extract-video-frames - SKILL.md Agent Skill

name: extract-video-frames description: Extracts frames and timestamped audio segments from video files (GIF, MP4, MOV) at configurable intervals and stores them in a directory with a manifest file. Use when analyzing video content, preparing frames for visual review, extracting audio for transcription, or creating frame+audio sequences for another agent to process. argument-hint: [video-path] [interval-seconds] [output-dir]

Extract frames and aligned audio segments from video files (GIF, MP4, MOV) at specified intervals using ffmpeg. Outputs PNG frames, per-segment AAC audio clips, a full continuous audio track, and a JSON manifest containing timestamps and paths -- ready for handoff to another agent for visual and audio analysis. Extract frames and audio from a video:

# Extract 1 frame + audio segment per second (default)
~/.claude/skills/extract-video-frames/scripts/extract-frames.sh input.mp4

# Extract 1 frame + audio segment every 2 seconds
~/.claude/skills/extract-video-frames/scripts/extract-frames.sh input.mp4 2

# Specify custom output directory
~/.claude/skills/extract-video-frames/scripts/extract-frames.sh input.mp4 1 ./my-frames

The script creates:

frames/ directory with PNG files named frame_001.png, frame_002.png, etc.
frames/audio_001.aac, audio_002.aac, etc. (when audio exists, aligned 1:1 with frames)
frames/full_audio.aac (complete audio track, when audio exists)
frames/manifest.json with frame and audio metadata for the reviewing agent

1. **Verify ffmpeg is available** ```bash which ffmpeg || echo "ffmpeg not found - install with: brew install ffmpeg" ```

Run extraction script
```
~/.claude/skills/extract-video-frames/scripts/extract-frames.sh <video-path> [interval-seconds] [output-dir]
```
- video-path: Path to GIF, MP4, or MOV file (required)
- interval-seconds: Extract one frame every N seconds (default: 1)
- output-dir: Where to store frames and audio (default: ./frames)

Review manifest The manifest.json contains:

{
  "source": "recording.mp4",
  "source_path": "/path/to/recording.mp4",
  "interval_seconds": 5,
  "total_frames": 12,
  "has_audio": true,
  "audio_codec": "aac",
  "total_audio_segments": 12,
  "full_audio_path": "full_audio.aac",
  "output_directory": "./frames",
  "frames": [
    {
      "index": 1,
      "timestamp": "00:00:00",
      "timestamp_seconds": 0,
      "path": "frame_001.png",
      "audio_path": "audio_001.aac"
    },
    {
      "index": 2,
      "timestamp": "00:00:05",
      "timestamp_seconds": 5,
      "path": "frame_002.png",
      "audio_path": "audio_002.aac"
    }
  ]
}

Hand off to reviewing agent Pass the output directory path to the reviewing agent. The agent can read manifest.json to understand the frame sequence with audio alignment and use the Read tool to analyze individual frames.

- **GIF**: Animated GIFs (extracts frames; no audio stream, gracefully skipped) - **MP4**: Standard video format (frames + audio) - **MOV**: QuickTime format (frames + audio) - **Other**: Any format ffmpeg supports (AVI, WebM, MKV, etc.) ``` output-dir/ ├── manifest.json # Frame + audio metadata for reviewing agent ├── full_audio.aac # Complete audio track (when audio exists) ├── frame_001.png # First extracted frame ├── frame_002.png # Second extracted frame ├── audio_001.aac # Audio segment for frame 1 (when audio exists) ├── audio_002.aac # Audio segment for frame 2 └── ... ``` **Format**: AAC (.aac) -- chosen for universal compatibility, small file size, and broad tooling support.

Codec strategy:

If the source audio is already AAC, segments are stream-copied (no re-encoding) for speed and quality preservation.
If the source audio is any other codec (e.g., PCM, MP3, Opus), segments are re-encoded to AAC at 128kbps.

Alignment semantics: Each audio_NNN.aac segment covers the same time window as its corresponding frame_NNN.png. For a 5-second interval, audio_001.aac covers 0:00-0:05, audio_002.aac covers 0:05-0:10, etc.

No-audio handling: When the source has no audio stream (GIFs, silent videos), has_audio is false, no audio files are created, and all audio_path fields in the manifest are null. The frame extraction works identically regardless.

Full audio track: full_audio.aac contains the entire audio from the source file as a single continuous track, useful for full transcription or background listening.

When passing frames and audio to another agent, include:

Output directory path: Where frames and audio are stored
Manifest location: {output-dir}/manifest.json
Context: What the reviewing agent should look for

Example prompt for reviewing agent:

Analyze the frames and audio extracted from the screen recording at ./frames.
The manifest at ./frames/manifest.json lists all frames with timestamps and paired audio segments.
The full audio track is at ./frames/full_audio.aac for continuous listening or transcription.
Look for: [specific things to identify or analyze]

Example prompt for audio-focused agent:

Transcribe the audio segments from the recording at ./frames.
Read ./frames/manifest.json to get the list of audio files with timestamps.
For each audio_NNN.aac segment, provide a timestamped transcript.
The full continuous audio is also available at ./frames/full_audio.aac.

Frame and audio extraction is complete when: - Output directory contains PNG frames - `manifest.json` exists with valid frame and audio metadata - Frame count matches expected based on video duration and interval - Frames are readable by the Read tool for visual analysis - When source has audio: audio segments exist (1:1 with frames), `full_audio.aac` exists, `has_audio` is `true` - When source has no audio: `has_audio` is `false`, `audio_path` values are `null`, no audio files created