Youtube Transcript

YouTube transcript & search

Two capabilities, both backed by yt-dlp, no API key and no quota:

1. Search YouTube by topic → ranked candidate list (find videos yourself).
2. Transcribe a known URL → clean plain text (read a video).

The natural pipeline: search.sh "topic" → pick the relevant hits → get_transcript.sh on each → summarize / give verdicts.

The point: an agent that can only see a video's title is guessing. With this it reads what was actually said, so "watch these 20 videos and tell me which are worth my time" becomes a real instruction.

Requirements

yt-dlp must be on PATH:

brew install yt-dlp        # macOS
pipx install yt-dlp        # anything with Python

Keep it fresh — YouTube changes things and a stale yt-dlp starts failing (brew upgrade yt-dlp / pipx upgrade yt-dlp).

Search

/search.sh "" [count] [today|week|month|year]

• query – free-text topic, quote it. Works in any language (Cyrillic fine).
• count – how many results (default 10).
• when – optional freshness window. Omit for relevance, all-time.

Output: a numbered list, each entry = title, then channel · duration · views (and the upload date when a window is set), then the full youtube.com/watch?v=… URL on its own line (ready to pipe into get_transcript.sh).

Two modes:

• No window → relevance order, flat & fast, no dates shown.
• With window (today/week/month/year) → uses YouTube's own sp= upload-date filter and full extraction, so real upload dates appear. Slightly slower (worth it — that's the point of asking for fresh). Use this for "what's new in this niche" sweeps.

Note: ytsearchdate is unsupported in many packaged yt-dlp builds, so freshness goes through the sp= filter instead — which is richer anyway (time windows, not just sort).

Transcribe

/get_transcript.sh [--save] "" [sub-langs]

Prefer --save when you (the agent) will read the transcript. Without it the full transcript streams to stdout, and a long one (tens of KB) gets spilled by the harness into a random-hash file that's awkward to track. With --save the script writes to /.cache/ [].txt and prints only a short summary plus that path, so you then Read a file named after the actual video. It doubles as a cache: re-running the same URL rewrites the same well-named file.

• url – full video or Short URL (youtu.be/…, youtube.com/watch?v=…, youtube.com/shorts/…).
• sub-langs – optional yt-dlp --sub-langs override. By default no language is needed: the script reads the video's original audio language from YouTube metadata (%(language)s) and fetches that track, preferring the original -orig captions over auto-translations. If metadata has no language, it falls back to a Cyrillic-in-title heuristic (→ ru), then English. Pass a spec (e.g. ru.*,ru or .*) only to force something.

Output: a header line (title · channel), the URL, a blank line, then the transcript as one flowing text block. Exit code 2 + NO_SUBTITLES on stderr means the video has no captions at all — retry with an explicit --sub-langs ".*".

Notes

• Auto-generated captions have no punctuation-perfect formatting; the cleaner strips timestamps, inline tags, and rolling duplicate lines, but the text stays spoken-language rough. Good enough to read and summarize, not a polished quote source.
• Works on any video that has captions — including auto-generated ones, which is most of YouTube. Videos with captions disabled entirely are the exception, and the script tells you so instead of failing silently.
• This is for reading YouTube content. Transcribing arbitrary audio files is a different job — use a speech-to-text tool for that.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: grigorytal
• Source: grigorytal/youtube-transcript-skill
• License: MIT

Install and usage instructions live in the source repository linked above.