The AI Eyes Kit

# The AI Eyes Kit: let your AI watch the videos you never will — for about a cent each

Your "watch later" list is a graveyard. You save the tutorial, the two-hour lecture, the
conference talk you swear you'll get to — and you never watch a single one. Deep down you already
know it. So I stopped watching, and I let my AI watch them for me instead. It hands back the
notes; I do something else. This is the exact setup I use, packaged so you can run it in about
ten minutes on any machine with Node and ffmpeg.

The kit is free. On Instagram you can also comment EYES and the bot will DM it to you;
everywhere else the download button above is it.

You don't need a fancy "video understanding" API that charges by the minute. You need four steps:

1. Pull a handful of frames out of the video with ffmpeg — a dozen evenly-spaced stills catch
the slides, the code, the diagrams, the on-screen text.
2. Pull the transcript — the spoken words (a .vtt/.srt sidecar; for YouTube, yt-dlp
grabs the auto-caption without downloading the video).
3. Put the frames and the transcript into one multimodal prompt to a cheap vision model.
4. Ask for structured notes — a summary, an action list, or the claims worth verifying.

A dozen frames plus a transcript is a few thousand tokens. At the price of a model like
gemini-2.5-flash-lite, that lands around one cent a video — versus the twenty minutes it
would cost *you* to watch it. That arbitrage is the whole idea.

It's three small pieces — no account beyond a free OpenRouter key, no subscription:

- `watch.js` — the watcher, about 80 lines. Point it at a video, get notes:
node watch.js my-tutorial.mp4 --mode summary > notes.md.
- `ai-eyes-setup.md` — the setup: create a free OpenRouter account, make a key, set it as an
environment variable (never hard-coded), and watch your first video. About ten minutes.
- `note-formats.md` — the three modes, and how to add your own.

Same frames, same transcript, three different questions — so you get the shape of notes you
actually need instead of a wall of text:

- `summary` — one line on what it is, then 5-8 bullets of the real substance. The "I never
have to watch it now" mode. Best for talks, tutorials, and lectures.
- `actions` — only the concrete things to *do*, as a numbered checklist. Turns a how-to video
straight into a to-do list.
- `claims` — every factual claim worth verifying, one per line, with a note on why. A
skeptic's reading list for hot-takes, pitches, and health/finance content.

Each mode is one line in the script's PROMPTS object. Add a key, write the instruction, call it
with --mode yourname. The pattern never changes: **frames + transcript + one clear instruction →
structured notes.**

1. Get the tools. Node 18+ and ffmpeg (node --version, ffmpeg -version).
2. Get a free OpenRouter key. Sign up at openrouter.ai, create a key (starts with sk-or-...).
At a cent a video, a couple of dollars of credit is hundreds of videos.
3. Set the key as an environment variable — export OPENROUTER_API_KEY="sk-or-...". Never
paste a key into the script.
4. Watch a video — node watch.js my-tutorial.mp4 --mode summary > notes.md.
5. Add a transcript for sharper notes — `yt-dlp --write-auto-sub --sub-format vtt
--skip-download "<url>", then --transcript the-file.vtt`.

The setup is a one-time, ten-minute cost. After that, the watcher is a standing machine: point it
at any video — a new tutorial, a two-hour lecture, your entire watch-later list — and each one
becomes notes for a cent, with none of your attention. The pile that used to grow forever now
clears itself. Set it up once, and every video from here on is one command and a cent.

- [OpenRouter](https://openrouter.ai) — the single API that routes to the cheap vision model.
- ffmpeg — pulls the frames (and yt-dlp grabs the transcript). Both free, both standard.
- The kit itself — watch.js, the setup guide, and the note formats, in one zip:
download it here. On Instagram, comment
EYES and the bot DMs it to you.

The one-cent cost is real — it's the model's own per-call price, which OpenRouter shows you.
The "nine videos overnight" in the video is a representative good night, not an audited stat; how
many you clear depends on how big your backlog is and how long the videos are. The mechanism is
the same either way: your watch-later list stops being a graveyard and starts being a folder of
notes.

Install Node and ffmpeg, grab a free OpenRouter key, and run watch.js on one video sitting in
your "watch later" list right now. The moment a video you were never going to watch turns into a
clean set of notes for a cent, that whole guilt-pile stops being scary — and it starts clearing
itself, one command at a time.