YouTubeSearch

Search YouTube by query and return structured, human-readable results with metadata and engagement metrics.

What It Does

Runs a yt-dlp search against YouTube, returning the top N results (default 20) filtered to a recent time window (default 6 months). Each result includes:

• Title and URL
• Channel name and subscriber count
• View count and duration
• Upload date
• Engagement ratio (views / subscribers) — a quick signal for whether a video over- or under-performed relative to the channel's audience

Numbers are human-readable (e.g., 1.2M, 45.3K). Results are separated by dividers for easy scanning.

Requirements

• yt-dlp installed and in PATH
• jq installed and in PATH
• bc installed (standard on macOS/Linux)

Usage

Run the bundled script:

bash ~/.claude/skills/YouTubeSearch/scripts/yt-search.sh "<search query>" [--count N] [--months N]

Parameters

Examples

# Basic search — top 20 results from last 6 months
bash ~/.claude/skills/YouTubeSearch/scripts/yt-search.sh "kubernetes security best practices"

# Narrow to 5 results from the last month
bash ~/.claude/skills/YouTubeSearch/scripts/yt-search.sh "rust async tutorial" --count 5 --months 1

# Broader window — last 2 years
bash ~/.claude/skills/YouTubeSearch/scripts/yt-search.sh "home lab setup" --months 24

Interpreting the Engagement Ratio

The views-to-subscribers ratio helps identify standout content:

• > 1.0x — The video got more views than the channel has subscribers. Strong signal that the topic resonated or the algorithm boosted it.
• 0.3x - 1.0x — Typical range for established channels.
• < 0.3x — Below average reach. Could mean the topic is niche, the thumbnail/title underperformed, or the channel's audience has moved on.

This metric is most useful for comparing videos on the same topic — a 5.0x ratio on a small channel often means the content hit a nerve.

How Claude Should Use This

When the user asks to search YouTube or find videos:

1. Run the script with the user's query and any specified flags
2. Present the results — the script output is already formatted for terminal reading
3. If the user wants analysis (e.g., "which of these are worth watching?"), use the engagement ratio and view counts to highlight standouts
4. If subscriber count shows "N/A" for many results, that's normal — yt-dlp can't always fetch channel metadata from search results

Troubleshooting

• No results: Try broadening the query or increasing --months
• Slow: Each result requires a metadata fetch. Reduce --count for faster results.
• "N/A" for subscribers: yt-dlp sometimes can't resolve channel follower counts from search. The other fields will still populate.

Gotchas

• API order=date still applies relevance ranking — recent uploads can be deprioritized. Two-query merge (date + relevance) is the workaround.
• Quota is per-project per-day — search is expensive (100 units/query); bursting locks the project out for the day.
• regionCode parameter changes results — same query from different regions returns different videos; default is viewer's IP geo.
• Transcript API is separate from search API — transcript availability is per-video; some videos have none and the API returns no error.
• Channel ID vs Channel username: handles (@name) vs legacy usernames vs IDs are three different identifiers; resolution requires explicit channel-list call.
• Embedded videos blocked by uploader: search may return them, but a downstream tool that embeds will fail without warning — check status.embeddable.