/youtube-seo-video       https://youtube.com/watch?v=VIDEO_ID
/youtube-seo-optimize    https://youtube.com/watch?v=VIDEO_ID
/youtube-seo-keywords    "your niche"
/youtube-seo-thumbnail   thumbnail.jpg
/youtube-seo-competitor  https://youtube.com/@competitor

Suggested routes: new video launch is keywords → optimize → thumbnail; a stalled video is video (with the Studio CSV) → optimize; a channel overhaul is audit, which orchestrates the rest.