video-analyse

name

video-analyse

description

Analyse a video file by extracting timestamped frames and a Whisper transcript, then producing a structured analysis document. Use when the user asks to analyse a video, summarise a recording, get meeting notes from a video, or understand what happens in a video file. Triggers on phrases like "analyse this video", "video analysis", "what happens in this video", "summarise this recording", "meeting notes from this video", "video-analyse", "break down this video".

Video Analyse

Instructions

You are analysing a video file for the user. Follow these steps precisely.

Step 1: Find the video

If the user provided a path as an argument, use that. Otherwise, look for video files in the current directory:

ls -1 *.{mp4,mov,mkv,avi,webm,m4v} 2>/dev/null

If multiple videos are found, ask the user which one. If none are found, ask for a path.

Step 2: Check dependencies

Before running anything, verify the tools are available:

which ffmpeg && python3 -c "import whisper; print('whisper OK')"

If ffmpeg is missing: tell the user to run brew install ffmpeg (macOS) or apt install ffmpeg (Linux). If whisper is missing: tell the user to run pip install openai-whisper.

Step 3: Determine frame rate

The script at ~/.claude/skills/video-analyse/analyse_video.py has auto mode built in. Unless the user specified a frame rate, use --auto which selects based on duration:

Duration	FPM	Interval	Approx frames
< 5 min	2	30s	~10
5–15 min	1	60s	~15
15–45 min	0.5	2 min	~15–22
45–90 min	0.33	3 min	~15–30
90+ min	0.2	5 min	~18–24

If the user asked for a specific rate, pass --fpm <value> instead.

Step 4: Run the extraction pipeline

python3 ~/.claude/skills/video-analyse/analyse_video.py "<video_path>" --auto

This creates a {video-stem}-analysis/ directory containing:

• summary.md — frame index and metadata
• transcript.md — timestamped Whisper transcript
• frames/ — numbered JPG stills with timestamps

The script will take a while for long videos (transcription is the slow part). Let the user know it's running.

Step 5: Read all outputs

Once the script completes, read everything into context:

1. Read {video-stem}-analysis/summary.md
2. Read {video-stem}-analysis/transcript.md
3. Read every frame image in {video-stem}-analysis/frames/ (use the Read tool on each .jpg — Claude can see images)

This gives you full multimodal context: what was said (transcript) and what was shown (frames).

Step 6: Write the analysis

Create analysis.md in the current working directory with the following structure:

# Video Analysis: {filename}

**Duration:** {duration}
**Date analysed:** {today}

## Participants

List everyone visible or audible in the video. Note speaker identification confidence.

## Timeline Summary

For each major section/topic change, write a timestamped entry:

### [{timestamp range}] Section Title

What happened in this segment. Reference both audio (what was said) and visual (what was on screen) evidence.

> Key quote if relevant — [{timestamp}]

## Key Decisions

Bulleted list of decisions made (if applicable, e.g. for meetings).

## Action Items

- [ ] Action item with owner if identifiable — [{timestamp}]

## Visual Reference Points

Notable moments where the screen content is important:
- [{timestamp}] Description of what's shown and why it matters

## Whisper Accuracy Notes

Flag any sections where the transcript seems unreliable (garbled text, obvious misheard words, low-confidence sections).

Adapt the structure to fit the video type — a meeting needs decisions/action items, a tutorial needs step-by-step breakdowns, a presentation needs slide-by-slide coverage, etc.

Step 7: Stay in context

After writing the analysis, tell the user it's done and that you have the full transcript + frames in context. Offer to:

• Answer questions about specific moments
• Find when a particular topic was discussed
• Extract more detail from any section
• Compare what was said vs what was shown on screen

Options

• --fpm <number>: Override the auto frame rate (e.g. --fpm 2 for one frame every 30s)
• --whisper-model <size>: Use a different Whisper model (tiny, base, small, medium, large). Default is base. Use small or medium for better accuracy on difficult audio.
• --no-transcribe: Skip Whisper transcription (frames only). Useful if the video has no speech or you only care about visuals.