razbakov/summarize-youtube

Summarize YouTube

Turn YouTube links into faithful, skimmable summaries.

Why a skill (and not just "fetch the page")

A plain web fetch of a YouTube watch page returns nav chrome and, at best, the title — never the transcript. The reliable source of truth is the video's captions, pulled with yt-dlp. Auto-generated captions arrive as rolling lines (each cue repeats the tail of the previous one), so raw SRT is ~3x duplicated noise. The bundled script handles both problems, so you summarize from clean prose instead of fighting caption formatting every time.

Workflow

1. Collect the links and the intent

Pull every YouTube URL from the user's message. Note any delivery instruction ("to telegram", "in chat", "email me") and any focus ("just the action items", "compare them"). No instruction → default to a Telegram summary (see Delivery).

2. Fetch metadata + transcripts

Run the bundled script with all the links at once (it dedupes captions for you):

python3 <skill>/scripts/fetch_youtube.py <url1> <url2> ... --outdir /tmp

It prints one tab-separated line per video — id TITLE CHANNEL DURATION UPLOAD_DATE /tmp/clean_<id>.txt — and writes the cleaned transcript to each clean_<id>.txt. A video with no captions gets an empty path; for those, fall back to transcribing the audio (see /transcribe-via-faster-whisper) or tell the user it has no captions.

Read each clean_<id>.txt. These are full transcripts — read the whole thing, don't skim, so the summary is grounded in what was actually said.

3. Summarize each video

Write the summary from the transcript, not from the title or your priors. Lead with the concrete claims, structure, and any conclusions the video reaches.

Per video, use this shape (scale length to the video — a 5-min clip gets a few lines, a 40-min talk gets the bullets):

<Title> — <Channel> (<duration> · <date>)
https://youtu.be/<id>
<1-2 sentence thesis: what the video is actually about / its main claim>
• <key point / argument / step>
• <key point>
• <takeaway or conclusion>
Relevance: <why this matters to *this* user — tie to their projects, role, or
known interests; say "no direct tie — <reason to watch anyway>" when there isn't one>

Two non-negotiable fields, because users ask for them every time if omitted:

• The link. Always include a clickable URL per video (https://youtu.be/<id>). Paste it as a raw URL on its own line, not a markdown [title](url) — chat surfaces like Telegram render raw URLs but not markdown link syntax.
• Relevance to the user. One line per video connecting it to what you know about them (their projects, work, stated interests). This is what turns a generic summary into a personal feed digest. If you don't know the user's context, infer from the conversation/repo or ask once — don't skip it.

Grounding rules that keep summaries trustworthy:

• Don't characterize what you haven't verified. If a video's framing makes you want to call it "satire", "fake", "outdated", or "obviously true", resist unless the transcript itself establishes it. Report what it says; flag genre ("this is presented as comedy/commentary") only when the transcript supports it. Tone is not proof.
• Attribute claims to the video ("the host argues…", "she claims…"), don't launder them into your own assertions of fact.
• Keep names, numbers, and quotes accurate to the transcript.

4. Deliver

Default — Telegram. Do not ask which surface. When the user gives no delivery instruction, send the summary to Telegram and then recap in chat — asking "should I send this to Telegram?" defeats the point, because Telegram is already the answer. Only deliver elsewhere if the user explicitly named a different surface in their request.

Send via the instance's proactive Telegram sender (in this setup: .bin/telegram-send.py; the agent/bot mapping lives in your private CLAUDE.md — pick the agent whose domain fits the video topic). Telegram renders HTML only, not Markdown — use <b>, <i>, <code>, line breaks; no # headers, no * bullets (use •), no code fences. Always write the message body to a temp file and pass it with --file; inline multi-byte text on the command line corrupts. Example:

python3 .bin/telegram-send.py --agent <agent> --file /tmp/yt_summary.html

Other surfaces. If the user named one, deliver there instead (a saved file, a chat reply, an email draft, a note). Match that surface's formatting.

After delivering, give the user a short in-chat recap too, so they see the result without leaving the conversation.

Notes

• Pass all links to the script in one call — it's faster and keeps the videos grouped.
• Non-English captions: pass --lang <code> (e.g. --lang de). Summarize in the user's language unless told otherwise.
• This skill is for consuming a video (understanding/summarizing). For editing a video's own title/description/chapters, use /youtube-metadata-updater; for deep multi-source research off a video, use /youtube-notebooklm-research.
• Requires yt-dlp on PATH.