/video-watch

Video Watch

Prepare video artifacts that GitHub Copilot can inspect: captions, frame contact sheets, a frame index, a manifest, and a prompt packet.

This skill is inspired by bradautomates/claude-video, but it targets GitHub Copilot workflows. Do not assume Claude Code Read parallelism. Generate files first, then inspect only the artifacts needed for the user's question.

When to Use

• The user asks to watch, inspect, summarize, compare, or diagnose a video.
• The input is a video URL supported by yt-dlp, or a local .mp4, .mov, .mkv, or .webm file.
• The task depends on both what is said and what appears on screen, such as bug reproductions, product demos, launch videos, tutorials, webinars, ads, or screen recordings.

Not the Best Fit

• Do not bypass DRM, paywalls, authentication, or site terms.
• Do not request or store cookies, bearer tokens, signed URLs, or private media credentials.
• If an official API or existing transcript answers the task without video processing, prefer that lighter path.

Core Workflow

1. Run preflight if dependencies are uncertain: python .github/skills/video-watch/scripts/preflight.py Use python .github/skills/video-watch/scripts/preflight.py --url-mode for URL inputs.
2. Produce artifacts: python .github/skills/video-watch/scripts/video_watch.py <url-or-path> --question "<what to answer>" Use --transcript-file <path> when another tool, Azure Speech, or Speech Translator Desktop Plus already produced a transcript. For private or authenticated media, prefer --detail transcript --metadata-only --transcript-file <path> and do not pass cookies, bearer tokens, or signed URL credentials.
3. Read artifacts in this order: manifest.json → prompt.md → transcript.md → frame-index.md → contact-sheet.jpg → selected files in frames/ only if needed. manifest.json records SHA-256 for local source bytes, or for downloaded URL video bytes when a download occurred. Metadata-only URL runs leave the hash null rather than fetching media.
4. Answer from the artifacts. State when the answer is transcript-only, frame-only, or limited by sparse sampling.
5. For named moments, rerun with --start / --end to focus the frame budget.

Detail Modes

Rubber Duck Checkpoints

• Before changing this skill's workflow, review the plan for primitive choice, hidden dependencies, safety boundaries, and Copilot artifact order.
• After editing, review SKILL.md, scripts, and references for self-contained behavior, trigger quality, line count, and executable validation.
• Treat the critic as read-only. The producer applies changes and reruns validation.

References