Give Claude Eyes — watch any video frame-by-frame, 100% local (buildwith.conrad)
🛠️ Want the rest of my tools? Grab them at get.caffier-agency.com/start
Freebie for the EYES keyword (reel-47, "Give Claude eyes"). Deliver as a public GitHub Gist — numbered steps the user can run today, not a raw link. Value-first: by the end you have a Claude Code skill that sees a video (every cut, every on-screen detail), not just reads its transcript.
Claude has no native video model. So every "analyze this video" tool just pulls the transcript — and misses everything that's only on screen: the cuts, the text overlays, the visual hook, the b-roll. Half the meaning of a good video lives in the pixels, not the words.
This skill fixes that. It splits any video into the two things Claude does understand — images + text — and hands both over with the timestamps lined up. Claude flips through the frames like a flip-book while it reads the script, so it knows exactly what's on screen the moment something is said. All local. Zero API cost.
Three free command-line tools:
- yt-dlp — pulls the video from almost anywhere (YouTube, Instagram, Loom, TikTok, a direct URL).
- FFmpeg — rips frames + a clean audio track out of the video.
- whisper.cpp — transcribes the audio locally (free), if the source has no captions.
# macOS (Homebrew)
brew install yt-dlp ffmpeg whisper-cpp
# grab a small, fast local transcription model (~150 MB, one time)
mkdir -p ~/.claude/models
curl -L -o ~/.claude/models/ggml-base.en.bin \
https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin(Windows/Linux: install yt-dlp, ffmpeg, and whisper.cpp from their repos — same three tools.)
Save the file below as ~/.claude/skills/watch-video/SKILL.md, then restart Claude Code. That's it —
Claude now has a watch-video skill it will use whenever you hand it a video URL.
In Claude Code, just say:
/watch-video https://www.instagram.com/reel/XXXXXXXXX/
or "watch this video and break down why it works: ". Claude downloads it, samples the frames, transcribes the audio, and gives you a scene-by-scene breakdown — hook, cuts, on-screen text, the payoff.
The trick that makes it accurate: sample the first ~15 seconds at 15 fps (the hook carries the most motion and decides retention), and the rest at 1 frame / 3–4s. One frame every few seconds only catches end-states and misses the word-by-word reveals, the fast cuts, and the count-up animations. Dense on the hook, sparse on the body.
---
name: watch-video
description: Use when the user gives a video URL (YouTube, Instagram, Loom, TikTok, direct link) and wants Claude to actually SEE the video — a scene-by-scene / frame-by-frame breakdown, not just the transcript. Triggers on "/watch-video URL", "watch this video", "break this reel down".
---
# watch-video — see the video, not just the transcript
Claude has no native video model, so this skill splits a video into images + text and reads both with
timestamps aligned. Run every step; never skip the dense-hook sampling.
## Step 1 — Download the video
```bash
mkdir -p /tmp/watch-video && cd /tmp/watch-video
yt-dlp -o video.mp4 -f "mp4" "<URL>"
ffprobe -v error -show_entries format=duration -of csv=p=0 video.mp4 # duration (seconds)
```
## Step 2 — Two-stage frame sampling (the important part)
The hook holds the most motion and decides retention — sample it densely; sample the body sparsely.
```bash
# HOOK: first 15s at 15 fps → tiled contact sheets (5 cols x 9 rows = 45 frames = 3s per sheet)
mkdir -p hook && ffmpeg -y -t 15 -i video.mp4 \
-vf "fps=15,scale=200:-1,tile=5x9:margin=6:padding=4:color=white" -an hook/sheet_%02d.jpg
# BODY: 1 frame / 3.5s across the whole video
mkdir -p body && ffmpeg -y -i video.mp4 -vf "fps=1/3.5,scale=360:-1" body/f_%02d.jpg
```
Read the sheets in order; frame N in the hook is at t = N / 15 seconds (sheet k, cell i → frame (k-1)*45 + i).
## Step 3 — Transcribe the audio (free, local)
```bash
ffmpeg -y -i video.mp4 -ar 16000 -ac 1 audio.wav
whisper-cli -m ~/.claude/models/ggml-base.en.bin -f audio.wav -otxt -of transcript
```
If YouTube has captions, use those (free, exact). Always sanity-check the transcript against the frames —
local models mishear names (they'll hear "Claude" as "plot").
## Step 4 — Read frames + transcript together, then output
Look at the hook sheets and body frames, line them up with the transcript timestamps, and produce:
1. **Transcript** (corrected against what's on screen).
2. **Scene-by-scene breakdown** — for each chapter: time, what's on screen, the on-screen text, the cut.
3. **Why it works** — 2–3 concrete mechanisms (hook structure, cut cadence, the one visual payoff, the CTA).
4. **Steal-the-structure notes** — what to lift into your own video (change the content, keep the structure).
Clean up `/tmp/watch-video` when done.Want Claude to run workflows like this across your whole business — remembering your projects, your brand, and your taste every session instead of starting cold? That's what my get. kits set up (the CLAUDE.md + skills + memory system). But this skill alone will already change how you study every video you save. 🚀
— Conrad · @buildwith.conrad