demo-video-watcher - SKILL.md Agent Skill

name: demo-video-watcher description: Watch demo videos linked in GitHub issue/PR comments. Downloads the video, transcribes audio with whisper and extracts frames. Use when encountering a video URL in a teamniteo or mayetrx repository. argument-hint: "" allowed-tools:

Bash(curl *)
Bash(file *)
Bash(gh api*)
Bash(gh auth token*)
Bash(ls *)
Bash(mkdir *)
Bash(nix-shell *)
Glob
Read

Demo Video Watcher

Extract context from demo videos linked in GitHub issues or PR comments. These are typically User Story demos — screen recordings with narration showing the implemented feature.

Parse $ARGUMENTS to extract a video URL. Accept:

GitHub comment URL: https://github.com/<owner>/<repo>/issues/<n>#issuecomment-<id> — fetch the comment body to find the video attachment URL with gh api repos/<owner>/<repo>/issues/comments/<id> --jq '.body'.
GitHub PR comment URL: https://github.com/<owner>/<repo>/pull/<n>#issuecomment-<id> — #issuecomment-<id> is an issue comment even on a PR, so use the same endpoint: gh api repos/<owner>/<repo>/issues/comments/<id> --jq '.body'.
Direct video URL: https://github.com/user-attachments/assets/<uuid>

If $ARGUMENTS is empty, use AskUserQuestion to ask for the URL.

GitHub user-attachment URLs require authentication. Download with:

curl -sL -H "Authorization: token $(gh auth token)" \
  -o /tmp/demo-video.mp4 "<VIDEO_URL>"

Verify the download:

ls -lh /tmp/demo-video.mp4 && file /tmp/demo-video.mp4

If the file is tiny (< 1KB) or contains "Not Found", the download failed.

Audio transcription gives the best context. Always do this first.

Check for audio track:

nix-shell -p ffmpeg --run "ffprobe -i /tmp/demo-video.mp4 2>&1 | grep Audio"

Extract audio:

nix-shell -p ffmpeg --run \
  "ffmpeg -y -i /tmp/demo-video.mp4 -vn -acodec pcm_s16le -ar 16000 -ac 1 /tmp/demo-audio.wav"

Transcribe with whisper:

nix-shell -p openai-whisper --run \
  "whisper /tmp/demo-audio.wav --model tiny --language en --output_format txt --output_dir /tmp/demo-transcript"

Present the transcription with timestamps.

After audio transcription, extract frames to capture visual context.

mkdir -p /tmp/demo-frames
nix-shell -p ffmpeg --run \
  "ffmpeg -i /tmp/demo-video.mp4 -vf 'fps=1' -q:v 2 /tmp/demo-frames/frame_%03d.jpg"

Use fps=0.5 for longer videos (> 60s) or fps=2 for short ones (< 10s) where detail matters.

View frames with the Read tool. Start with a sample (every 5th frame) to get an overview, then fill in gaps if needed.

Provide a concise summary of what the demo shows and append the full transcript and key frames.