What this agent does
One good long-form video usually has five short clips hiding inside it, but cutting them by hand means re-watching the whole thing with a stopwatch. This agent finds the strongest moments in a long-form video and turns each one into a short, captioned clip ready to post.
stage-1-highlight-detect— analyzes the video's transcript and pacing to identify 5 standalone moments that work without the surrounding context (a strong hook, a complete thought, a punchline).stage-2-clip-caption— cuts each identified segment, generates burned-in captions synced to the audio, and writes a short platform-native caption + hashtag set for each clip.
Every run saves the 5 clips plus their captions into a dated output folder, ready to review and post.
Requirements
- A video file (local, or a URL Claude AI can fetch) with clear spoken audio — this agent relies on the transcript for highlight detection, so heavily music-only content won't work well.
- Works on Claude AI (Claude Code, with a local video-processing tool like
ffmpegavailable on the machine) for the cutting/captioning step. ffmpeginstalled locally for video cutting and caption burn-in — this is the only external dependency.
Setup
- Unzip the package into your workspace, keeping the folder structure intact.
- Confirm
ffmpegis installed and on yourPATH(ffmpeg -versionto check). - Fill in your preferred clip length range and caption style.
- Run the starter prompt against a sample video to see the full pipeline end to end.
- For a recurring workflow (e.g. every new upload in a watched folder), adapt the automation pattern to your folder-watching setup.
What you get
- The 5 strongest standalone moments automatically identified in any long-form video, no manual scrubbing
- Cut clips with burned-in, audio-synced captions ready to post
- A platform-native caption and hashtag set written for each clip
- A dated output folder with all 5 clips and captions ready to review and post
