--- name: search-youtube description: > This skill should be used when the user asks to "search YouTube", "find videos about", "get a transcript", "download subtitles", "extract audio from YouTube", "scan a channel", "research a topic on YouTube", "get video metadata", "what videos exist about", "download YouTube audio", "YouTube research", "summarize this video", "what is this video about", "pull captions from", "grab the audio from", or provides a YouTube/Vimeo/video URL and wants to extract information from it. Also triggers on "batch download transcripts", "analyze a channel", or any multi-video research workflow. argument-hint: "" allowed-tools: - Read - Grep - Glob - Bash(python3:*) - Bash(yt-dlp:*) - Bash(jq:*) - AskUserQuestion - WebSearch - Task --- # YouTube Research Multi-platform video research toolkit. Operates in two modes: toolkit (individual operations) and research (autonomous search-to-synthesis pipeline). All operations use a single CLI at `${CLAUDE_PLUGIN_ROOT}/skills/search-youtube/scripts/yt_research.py`. Run `python3 ${CLAUDE_PLUGIN_ROOT}/skills/search-youtube/scripts/yt_research.py --help` for full flag reference on any subcommand. ## Toolkit Mode Invoke individual subcommands for targeted operations. Default mode when the user requests a specific action (transcript, search, metadata, audio, channel scan). ### Search Find videos matching a query. Returns structured results with metadata. ```bash python3 ${CLAUDE_PLUGIN_ROOT}/skills/search-youtube/scripts/yt_research.py search "" --count 10 ``` Add filters to narrow results: `--min-duration 600` (seconds), `--after 20250101` (YYYYMMDD), `--min-views 50000`. Filters are applied client-side after fetching, so the tool over-fetches automatically to compensate. Output is JSON by default; add `-f text` for human-readable. ### Transcript Download and clean subtitles to LLM-ready text. ```bash python3 ${CLAUDE_PLUGIN_ROOT}/skills/search-youtube/scripts/yt_research.py transcript "" ``` Outputs clean prose to stdout by default. Add `--timestamps` for SRT with timing cues. Add `--save -t ` to persist to `~/youtube-research//`. Use `--lang all` to list available subtitle languages before downloading. Fallback chain: manual subs then auto-generated. Exit 4 if no subtitles exist in the requested language. ### Metadata Extract full video information without downloading. ```bash python3 ${CLAUDE_PLUGIN_ROOT}/skills/search-youtube/scripts/yt_research.py metadata "" ``` Returns: title, description, channel, duration, chapters, view/like counts, tags, available subtitle languages, thumbnail URL. Add `--playlist` for playlist entry listings. ### Audio Download audio in the requested format. ```bash python3 ${CLAUDE_PLUGIN_ROOT}/skills/search-youtube/scripts/yt_research.py audio "" --audio-format mp3 -t ``` Saves to `~/youtube-research//audio/`. Supported formats: mp3, m4a, opus, wav. Always saves to disk (audio cannot go to stdout). Prints the file path on success. ### Channel Scan a channel's content. ```bash python3 ${CLAUDE_PLUGIN_ROOT}/skills/search-youtube/scripts/yt_research.py channel "" --limit 20 ``` Supports tabs: `--tab videos` (default), `shorts`, `streams`, `playlists`. Filter with `--after`/`--before` (YYYYMMDD). Sort with `--sort views` for most-viewed-first. ### Batch Processing Any subcommand except `search` accepts `--batch ` (or `--batch -` for stdin) to process multiple URLs. One URL per line; lines starting with `#` or `;` are skipped. ```bash python3 ${CLAUDE_PLUGIN_ROOT}/skills/search-youtube/scripts/yt_research.py transcript --batch urls.txt --save -t ``` ## Research Mode Activate when the user asks to "research", "investigate", or "find out about" a topic using YouTube as a source. This is an adaptive multi-round discovery pipeline designed for niche and emerging topics where popular videos often under-serve. ### Round 1: Divergent Search Generate 4-6 query variants that cover different angles of the topic: - Exact tool/concept name (e.g., `"openclaw"`) - Tool + ecosystem context (e.g., `"openclaw claude code"`) - Problem the tool solves (e.g., `"claude code documentation framework"`) - Workflow/demo framing (e.g., `"openclaw walkthrough demo"`) - Alternative names, abbreviations, or common misspellings if applicable Spawn one `Task` agent per query variant simultaneously, each running: ```bash python3 ${CLAUDE_PLUGIN_ROOT}/skills/search-youtube/scripts/yt_research.py search "" --count 15 ``` Collect all results and deduplicate by video ID. Aim for 30-50 unique candidates across all threads. ### Round 1 Evaluation: Niche-First Heuristics For niche, emerging, or edge-of-tech topics, these signals predict quality: **Positive signals (use these):** - Title specifically names the tool or concept (not "top AI tools 2025") - Small channel (< 50K subscribers) — for new tech, practitioners publish before educators discover the topic - Technical, specific description (mentions code, config, architecture, or commands) - Structured content: chapters, timestamps, or detailed description - Recent upload date **Negative signals (treat as red flags on niche topics):** - View count > 100K — on a narrow topic this usually means beginner-level or clickbait - "Tutorial for beginners" / "complete guide" in the title for brand-new tools - Large generalist channel covering many unrelated topics Select 6-10 candidates from the combined pool. Note which channels produced the strongest results — those are targets for Round 2. ### Round 2: Channel Discovery and Refinement For each channel that surfaced a strong Round 1 result, scan its recent videos: ```bash python3 ${CLAUDE_PLUGIN_ROOT}/skills/search-youtube/scripts/yt_research.py channel "" --limit 20 ``` Also run 1-2 refined search queries using specific terminology that appeared in strong Round 1 titles or descriptions. Add any new candidates to the pool. **Quality gate:** If Round 1 candidates are weak (generic titles, all high-view generalist content, nothing specifically about the topic), surface this to the user and run another search round with reformulated queries before proceeding to transcripts. ### Round 3: Confirm and Transcribe From the enriched candidate pool, select 4-7 videos using the Round 1 criteria. Extract metadata to confirm relevance before committing to downloads: ```bash python3 ${CLAUDE_PLUGIN_ROOT}/skills/search-youtube/scripts/yt_research.py metadata "" ``` Download transcripts in parallel — always spawn one `Task` agent per video, even for 2: ```bash python3 ${CLAUDE_PLUGIN_ROOT}/skills/search-youtube/scripts/yt_research.py transcript "" --save -t ``` Read each transcript and synthesize findings into a report. Use `WebSearch` or brave-cli to cross-reference claims or fill gaps when YouTube sources disagree or leave questions unanswered. ### Research Report Format Present the synthesis as a structured markdown report: - Title and one-paragraph summary of the research question - Key findings (3-7 bullet points of the most important takeaways across all sources) - Points of agreement between sources (what multiple videos confirm) - Points of disagreement (where sources contradict, with attribution) - Unique insights (notable points from individual videos not repeated elsewhere) - Gaps in coverage (what the sources collectively missed) - Sources table: video title, channel, duration, and URL for each video used Attribute specific claims to their source video. Include timestamps when the transcript preserves them. Save all transcripts under the topic directory for future reference. See `examples/research-report.md` for a sample report structure. ### Research Composability See `references/cli-reference.md` for pipeline patterns that chain subcommands with standard Unix tools (search → jq → batch transcript). ## Error Recovery | Exit Code | Meaning | Recovery Action | |-----------|---------|-----------------| | 0 | Success | — | | 1 | Usage error | Check `--help` for correct syntax | | 2 | yt-dlp not found | Tell user to install: `pip install yt-dlp` | | 3 | Network/download error | Check URL validity; try `--cookies ` for private/restricted content | | 4 | No results | For transcripts: try `--lang all` to list available languages. For search: broaden query or remove filters | ## Platform Notes Load `references/platforms.md` when processing a non-YouTube URL or when a yt-dlp command fails with exit code 3 on an unfamiliar platform. YouTube is the primary platform, but any yt-dlp-supported URL works (Vimeo, Twitter, Twitch, etc.). After extracting a transcript, read the output and summarize key points for the user unless they asked for raw output only.