Telling Technology · Learning in public
All episodes
EP.08 — NOT WATCHING
A Claude Code workflow, on screen

Learn things by
not watching

A 20-minute YouTube tutorial was on my list. I never pressed play. One slash command downloaded it, cut it into 80 frames, read all 565 transcript segments — and then three research agents fact-checked the video's money claims. Spoiler: the numbers were ~4× inflated.

A paused YouTube player dissolving into a grid of extracted film frames feeding a terminal window
⌁ /watch — 20:27 video · 80 frames · 565 segments · never pressed play
it started with a question about a tool I didn't have
01 The question

"Can you watch YouTube?" — "No. But…"

I'd seen a video claiming Claude Code could watch videos, so I asked mine if it had a fetch tool or a YouTube MCP. It didn't — and said so. Then it did something better: it pulled the video's auto-captions in one Python call and summarized the whole thing in seconds. Text only. No frames. And honest about exactly that limit.

claude code — desktop
> Do you have the tool called fetch? Can you read YouTube content?
Short answer: no — no fetch MCP and no YouTube MCP in this session. WebFetch gets the watch-page HTML, not the transcript. But I can pull the caption track directly… ✓ captions fetched (~2s) — here's the summary.
Terminal showing a caption-only YouTube summary, generated in seconds
captions only · ~2 s
round 1: transcript-only — fast, but blind
02 The difference

So what does actually watching add? We tested it.

Before installing anything, I made Claude prove the upgrade was worth it. It rebuilt the pipeline by hand — yt-dlp download, ffmpeg cut into 80 frames, frames + transcript fed back to itself as one multimodal prompt. The frames caught what the transcript never said: on-screen comparison tables, UI details, even that the video's intro was itself made with the same /watch skill, run on a different video.

claude code — desktop
> What's the difference between that skill and what you just did? Do I gain anything?
Honest comparison: I read text only. The skill adds yt-dlp download, ~80 ffmpeg frames, Whisper fallback for caption-less videos, and works on TikTok / Loom / local MP4s. For talking heads: no difference. For anything shown on screen: big difference. Let me prove it…
A contact sheet of 80 frames extracted from a YouTube video by ffmpeg
80 frames · 0.137 fps
what the transcript can't see: everything on screen
03 The glitch

Installed it — and /watch didn't exist

Two slash commands installed the plugin: marketplace add, then install. Then I pasted my first /watch URL and got… "Unknown command: /watch. Did you mean /batch?" The plugin was fully installed — config confirmed it. The catch: plugin slash commands load once, at session start. Installing mid-session means the running session never hears about it. Restart, and it just works.

claude code — desktop
> /watch https://www.youtube.com/watch?v=nQOGK72IHx8 summarize this for me
● Unknown command: /watch. Did you mean /batch?
Claude Code replying Unknown command: /watch. Did you mean /batch?
the real error
installed ✓ enabled ✓ loaded ✗ — restart required
⚠️

Transcripts lie a little

Bonus glitch from the same session: the video's auto-captions mangled the name of a keyword tool into clearer. The frames + a research agent revealed it was actually LowFruits.io. If we'd trusted the transcript alone, the episode's best takeaway would have pointed at a tool that doesn't exist.

04 The receipts

One command, four artifacts

Everything below came out of a single /watch run on a 20-minute video I never played. Tap any image to enlarge it and read exactly how it was produced.

FACT-CHECKED BY THREE RESEARCH AGENTS

The video said $500K/month. The agents said ~4× inflated.

This is the part watching can't give you. After the summary, I asked Claude to spin up research agents on the video's claims. Verdict: real traffic ~2.6M visits/mo (not 11M), realistic revenue $15–50K/mo (not $500K), and the video was one of 8–10 near-identical "boring website" videos — a coordinated affiliate push paying up to 60% commission. The business model is real. The headline numbers are hype. I learned more than the video knew about itself — by not watching it.

A verdict board: claimed 11M visits and $500K/mo versus verified 2.6M visits and $15–50K/mo
session 2026-05-21 · 1 × /watch · 3 research agents · 0 minutes of video watched
06 Steal this

The whole setup is two commands

The /watch skill is free and open-source (by Brad — credit where due). It needs yt-dlp + ffmpeg on your machine, and optionally a Groq or OpenAI key for the Whisper fallback when a video has no captions. Everything in this episode is free and open — install it, point it at any video, make it yours.

skill watch@claude-video cmd /watch <url or file> deps yt-dlp + ffmpeg optional GROQ_API_KEY → Whisper fallback
run it /plugin marketplace add bradautomates/claude-video && /plugin install watch@claude-video

No GitHub? Comment WATCH on the post and the bot DMs you the link.

Next episode

Then I turned the whole system into a game.

A command that watches videos is one card in a much bigger deck. Next: I batch-generated 18 gamified "unlock cards" for my whole operator setup — from my phone, on the couch, in one Telegram session.

Teaser: a deck of 18 gamified unlock cards for the whole AI-operator setup, dealt out across a dark table
drops next · follow @bitesizeai so you don't miss it