viewing-videos
Skill6 starsChanged 23 days agoMITView on GitHub
Use when the user wants Claude to view, watch, inspect, or extract information from a video file (mp4, mov, webm, mkv, screen recording, Zoom/Meet/Teams meeting recording, demo, screencast, phone clip) — e.g. "can you see this video", "what numbers are shown on screen", "what slide was up when X was said", "check the recording", or when a transcript exists but on-screen content (figures, slides, UIs, dashboards) must be read visually. Claude cannot ingest video natively; this skill is how Claude sees one.
- Deletes or force-pushes
In this repository29 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- skills/prompt-engineer/references/claude.mdCLAUDE.md
- brand-assetsSkill
- brand-systemSkill
- codex-reviewSkill
- deck-to-videoSkill
- elevenlabs-sttSkill
- elevenlabs-ttsSkill
- excalidraw-diagramsSkill
- gdoc-syncSkill
- generating-video-clipsSkill
- gsheetsSkill
- gslides-syncSkill
- image-generationSkill
- namecheap-domainsSkill
- narrating-pptxSkill
- office-renderSkill
- pair-intelligenceSkill
- presentation-generatorSkill
- producing-animated-episodesSkill
- prompt-engineerSkill
- proton-passSkill
- reviving-sessionsSkill
- self-presenting-decksSkill
- tavily-extractSkill
- tech-design-docSkill
- writing-plainlySkill
- writing-project-instructionsSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 53
If the question lives in one segment, sweep coarse first, then densify only that range: ffmpeg -ss <start> -to <end> -i video.mp4 -vf …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Viewing Videos
Overview
Claude can't watch video, but it reads images well. Turn the video into the fewest frames that answer the question with ffmpeg, triage them on a contact sheet, then Read the interesting ones at full resolution. Requires ffmpeg + ImageMagick montage (both via Homebrew).
Step 0: Probe first
ffprobe -v error -show_entries format=duration:stream=codec_type,width,height -of default=noprint_wrappers=1 video.mp4Duration picks the strategy; resolution sets expectations — 1080p+ screen content reads fine, at 720p big KPI numbers are fine but dense tables/spreadsheets need the crop+upscale step below.
Pick a strategy
| Situation | Strategy |
| You know the moment (timestamped transcript, user says "at 12:34") | Targeted single frames |
| Screen shares / slides / UI demos, unknown moments | Scene-change sweep (default) |
| Continuous motion (camera footage, gameplay) — scenes never "change" | Interval sampling (see frame budget below) |
Targeted frame
-ss before -i = fast seek:
ffmpeg -ss 12:34 -i video.mp4 -frames:v 1 /tmp/frames/t12m34s.pngScene-change sweep → timestamp names → contact sheet
rm -rf /tmp/frames && mkdir -p /tmp/frames
ffmpeg -hide_banner -i video.mp4 -vf "select='gt(scene,0.2)',showinfo" -fps_mode vfr /tmp/frames/frame_%03d.png 2> /tmp/frames/log.txt
cd /tmp/frames && grep -o 'pts_time:[0-9.]*' log.txt | cut -d: -f2 | awk '{n++; t=int($1); printf "mv frame_%03d.png t%02dm%02ds.png\n", n, t/60, t%60}' | sh
montage -label '%t' -font /System/Library/Fonts/Helvetica.ttc t*.png -tile 6x -geometry 320x180+4+4 -pointsize 18 contact_%d.pngRead contact_0.png to triage (timestamps are printed under each tile), then Read only the frames that matter at full resolution. Threshold 0.2 works for meetings: webcam segments cluster harmlessly, screen shares fire exactly on slide/page changes. Footage with constant motion → hundreds of frames; switch to interval sampling.
Frame budget
Aim for ≤ ~100 frames per sweep — enough to cover a long video, few enough to triage on 1–2 contact sheets. Tune to hit it, don't accept whatever falls out:
- Scene detection over-firing? Raise the threshold (
gt(scene,0.3)–0.4). - Interval sampling: derive the interval from duration instead of fixing it —
interval = max(5, duration_seconds / 100), i.e.-vf fps=1/$INTERVAL. A 10-min clip samples every ~6s; a 3-hour recording every ~108s (not 1080 frames at a blind 1/10s). - If the question lives in one segment, sweep coarse first, then densify only that range:
ffmpeg -ss <start> -to <end> -i video.mp4 -vf fps=1/2 ....
Small text unreadable? Crop + upscale
Re-extract just the region at 2× rather than squinting at the full frame:
ffmpeg -ss 17:40 -i video.mp4 -frames:v 1 -vf "crop=900:500:80:120,scale=2*iw:-1" /tmp/frames/zoom_t17m40s.pngGet crop coords by reading the full frame first.
With a transcript
Timestamped transcript (.vtt) → go targeted: find where figures/screens are discussed, seek those times. Plain text transcript (no timestamps) → sweep, then correlate by content order. If the user has a choice, ask for the timestamped export.
Common mistakes
- Globs in compound commands: zsh aborts the whole
&&chain on a no-match glob (rm -f *.png→ ffmpeg never runs). Userm -rf dir && mkdir -p dir. montagewithout-font: fails with "unable to read font" on macOS. Always pass-font /System/Library/Fonts/Helvetica.ttc.- Blind 1-frame-per-second sampling: 30 min → 1800 redundant frames. Scene detection or 1/10s intervals first; densify only where needed.
- Skipping the contact sheet: reading 60+ full frames one by one burns context. Triage on the sheet, read few.
- Trusting frame numbers instead of timestamps: rename to
tMMmSSs.pngimmediately, or you can't correlate with the transcript or seek back for a better frame.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

