Analyze Videos with Claude: New Open-Source Skill for AI Agents
I'm not going to watch all of this.
Now you don't even need to — an skill has appeared that allows Claude to analyze video via a link almost like a human.
It works with YouTube, Loom, TikTok, local files, and anything supported by yt-dlp.
With its help, an agent can deconstruct a competitor's launch by hooks, visuals, and structure, find a bug from a screen recording, create a short summary of a long presentation with precise timestamps, or extract only the actual changes from a product video without the marketing fluff.
Under the hood, yt-dlp first fetches available subtitles, and ffmpeg extracts frames while accounting for scene changes so as not to waste tokens on dozens of identical images. Claude receives frames with timestamps, and if there are no subtitles, Groq Whisper kicks in.
For Claude Code:
/plugin marketplace add bradautomates/claude-video
/plugin install watch@claude-video
For Codex, Cursor, and Gemini CLI:
npx skills add bradautomates/claude-video -g
p.s. For videos longer than 10 minutes, it is better to specify the required fragment using --start and --end; otherwise, a full analysis can consume a significant number of tokens.