video-understanding
Skill3 starsChanged 10 days agoMITView on GitHub
Turn a video file into a full AI understanding: use video-understanding.sh for frames + transcript, then review as agent to produce understanding.md. Local ffmpeg + dottie-talk (parakeet) or whisper.cpp. Use when user runs /video-understanding, or says "process this video with full understanding", "get frames and transcript then analyze".
2 other projects use a version of this.
Adopted by
Most recently taken up by
- zxyycom/skillstheir own version, first seen Nov 8, 2026
- ai-workspace-lab/xworkspace-core-skillstheir own version, first seen Oct 17, 2026
Other versions2 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 2
- ai-workspace-lab/xworkspace-core-skills★ 8same name
- zxyycom/skills★ 2same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
The file
Show its rules and text
We split this skill into the rules it gives an agent
13 rules. About 0 are common, 6 are distinctive, and 7 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
7
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 31
Do not freestyle alternate whisper installs unless the user asks — use local-ai-cli as the source of truth.
- Very specific: line 39
Use VU_PROFILE=local (default, fully local whisper) or grok.
- Very specific: line 40
Default frame interval: 500ms (use --interval 500ms / 0.5 / 2 etc).
- Very specific: line 42
Use --interval <val> or DEFAULT_INTERVAL.
- Very specific: line 46
Prefer local extraction: ffmpeg + dottie-talk (parakeet), else whisper.cpp (whisper-cli).
- Distinctive: line 47
Never yt-dlp or YouTube links.
- Very specific: line 48
Use video-understanding (on PATH → skill video-understanding.sh) for stage 1 (supports --interval 500ms etc).
- Very specific: line 50
Follow the generated AGENT.md exactly for stage 2.
- Distinctive: line 54
Run the script on a local video file:
- Very specific: line 61
Default interval: 500ms (override via --interval, DEFAULT_INTERVAL in profile/env/positional.
- Distinctive: line 66
After the script runs, point to the output folder:
- Distinctive: line 72
Read every frame image in order (filenames = timestamps).
- Distinctive: line 92
After extraction, the agent works entirely from the output folder files.
≈ “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
video-understanding
Full pipeline for turning videos into agent-understandable output: timestamped frames + SRT transcript, followed by agent review that correlates visuals with speech and writes a complete understanding.md.
Stage 1 is mechanical extraction. Stage 2 is the agent analysis. Generic for any video content.
Skill root: ~/.agents/skills/video-understanding/ (also linked as Claude skills via ~/.agents/skills).
Prerequisites (local STT + ffmpeg)
Stage 1 needs ffmpeg + a local STT: dottie-talk 0.26+ (parakeet, used automatically when dottie-talk transcribe --json works; point DOTTIE_TALK_BIN at a checkout if the PATH copy is older) or whisper.cpp (whisper-cli) + a ggml model.
Install / repair the local stack with local-ai-cli (builds whisper.cpp + models, puts tools on PATH):
git clone https://github.com/stevederico/local-ai-cli.git && cd local-ai-cli && bash setup.sh
# ensures whisper-cli, ffmpeg (via brew), model under ~/.local/opt/whisper.cpp/Do not freestyle alternate whisper installs unless the user asks — use local-ai-cli as the source of truth.
Optional: BYOK mode (install-free)
The CLI also supports VU_PROFILE=byok (xAI STT + optional Mux frames) when keys are set — see repo README. Stage-1 extraction script still implements this; prefer local + local-ai-cli unless the user asks for BYOK.
Configuration
Use VU_PROFILE=local (default, fully local whisper) or grok.
Default frame interval: 500ms (use --interval 500ms / 0.5 / 2 etc).
Profiles: ~/.agents/skills/video-understanding/config/profiles/<name>.sh (or use the video-understanding command on PATH)
Use --interval <val> or DEFAULT_INTERVAL. Env/flag override.
See config/profiles/local.sh for the current local-focused setup.
Core rules
- Prefer local extraction: ffmpeg + dottie-talk (parakeet), else whisper.cpp (
whisper-cli). - Never yt-dlp or YouTube links.
- Use
video-understanding(on PATH → skillvideo-understanding.sh) for stage 1 (supports --interval 500ms etc). Explicit path:~/.agents/skills/video-understanding/video-understanding.sh - Follow the generated
AGENT.mdexactly for stage 2. - Output goes to
<video>_understand/(or custom dir). The agent then works inside that folder.
Stage 1: Mechanical extraction
Run the script on a local video file:
video-understanding /path/to/video.mp4 [interval] [output_dir] [--interval <val>]
# default 500ms. e.g. --interval 500ms or --interval 2- Default interval: 500ms (override via --interval, DEFAULT_INTERVAL in profile/env/positional; accepts 500ms, 0.5 etc)
- Outputs:
frames/tNNmNNs.jpg(or ...sNNNms.jpg for subsec),transcript.srt,transcript.txt,transcript.json,manifest.json,AGENT.md - Env:
VU_MODEL=large-v3-turbo,WHISPER_MODEL=...etc.
Stage 2: Agent review
After the script runs, point to the output folder:
"read <output>/AGENT.md and do it."
The AGENT.md instructs:
- Read transcript.srt fully.
- Read every frame image in order (filenames = timestamps).
- Correlate on-screen content (text, UI, visuals, actions) with spoken words at each time.
- Note scene changes, on-screen text not in audio, key demos, etc.
- Write
understanding.mdwith:- Summary
- Timeline table (time | on-screen | spoken)
- Key visuals
- Full takeaways
- Corrected transcript (if frames reveal the audio transcription got a name/term/number wrong; list the fixes).
Full example flow
- User: "video understand this: ~/Movies/talk.mov"
- Agent:
video-understanding ~/Movies/talk.mov --name talk --interval 500ms - Agent: read the generated AGENT.md in
talk_understand/(or custom suffix) and produce understanding.md - Output the summary or full understanding.md to user.
The tool works for any video content. Use it to deeply analyze talks, demos, interviews, etc.
Tips
- Adjust interval: --interval 500ms (or 0.5) for fast cuts/action, 1-3s for typical, larger (5s+) for slow talking-head.
- After extraction, the agent works entirely from the output folder files.
- Stage 1 is deterministic given the same model + inputs.
Transcript quality (adopted from x-studio patterns)
- Script now applies post-processing: removes non-speech markers, simple repetitions.
- Outputs
segments.jsonwith structured {timestamp, speaker, text} for better agent input. - Whisper tuned with DTW, no-speech-thold 0.68, logprob -0.9, VAD if available.
Config profiles
- Source with
VU_PROFILE=local(default) orgrok. - Profiles live in
~/.agents/skills/video-understanding/config/profiles/<name>.sh - Local profile: fully local whisper via local-ai-cli / whisper.cpp.
- Grok profile: same local extraction; agent may resolve X/media differently.
- Profiles also control DEFAULT_INTERVAL and DEFAULT_OUTDIR_SUFFIX (default 500ms).
- Override via --interval, env (DEFAULT_INTERVAL=500ms), or positional. Env takes precedence.
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.

