ffmpeg-analyse-video

FFmpeg Video Analysis

Extract frames from video files with ffmpeg. Delegate frame reading to sub-agents to preserve the main context window. Synthesise a structured timestamped summary from text-only sub-agent reports.

Architecture: Context-Efficient Sub-Agent Pipeline

Problem: Reading dozens of images into the main conversation context consumes most of the context window, leaving little room for synthesis and follow-up.

Solution: A 3-phase pipeline:

Main Agent                          Sub-Agents (disposable context)
──────────                          ──────────────────────────────
1. ffprobe metadata        ───►
2. ffmpeg frame extraction ───►
3. Split frames into batches ──►   4. Read images (vision)
                                      Write text descriptions
                                      to batch_N_analysis.md
5. Read text files only    ◄───    (context discarded)
6. Synthesise final output

Images only ever exist inside sub-agent contexts. The main agent only reads lightweight text files. This cuts context usage by ~90%.

1. Prerequisites

which ffmpeg

If either is missing, show platform-specific install instructions and STOP:

• macOS: brew install ffmpeg
• Ubuntu/Debian: sudo apt install ffmpeg
• Windows: choco install ffmpeg or winget install ffmpeg

2. Setup Temp Directory

3. Extract Video Metadata

Extract and report: duration, resolution (width x height), fps, codec, file size, whether audio is present.

If no video stream is found, report "audio-only file" and STOP. If file size > 2GB, warn the user and suggest analysing a time range with -ss START -to END.

4. Extract Frames

Choose strategy based on duration:

For thumbnail filter, calculate SEGMENT_FRAMES = total_frames / 60 to cap output at ~60 frames.

Fallbacks:

• Scene detection yields 0 frames → retry with interval at 1 frame/5s
• More than 100 frames extracted → subsample evenly to 80
• Frame extraction fails → try the next simpler strategy (scene → interval, keyframe → interval)

Time range analysis: When user specifies a range, prepend -ss START -to END before -i. Higher detail mode: If requested, double the fps rate and lower scene threshold to 0.2.

After extraction, list all frame files and calculate each frame's timestamp from its sequence number and the extraction rate.

5. Delegate Frame Analysis to Sub-Agents

This is the critical context-saving step. Do NOT read frame images in the main conversation. Instead, split frames into batches and delegate each batch to a sub-agent.

5a. Prepare Batch Manifest

Split the extracted frame file list into batches of 8-10 frames each. For each batch, record:

• Batch number (1, 2, 3, ...)
• Frame file paths (absolute)
• Frame timestamps (calculated from sequence number)
• Output file path: TMPDIR/batch_N_analysis.md

5b. Spawn Sub-Agents

For each batch, spawn a sub-agent with the prompt below. Launch all batches in parallel where the tool supports it — they are fully independent.

Sub-Agent Prompt Template

Use this prompt verbatim, substituting the placeholders:

How to Spawn

Use whatever sub-agent, background task, or independent agent mechanism your tool provides. The requirements are simple — each sub-agent needs to:

1. Read image files (the frame JPEGs)
2. Write a text file (the batch analysis markdown)

Launch all batches in parallel if your tool supports it — they are fully independent with no shared state.

If your tool has no sub-agent mechanism, fall back to reading frames directly in the main context but limit to 20 frames maximum and warn the user about context usage.

5c. Collect Results

After all sub-agents complete, read the text analysis files. These are lightweight markdown — no images enter the main context.

Read each batch_N_analysis.md file in order. These contain only text descriptions — the context cost is minimal compared to reading the original images.

6. Synthesise Output

Using only the text from the batch analysis files, perform synthesis in the main context:

1. Merge all frame descriptions into a single chronological timeline
2. Group frames into natural segments (same scene, slide, or screen)
3. Detect the dominant content type across all batches
4. Identify 3-7 key moments
5. Extract all quoted text, prompts, or commands the user typed
6. Write a 2-5 sentence narrative summary

Format the output as:

7. Cleanup

Remove the temp directory after output is complete:

Skip cleanup if the user asks to keep frames.

Advanced Options

• Time range: "Analyse 2:00 to 5:00 of video.mp4" → use -ss 120 -to 300
• Higher detail: "Analyse in high detail" → double frame rate, lower scene threshold to 0.2
• Focus area: "Focus on the code shown" → prioritise text/code extraction in sub-agent prompts
• Sprite sheet: For a visual overview, generate a contact sheet:

  ffmpeg -hide_banner -y -i INPUT -vf "select='not(mod(n,EVERY_N))',scale='min(320,iw)':-2,tile=5xROWS" -frames:v 1 DIR/sprite.jpg

Error Handling

• ffmpeg not found → install instructions per platform, STOP
• No video stream → report audio-only, STOP
• Scene detection yields 0 frames → fallback to interval
• Too many frames (>100) → subsample to 80
• Large files (>2GB) → warn, suggest time range
• Sub-agent fails or times out → read that batch's frames directly as fallback, warn about context usage
• Frame read failure in sub-agent → skip frame, note gap in batch analysis file