Free Resources@dylan_j_watkins
← All resources
WATCHUnlocked

Give Claude Eyes: The 10-Minute Install + Wire-In Playbook

The outcome: in about 10 minutes, your Claude can watch any video on the internet (YouTube, TikTok, IG reels, Looms, local screen recordings) and answer questions about what's actually on screen. Free, open source, runs on your machine.

This is the exact setup I run. I audited every line of this skill's code before installing it, and the same afternoon it was breaking down competitor reels for me frame by frame.

1 — The skill (free, open source)

Repo: https://github.com/bradautomates/claude-video

Built by Brad Bonanno (@bradbonanno on YouTube). MIT licensed. It works by splitting any video into the two things Claude already understands: frames (images) and a timestamped transcript. No expensive video-model API in the middle.

2 — Install in Claude Code (2 commands)

Open Claude Code and run: /plugin marketplace add bradautomates/claude-video
Then run: /plugin install watch@claude-video
First use: type /watch <any video URL> <your question> — the skill auto-installs its two dependencies (ffmpeg + yt-dlp) on first run and walks you through anything missing.

Other surfaces:

3 — The 2-minute upgrade: Groq key (for videos without captions)

YouTube videos usually have free captions, so they cost nothing. TikToks, reels, and screen recordings usually don't — the skill falls back to Whisper transcription for those.

Grab a free key at https://console.groq.com/keys (their free tier covers ~2 hours of transcription per hour — I use this daily and have never paid)
Put it in ~/.config/watch/.env on the line GROQ_API_KEY=your_key_here

That's the whole setup. Paste a URL, ask a question, Claude watches it.

4 — My wire-in playbook (the part nobody else gives you)

Installing the skill is step one. The real unlock is wiring it into a system so it runs without you. Here's the pattern I used:

  1. Build one adapter script that calls the skill's engine as a library and always outputs the same files: source.mp4, source.srt, transcript.json, hero frames, and a media.json with cut timestamps. One media layer, every pipeline shares it.
  2. Keep your metrics scraper separate. The watch skill sees the video; it can't see views, likes, saves, or follower counts. I still pull those via Apify. Video understanding + metrics = a complete teardown.
  3. Point your existing workflows at the adapter. My competitor-teardown, swipe-file canvas, and video-deconstruct flows all call the same adapter now. One afternoon of rewiring replaced three separate scraping mechanisms.
  4. Let it compound. Every video your AI watches becomes searchable notes in your knowledge base. The system gets smarter every day without you watching anything manually.
💡 Pro tip: for a long video, run /watch <url> --start 2:15 --end 2:45 to zoom into one section instead of burning tokens on a sparse full scan.

Want this running in YOUR business?

If you run a clinic or practice and want a system like this doing your content research automatically — watching what works in your niche and turning it into content ideas while you see patients — that's literally what I build. DM me BUILD on Instagram (@dylan_j_watkins) and I'll show you what it looks like.

— Dylan