| name | deep-research |
| description | Multi-source deep research combining YouTube, web search, and article URLs into NotebookLM as a free RAG backend. Use when user wants to research a topic in depth, find up-to-date information, generate analysis, infographics, podcasts, quizzes, or any NotebookLM artifact. Triggers on "research X", "deep dive into X", "find everything about X", or /deep-research. |
| user_invocable | true |
| arguments | topic to research |
Deep Research Skill
Research any topic using multiple sources (YouTube, web articles, web research) fed into NotebookLM as a free RAG analysis engine. Zero token cost for the research itself — only tokens used are for sending/receiving commands.
Why This Works
- YouTube = expert video content, tutorials, talks, interviews
- Web research = NotebookLM's built-in web search finds and indexes articles automatically
- WebSearch = Claude's web search finds specific articles, blog posts, papers to add as URL sources
- NotebookLM = free Google-powered RAG that indexes all sources and generates artifacts
- Claude Code = orchestrator that runs all source-gathering in parallel
- Result: comprehensive, multi-perspective research without burning Claude tokens on analysis
Workflow
User query
├── YouTube search (yt-search) ──────────────┐
├── Web research (notebooklm add-research) ──┼── NotebookLM notebook ── Analyze/Generate
└── Article search (WebSearch + add URLs) ───┘
Step-by-Step Execution
Step 1: Create NotebookLM Notebook
Create the notebook first so all source-adding steps can target it.
notebooklm create "<TOPIC> Research" --json
notebooklm use <notebook_id>
Step 2: Gather Sources (run in parallel)
Launch three parallel tracks to maximize source diversity:
Track A: YouTube Videos
Use /yt-search to find relevant videos:
PYTHONIOENCODING=utf-8 python -m yt_dlp "ytsearch10:<TOPIC>" --flat-playlist -j 2>/dev/null
Track B: NotebookLM Web Research (automatic)
Let NotebookLM search the web and import sources itself:
notebooklm source add-research "<TOPIC>" --notebook <notebook_id>
notebooklm source add-research "<TOPIC>" --mode deep --no-wait --notebook <notebook_id>
- Fast mode is good for focused topics
- Deep mode is better for broad or complex topics
- Deep mode runs async — use
notebooklm research wait --notebook <notebook_id> --import-all to wait
Track C: Targeted Article URLs (optional but powerful)
Use Claude's WebSearch to find high-quality articles, then add them:
notebooklm source add "https://example.com/article" --notebook <notebook_id>
notebooklm source add "https://blog.example.com/post" --notebook <notebook_id>
Good for:
- Industry reports and whitepapers
- Blog posts from known experts
- Documentation and official guides
- Academic papers and case studies
Step 3: Wait for Sources
Ensure all sources are indexed before analysis:
notebooklm source list --notebook <notebook_id> --json
notebooklm source wait <source_id> -n <notebook_id>
For deep research, also wait for research completion:
notebooklm research wait -n <notebook_id> --import-all --timeout 300
Step 4: Analyze with NotebookLM
notebooklm ask "Analyze the key strategies, frameworks, and trends discussed across all sources. Provide a comprehensive summary." --notebook <notebook_id>
notebooklm ask "<specific question>" --notebook <notebook_id>
Step 5: Generate Artifacts (Optional)
| Artifact | Command | Output |
|---|
| Audio podcast | notebooklm generate audio | MP3 |
| Video overview | notebooklm generate video | MP4 |
| Infographic | notebooklm generate infographic | PNG |
| Slide deck | notebooklm generate slide-deck | PDF/PPTX |
| Quiz | notebooklm generate quiz | JSON/MD |
| Flashcards | notebooklm generate flashcards | JSON/MD |
| Report | notebooklm generate report --format briefing-doc | Markdown |
| Mind map | notebooklm generate mind-map | JSON |
| Data table | notebooklm generate data-table "description" | CSV |
Download artifacts:
notebooklm download audio output.mp3
notebooklm download infographic output.png
Step 6: Return Results to User
- Present the analysis in a clean, structured format
- List source breakdown: X YouTube videos, Y web research sources, Z articles
- Offer to generate additional artifacts if useful
Parallel Execution Strategy
For maximum efficiency, use subagents:
Main agent: Create notebook
├── Subagent 1: yt-search → add YouTube videos → wait for indexing
├── Subagent 2: add-research (deep) → wait → import-all
└── Subagent 3: WebSearch → add article URLs → wait for indexing
Main agent: Once all ready → analyze + generate artifacts
Use the --notebook <id> flag on all commands so subagents don't conflict on context.
Research Depth Levels
| Level | YouTube | Web Research | Articles | Best For |
|---|
| Quick | 5 videos | fast mode | skip | Quick overview, familiar topic |
| Standard | 10 videos | fast mode | 3-5 URLs | Most research tasks |
| Deep | 15-20 videos | deep mode | 5-10 URLs | Comprehensive analysis, new domain |
Default to Standard. Use Deep when user says "thorough", "comprehensive", or "deep dive".
Tips
- Diversity wins — mixing YouTube + web articles + auto-research gives the richest analysis
- Filter YouTube by views — higher view count usually means better quality
- Use WebSearch strategically — search for "[topic] guide 2026" or "[topic] case study" to find written content YouTube doesn't cover
- Deep research is async — start it first, add YouTube/articles while it runs
- Source limits — Standard plan: 50 sources/notebook. Don't overshoot.
- Save results —
notebooklm ask "..." --save-as-note preserves Q&A in the notebook
Example Prompt
/deep-research meta ad strategies 2026
This will:
- Create a NotebookLM notebook
- Search YouTube for top 10 videos on meta ad strategies
- Run deep web research via NotebookLM
- Find 5 top articles/guides via WebSearch
- Wait for all sources to index
- Generate comprehensive analysis
- Optionally create infographic/slides/podcast
Credits
Workflow inspired by Erik Lazar. NotebookLM Python API by teng-lin/notebooklm-py.