/watch (Claude Video)

Why it's ranked here

The pitch is unusually well engineered for a single-purpose skill. The frame budget scales with video length instead of blindly sampling at a fixed rate, a visual near-duplicate check keeps a ninety-second static slide from burning the same tokens as ninety seconds of dense motion, and the same package installs cleanly across Claude Code and other agent hosts, not just one. Licensed MIT, with a documented fix that closed a subprocess argument-injection path in how it shells out to the downloader. That is the kind of detail that separates a maintained utility from a one-off script.

What's good

The engineering shows in details that matter for a tool that shells out to external binaries. The near-duplicate frame filter compares each candidate to the last frame actually kept, not the previous one, so it still catches a slow fade even though a simple frame-to-frame threshold would miss it. Frame counts are capped by minute of runtime rather than one flat number, so a thirty-second clip gets dense coverage and a fifty-minute one does not silently blow past a reasonable budget. Users can also pin exact timestamps a transcript calls out, so a moment the speaker flags by name is never sampled away by chance.

Tradeoffs

It leans on two external binaries and, for videos without captions, a paid speech-to-text API key, so a fresh install is not truly zero-effort: the automatic installer only covers macOS, and Linux or Windows users have to run printed commands themselves. The most thorough detail mode is explicitly uncapped, so a long, high-motion video can spend far more in image tokens than expected unless the user knows to ask for a narrower window first. Coverage of a long video is inherently sparse in every mode except that priciest one, which trades the sparsity for cost.

How to use it well

This fits best as an on-demand tool for a specific clip: debugging a screen recording someone sent you, checking what a competitor's demo actually shows, or pulling the real structure out of a talk instead of trusting its title. Start cheap, with captions only, and step up to frame sampling only when the question needs to see the screen. For anything over ten minutes, naming the section you care about gets far better answers than asking for the whole video, since full coverage past that length gets sparse by design. It is not a fit for exhaustive frame-by-frame motion analysis unless you accept the uncapped mode's cost.

Observed

License

MIT licensed

Packaging

Packaged as a Claude Code plugin/skill, and separately installable into other Agent-Skills-compatible hosts such as Codex, Cursor, Copilot, and Gemini CLI via a dedicated installer CLI

Interfaces

Ships as a Python command-line script invoked from within the skill, not exposed as an MCP server or hosted API

Dependencies

Requires the external binaries yt-dlp and ffmpeg on the host machine; the bundled installer auto-installs both via Homebrew on macOS and prints manual install commands on Linux and Windows

Speech-to-text

Speech-to-text fallback supports two interchangeable backends, Groq and OpenAI, selected by which API key is configured

Frame processing

Frame near-duplicate detection uses only the Python standard library after an initial ffmpeg downscale step, with no image-processing library dependency

Lifecycle hook

Registers a SessionStart hook that runs a setup-check script on session launch