Capture and describe video frames

Step 1: Capture frames

Run the capture_video_frames.py script:

uv run .agents/skills/capture-video-frames/capture_video_frames.py <youtube_url> <output_dir> [--interval SECONDS]

Arguments

• youtube_url (required): YouTube video URL (same formats accepted by the extract-transcript skill).
• output_dir (required): Directory to save frames and the manifest file. Created if it doesn't exist.
• --interval (optional): Seconds between captured frames. Defaults to 30.

Outputs

• frame_0000.png, frame_0030.png, … — PNG images named by their timestamp in seconds (zero-padded to 4 digits).
• frames_manifest.md — A markdown file listing each frame with its timestamp and a placeholder for descriptions.

Example frames_manifest.md:

| File | Timestamp | Description |
|------|-----------|-------------|
| frame_0000.png | [00:00] | |
| frame_0030.png | [00:30] | |
| frame_0060.png | [01:00] | |

Prerequisites

• yt-dlp: brew install yt-dlp or pip install yt-dlp
• ffmpeg: brew install ffmpeg or apt-get install ffmpeg

Step 2: Describe frames using the describe-frame subagent

After capturing frames, describe each frame by running the describe-frame custom agent as a subagent. Each subagent invocation gets an isolated context, so frame images won't accumulate and exhaust the context window.

The describe-frame agent is defined in .github/agents/describe-frame.md.

Procedure

1. Read frames_manifest.md from the output directory to get the full list of frames.
2. For each frame, run the describe-frame agent as a subagent with a prompt that includes:
3. The absolute path to the current frame image to view.
4. The absolute path to the previous frame image to view (if one exists).
5. The previous frame's description as text (if one exists).
6. The subagent will return a plain-text description (or (same as previous) if the frame is essentially identical to the previous one).
7. After each subagent returns, update the Description column for that row in frames_manifest.md immediately.
8. Continue until all frames are described.

Subagent prompt template

Use this as the prompt when invoking the describe-frame subagent (fill in the bracketed values):

Describe the current frame image at: [CURRENT_FRAME_ABSOLUTE_PATH]

[If previous frame exists, include these two lines:]
The previous frame image is at: [PREVIOUS_FRAME_ABSOLUTE_PATH]
The previous frame was described as: "[PREVIOUS_DESCRIPTION]"

Example output

After describing all frames, frames_manifest.md should look like:

| File | Timestamp | Description |
|------|-----------|-------------|
| frame_0000.png | [00:00] | Title slide introducing "Building RAG apps with Python" |
| frame_0030.png | [00:30] | Speaker showing the agenda with four main topics |
| frame_0060.png | [01:00] | (same as previous) |
| frame_0090.png | [01:30] | Architecture diagram of a retrieval-augmented generation pipeline |

Step 3: Deduplicate frames and select best speaker faces

After all frames are described, groups of consecutive (same as previous) rows represent the same visual content captured at different moments. Within each group, speaker faces may differ — eyes open vs closed, mouth open vs closed, facing camera vs turned away.

Goal

For each group of duplicate frames, keep only one frame — the one with the best speaker face quality — and remove the rest.

Criteria for best face (in priority order)

1. The current speaker's mouth should be open (mid-speech). If you know who is speaking at that timestamp (from a transcript), prioritize that speaker.
2. Eyes open — no mid-blink frames.
3. Facing camera — not turned sideways or looking down.
4. If no speakers are visible (e.g., full-screen demo or slide without webcam feeds), all frames in the group are equivalent — keep the first one.

Procedure

1. Identify all groups of consecutive rows where the description is (same as previous). Each group starts with the "anchor" frame (the one with an actual description) followed by one or more (same as previous) rows.
2. For each group, run the describe-frame subagent to compare faces across the anchor frame and each duplicate. Use this prompt template:

Compare these two frames focusing ONLY on the speaker faces visible in webcam feeds. Which frame has better speaker faces — eyes open, facing camera, mouth open (mid-speech), not mid-blink or turned away?

Frame A: [ANCHOR_FRAME_ABSOLUTE_PATH]
Frame B: [DUPLICATE_FRAME_ABSOLUTE_PATH]

Reply with ONLY one of:
- "A BETTER" if the anchor frame has better speaker faces
- "B BETTER" if the duplicate frame has better speaker faces
- "EQUAL" if both are equivalent
- "NO SPEAKERS" if no speaker faces are visible in either frame

Then add a brief reason.

1. After comparing all duplicates in a group against the anchor (and the current best), determine the single best frame.
2. If the best frame is NOT the anchor:
3. Move the anchor's description to the best frame's row.
4. Add a face-quality note to the description, e.g., (better speaker faces than frame_XXXX: eyes open, facing camera)
5. Remove all other (same as previous) rows from the manifest.
6. If the anchor was already the best, just remove the duplicate rows.

Recapturing frames for closed mouths

After deduplication, if the best frame in a group still has the speaking person's mouth closed (both speakers have mouths closed), try recapturing at nearby timestamps:

1. Download the video if not already available: bash yt-dlp -f "bestvideo[height<=720]" --no-playlist -o "<output_dir>/video.%(ext)s" "<youtube_url>"
2. Capture alternative frames at +2s, +5s, +8s, and +10s offsets from the frame's timestamp: bash ffmpeg -ss <SECONDS> -i <output_dir>/video.mp4 -frames:v 1 -q:v 2 <output_dir>/alt_<FRAME>_<SECONDS>.png -y
3. Use the describe-frame subagent to check if the speaker's mouth is open in any alternative, AND that the slide/demo content is still the same.
4. If a better alternative is found, replace the frame file (cp alt_XXXX.png frame_XXXX.png).
5. Clean up: rm -f <output_dir>/alt_*.png <output_dir>/video.mp4