grw—lab Studio

General + Free · v0.1.0

Video Reader

Turns a video into a contact sheet, key frames, cut times and audio numbers so an agent can break it down without watching it.

  1. 01

    Add the catalog

    Open Claude Code and paste this line. Do this once. If you already added grw—lab for another skill, skip it.

    $ /plugin marketplace add grw-lab/skills

  2. 02

    Install Video Reader

    Paste this second line. It installs Video Reader and leaves the other skills out.

    $ /plugin install video-reader@grw-lab

Every skill instead: use /plugin install grw-skills@grw-lab as line 2. On claude.ai, zip the video-reader folder and upload it in Settings, then Capabilities, then Skills. In Cursor, point the agent at video-reader/SKILL.md.

View on GitHubRead as Markdown

What it does

Video Reader lets an agent study a video file. The agent cannot watch pictures move or hear the soundtrack, so the skill runs one script that makes things it can read:

  • A contact sheet of the whole video, with a timestamp under every frame.
  • Three large key frames, taken from the longest shots.
  • Cut times, and audio numbers: loudness per second, onset times, and a BPM only when those onsets are evenly spaced.

It then writes a structured breakdown: what happens, how it moves, the rhythm, the visual style, and what to take if you want to make something in the same spirit.

When to use it

  • You have an mp4, mov, webm or m4v and you want it broken down.
  • You want the cuts, the pacing, and the moves, with times attached.
  • You want to recreate the feel of a reference without copying its assets, logos or music.
  • You are cutting a film for Stylr, Qufy, Claude RTL, or the next tool, and you want the reference read the same way every time.

Install

This installs video-reader only. The other skills stay out.

  1. Open Claude Code.
  2. Paste line 1 and press Enter. This adds the catalog. Do it once, then skip it next time.
  3. Paste line 2 and press Enter. This installs this skill.
/plugin marketplace add grw-lab/skills
/plugin install video-reader@grw-lab

Claude confirms the install in the chat. The next message can use the skill.

Every skill instead: keep line 1, and use this as line 2:

/plugin install grw-skills@grw-lab

On claude.ai, zip the video-reader folder and upload it in Settings → Capabilities → Skills.

In Cursor or another agent, point it at video-reader/SKILL.md.

The script also needs ffmpeg on your machine, plus Python packages numpy and Pillow.

Usage

  • "Look at this video and tell me how it is cut."
  • "Break down the pacing. Where are the cuts, and do they land on the sound?"
  • "Make a film in this spirit for our tool. Take the timing, not their logo or their music."
  • "This is one continuous shot. Describe the move."

You can also run the script on its own:

python ~/.claude/skills/video-reader/scripts/analyze_video.py video.mp4 --out ./out

To pull one more frame:

python ~/.claude/skills/video-reader/scripts/analyze_video.py video.mp4 --out ./out --frame 2.5

Example

A generated 6 second vertical reel, three flat colors, with a click every half second. The script printed:

reel.mp4 · 6.0s · 720x1280 · vertical · 30.0fps · audio yes
cuts: 2.0, 4.0
contact sheet: 24 tiles at 4 fps
keyframes: 1.00s, 3.00s, 5.00s
tempo: about 120 BPM from regular onsets

The same file with the audio removed comes back audio no and no BPM. A landscape clip with no color change comes back cuts: none.

What's inside

File What it's for
SKILL.md The steps the agent follows, and the rules against inventing cuts, tempo or typography.
references/report-template.md The breakdown the agent fills in after reading the images and summary.json.
scripts/analyze_video.py The pipeline. Needs ffmpeg, numpy and Pillow. Writes only inside --out.
README.md Install and command line usage for a person running the script directly.

FAQ

Can it hear the voiceover? No. Loudness, onset times and a BPM (only when the onsets are regular) are the whole audio report. Speech, lyrics, instruments and genre have to come from you.

Why does a quiet talking video have no BPM? The tempo is the gap between sudden volume jumps. Speech is uneven, so the script refuses the BPM instead of printing a wrong one. The median gap is still in summary.json.

What if the video is one take? cuts is empty. The report should say it is one continuous shot, or that the transitions are too soft to count as cuts.

Will it copy the reference? The report has a "Don't copy" list: assets, logos, another product's interface, and the music. The skill is for reading the timing and the moves.

What do I need installed? ffmpeg (which includes ffprobe), Python 3.9 or newer, numpy and Pillow. No account and no network.