Watching Isn't Owning: Turn YouTube Gold into Project Knowledge

Here's an uncomfortable inventory: how many genuinely great YouTube videos did you watch last month — lectures, breakdowns, masterclasses? And how many changed anything about how you work?

That gap has a name: watching isn't owning. The video was gold; your memory is a sieve. Notes would help, but pausing every forty seconds to type is miserable, so nobody does it twice.

This week we closed that gap for real, with a plugin — and the last step is the part nobody talks about: the knowledge didn't stop at "notes." It got wired into the project so our agent applies it automatically.

The tool: claude-video-vision

claude-video-vision (MIT, by Jordan Vasconcelos) is a Claude Code plugin that gives your agent eyes and ears for video. Point it at a local file or a YouTube URL and it can:

• pull metadata (title, duration, whether captions exist) before spending any effort,
• extract frames when the visuals matter — slides, code on screen, demos,
• get the spoken content — from YouTube's captions when they exist, or via Whisper transcription when they don't,
• and run a structural analysis (scene changes, silences) so the agent knows where to look instead of chewing every frame.

Install it from the plugin marketplace inside Claude Code (/plugin → search "video vision"). It needs ffmpeg, plus yt-dlp for YouTube URLs. Whisper is optional — and here's the honest tip that saves you an hour: most YouTube videos already have auto-captions, so you often don't need Whisper at all. Our machine doesn't even have it installed; everything below ran on captions.

The real run: a marketing lecture series → our project's doctrine

The actual case from our kitchen this week. We're studying Seth Godin's This Is Marketing alongside a Korean lecture series on YouTube — podcast-style episodes, 15–17 minutes each, one static image and one voice. Watching them is pleasant. Retaining them is the problem.

The pipeline the agent ran, per episode:

Step 1 — Check before you cook. video_info on the URL: title, 17:28, captions available. One structural scan showed a single scene change in the whole video — meaning frames carry nothing, skip them entirely. That decision alone saves most of the tokens.

Step 2 — Get the words. Captions come down in seconds, deduplicated and timestamped. A 17-minute lecture becomes ~400 clean lines of text.

Step 3 — Organize like a student, not a stenographer. The agent turns the transcript into a study note: key claims with timestamps, the memorable quotes verbatim, and — because the note lives in the repo — each episode appended to the same growing file:

docs/MARKETING_STUDY.md
├─ #0 Author's note — marketing = change; "who's it for?"
├─ #1 Chapter 1 — culture beats ads; find your people first
└─ (next episode appends here)

Step 4 — The step that makes it owning: apply it. Every episode's note ends with a table: this concept → what it means for this project → adopted or declined. The adopted ones don't stay in the note — they get written into CLAUDE.md, the file our agent reads at the start of every session. Godin says find the smallest group you can serve? Our CLAUDE.md now defines exactly one reader. Godin says change happens through tension? Our CLAUDE.md now requires every post to open with tension and resolve it with a runnable prompt.

Read that last part again, because it's the whole trick: a YouTube video changed the standing instructions of a working project. The agent that wrote this very post follows rules distilled from those videos. That's the difference between a summary and a system.

Two honest caveats

1. Auto-captions garble domain terms. In our run, the captions consistently misheard a key term from the book's official translation — close enough in sound, wrong on paper. A human who knows the material caught it in one glance, and we added the correction to the note's terminology rules. Keep yourself in the loop for vocabulary; delegate everything else.
2. Respect the creators. This pipeline is for making your own study notes from content you watch — the notes live in your repo, not on your blog. Link the original, credit the channel, and if a series is good (ours is), say so where people can see it.

Order up

Next time you finish a video worth keeping, don't bookmark it — paste this into Claude Code (with the plugin installed):

Watch this YouTube video: [URL]

1. Check the metadata first. If it's talk-focused (few scene changes), work from
   captions only — don't extract frames unless the visuals carry information.
2. Write a study note to docs/STUDY_NOTES.md in this project: key claims with
   timestamps, memorable quotes verbatim, and a one-line "so what" per section.
3. End with the important part: propose 1-3 concrete changes to this project's
   CLAUDE.md based on what this video teaches. Show me the diff and apply only
   what I approve.

Step 3 is where watching becomes owning. A video that doesn't change anything in your project was entertainment — which is fine, but now it's a choice.

So — what's the one video you keep re-watching because you never wrote it down? Feed it to the kitchen tonight.