Free Claude Code skill
Claude Code can understand your video
Claude Code has no native video perception. Point its Read tool at an .mp4 and you get back "This tool cannot read binary files." Ask about the video anyway and you get a confident description of something nobody looked at. This skill runs perception passes over the file on your own machine and leaves Claude a report, a transcript and stills it can already read, so the answer is grounded in the file.
- Setup
- 2 min
- Download
- 13.9 KB
- Cost
- $0
- Required tools
- 2
What you will have when you finish
No multimodal model captions your frames. The passes turn a video into text, numbers and still images, then you ask Claude against them: where the cuts fall, what was said at 0:12, whether the on-screen text matches the audio.
- A transcript primed with your domain terms. Unprimed, Whisper turned "Sonnet" into "Sonified" and "stale processes" into "for sale processes".
- A shot list and pacing from PySceneDetect's adaptive detector, so the cut rhythm is visible as numbers.
- Burned-in on-screen text read by RapidOCR, which doubles as an independent cross-check on the transcript.
- A motion heatmap built from per-pixel standard deviation in NumPy, showing where the movement actually is.
- Audio levels in LUFS, silence gaps, freeze and black frame detection and content bounds from FFmpeg.
- A music-bed reading from Demucs, telling you whether there is music under the voice or voice alone.
- Contact sheets, so you and Claude can look at the frames.
Do this once
Four steps, about two minutes if FFmpeg is already on your machine.
Python 3 and FFmpeg on your PATH. Every other pass is optional.
- 01
Download the zip from this page. Four files, 13.9 KB.
Download the skill - 02
Unzip it into your skills folder. It expands to a video-recon/ folder that is already structured the way Claude Code expects.
~/.claude/skills/On Windows that folder is C:\Users\<you>\.claude\skills\. The skill ends up at ~/.claude/skills/video-recon/SKILL.md.
- 03
Ask Claude to analyse a video, or run the script yourself. No restart. Claude Code picks up a new skill inside the session you already have open.
- 04
Read the preflight. Every run marks each pass ok or MISSING and prints the exact pip install line for anything absent.
A missing optional pass is skipped, never fatal. The report names what was skipped and the command that enables it.
The skill diagnoses itself. Run it and it tells you what you are missing.
The specifics
- What it costs
- Free. Download the zip, keep it, change it. There is no account and no sign-up.
- What it needs
- Python 3 and FFmpeg. Whisper, PySceneDetect, RapidOCR, Demucs, NumPy and OpenCV are optional and skipped when absent.
- Where your data goes
- Every perception pass runs on your machine. The video file is never uploaded. Claude reads the report the passes leave on disk.
- What we tested it on
- The 29-second reel that announced it, captions burned in, with the transcript verified line by line against those captions.
- What is out of scope
- No multimodal model captions your frames. It will not name a person on screen or tell you whether a shot looks good.
- Checked on
- 5 September 2026, against Claude Code's own error output and Anthropic's published vision documentation.
Questions people ask
What stops Claude making things up about the video?
Every claim traces to a pass you can check. RapidOCR reads the burned-in text independently of what Whisper heard, so the two cross-check each other, and the contact sheets let you look at the frames yourself.
What if I do not have a GPU?
It still runs. Whisper and Demucs work on CPU and take longer there. The preflight prints what is installed before any of it starts, so you know the shape of the run up front.
What if the video has no speech in it?
The other passes still run. Shot detection, on-screen text, motion mapping and the audio readings do not depend on anyone talking, and the report says which passes produced nothing.
Get video recon
Four files, 13.9 KB, free. Unzip it into your skills folder and ask Claude about a video in the same session. The first run tells you if anything is missing and how to install it.