YouTube Watcher
Fetch and read transcripts from YouTube videos. Use when you need to summarize a video, answer questions about its content, or extract information from it.
clawhub install youtube-watcherYouTube 下载器:下载 YouTube 平台上的视频内容。
# Install Skill (downloads SKILL.md to .claude/skills/) clawhub install youtube-download # Then just tell Claude: "use YouTube Downloader to help me..."
# Same install command — works with all SKILL.md-compatible AI coding tools clawhub install youtube-download
This Skill is compatible with the OpenClaw standard. After installation, a SKILL.md file is auto-generated, usable by any OpenClaw-compatible AI Agent (Claude Code, Cursor, Windsurf, etc.).
{baseDir}/download.sh "https://youtube.com/watch?v=VIDEO_ID"
brew install yt-dlpclawhub install youtube-download && clawhub run youtube-download --url "https://www.youtube.com/watch?v=example" --format best --output-dir ./downloadsFetch and read transcripts from YouTube videos. Use when you need to summarize a video, answer questions about its content, or extract information from it.
clawhub install youtube-watcherReal cold-start runs of this skill on Claude Code and Codex.
Input: The site’s #6 homepage scenario ("Download Videos"): download a short, shareable ~6s 360p MP4 of the Creative-Commons film Big Buck Bunny from a YouTube URL — using yt-dlp’s `--download-sections "*0-6"` + format 18 to keep it small.
Single yt-dlp call — format 18 (360p h264+aac) + `--download-sections "*0-6"` + `--force-keyframes-at-cuts` for an accurate cut. Proxy respected automatically; no signature-extraction friction. ffprobe-verified a valid playable MP4.
claude-opus-4-8
Same approach — format 18 + `--download-sections "*0-6"`; 6 shell steps (more granular probing/verification), 58k tokens. Byte-identical output to Claude.
codex-cli 0.140
Output: clip.mp4 — 640×360 / 6.0s / 171 KB (H.264+AAC), the real Big Buck Bunny title sequence. Both agents byte-identical (md5 15be9cc7…).
Verdict: Both downloaded the same valid 360p YouTube clip via yt-dlp — byte-identical (another deterministic-task convergence, finding f3). Efficiency favors Claude (3 vs 6 actions). This completes the site’s six flagship homepage scenarios end-to-end (Add Subtitles · Process Images · Transcribe Audio · Convert & Compress · AI Generation · Download Videos). Content is CC-BY 3.0 (Blender Foundation).
Environment: macOS · yt-dlp 2026.06.09 + ffmpeg · symmetric cold start (codex exec vs a fresh Claude Code subagent). Downloads route through an HTTP proxy (yt-dlp respects it automatically). Deterministic — outputs byte-identical. Test content: "Big Buck Bunny" © Blender Foundation, CC-BY 3.0 (freely redistributable).