NotebookLM integration patterns for external RAG, research synthesis, studio content generation, and knowledge management. Use when creating notebooks, adding sources, generating audio/video, or querying NotebookLM via MCP.
NotebookLM = external RAG engine that offloads reading from your context window. Uses the notebooklm-mcp-cli MCP server (PyPI) to create notebooks, manage sources, generate content, and query with grounded AI responses.
Disclaimer: Uses internal undocumented Google APIs via browser authentication. Sessions last ~20 minutes. API may change without notice.
# 1. Create a notebook for your projectnotebook_create(title="Auth Refactor Research")# 2. Add sources (docs, articles, existing code analysis)source_add(notebook_id="...", type="url", url="https://oauth.net/2.1/")source_add(notebook_id="...", type="text", content="Our current auth uses...")source_add(notebook_id="...", type="file", path="/docs/auth-design.md")# 3. Query with grounded AI responsesnotebook_query(notebook_id="...", query="What are the key differences between OAuth 2.0 and 2.1?")# 4. Generate a deep dive audio overviewstudio_create(notebook_id="...", type="deep_dive")studio_status(notebook_id="...") # Poll until complete# 5. Capture insights as notesnote(notebook_id="...", action="create", content="Key takeaway: PKCE is mandatory in 2.1")
Forgetting auth expiry — Sessions last ~20 min. Always check with nlm login --check before long workflows. Re-auth with nlm login.
One giant notebook — Split by project/topic. One notebook with 50 sources degrades retrieval quality.
Huge single sources — Split documents >50K characters into logical sections for better chunking and retrieval.
Not polling studio_status — Studio content generation takes 2-5 minutes. Poll studio_status instead of assuming instant results.
Ignoring source types — Use type=url for web pages (auto-extracts), type=file for local files. Using type=text for a URL gives you the URL string, not the page content.
Deleting notebooks without checking — notebook_delete is irreversible. List contents with source_list and note(action=list) first.
Skipping research_import — research_start discovers content but does not add it. Use research_import to actually add findings as sources.
Raw queries on empty notebooks — notebook_query returns poor results with no sources. Add sources before querying.
# 1. Authenticate (opens browser for Google OAuth)nlm login# 2. Register MCP server with Claude Codenlm setup add claude-code# 3. Verify auth is activenlm login --check
Manual .mcp.json fallback (if nlm setup is unavailable):
Build dedicated notebooks as curated knowledge bases for debugging, security, and onboarding. Add incident reports, advisories, and runbooks as sources for grounded, verified answers.
Incorrect -- re-investigating a known issue from scratch:
User: "Production is throwing OOM errors again"Claude: "Let me research possible causes..."# Wastes time if this was already diagnosed and documented
Correct -- query the debugging knowledge base:
# 1. Create dedicated KB notebooksnotebook_create(name="Debugging KB")notebook_create(name="Security Handbook")# 2. Add incident reports and advisories as sourcessource_add(notebook_id="debug_kb", content="INC-042: OOM caused by unbounded cache. Fix: add TTL...")source_add(notebook_id="security_kb", content="SEC-007: SQL injection in search endpoint. Fix: parameterize...")# 3. Query for grounded answersnotebook_query(notebook_id="debug_kb", query="What causes OOM errors and how were they fixed?")notebook_query(notebook_id="security_kb", query="Known SQL injection patterns in our codebase")
Key rules:
Create separate notebooks for debugging, security, and onboarding domains
Add incident reports, post-mortems, and security advisories as sources
Query KB notebooks before re-investigating known issues
Keep sources current -- add new incidents as they are resolved
Use for onboarding: new team members query the KB instead of asking around
Use the research API for automated web and Google Drive discovery. The flow is async: start a research task, poll for status, then import discovered sources into your notebook.
Incorrect -- manually searching and adding URLs one by one:
# Tedious and misses relevant contentsource_add(notebook_id="...", url="https://example.com/article1")source_add(notebook_id="...", url="https://example.com/article2")source_add(notebook_id="...", url="https://example.com/article3")# Missed 20 other relevant articles
Correct -- automated research flow:
# 1. Start research (searches web and/or Google Drive)task = research_start( notebook_id="...", topic="Latest developments in WebAssembly component model", sources=["web", "drive"])# 2. Poll for completion (uses Google API quota)status = research_status(task_id=task.id)# status: "searching" | "analyzing" | "completed"# 3. Import discovered sources into notebookresearch_import(task_id=task.id, notebook_id="...")# Adds the most relevant discovered sources automatically
Key rules:
Use research_start for broad topic discovery instead of manual URL hunting
Always poll with research_status -- research takes 1-3 minutes
Research uses Google API quota -- avoid running many parallel research tasks
Import results with research_import to add discovered sources to your notebook
Combine web and Drive sources for comprehensive coverage
Follow up with notebook_query to synthesize the newly imported sources
Add large documents, codebases, and references as notebook sources instead of pasting them into chat. Use notebook_query for targeted synthesis without consuming context window.
Incorrect -- pasting large content directly into chat:
User: "Here's our entire codebase (100K chars)... now explain the auth flow"# Wastes context, may hit token limits, loses nuance in truncation
Correct -- add as source, query for synthesis:
# 1. Add large docs as sources (use RepoMix for codebases)source_add(notebook_id="...", url="https://docs.example.com/api-reference")source_add(notebook_id="...", content=repomix_output)# 2. Query for specific synthesisnotebook_query(notebook_id="...", query="How does the authentication middleware chain work?")# 3. Follow up with targeted questionsnotebook_query(notebook_id="...", query="What error codes does the auth endpoint return?")
Key rules:
Add large documents as sources rather than pasting into chat context
Use RepoMix to bundle codebases into a single source for onboarding
Query the notebook for synthesis -- NotebookLM reads the full source each time
Multiple targeted queries are cheaper than one massive context load
Combine with second-brain pattern to build persistent project knowledge
Create a dedicated notebook per project to capture decisions, design docs, and insights. Query the notebook for grounded answers instead of relying on ephemeral chat context.
Incorrect -- relying on Claude's memory for past decisions:
User: "What did we decide about the auth architecture last week?"Claude: "I don't have context from previous sessions..."
Correct -- add decisions as sources, query later:
# 1. Create a project notebooknotebook_create(name="Project Alpha Decisions")# 2. Add decision documents as sourcessource_add(notebook_id="...", content="ADR-001: Use JWT for auth because...")source_add(notebook_id="...", content="ADR-002: PostgreSQL over MongoDB for...")# 3. Capture new insights with notesnote(notebook_id="...", content="Perf test showed 2x latency with Redis cache miss")# 4. Query for grounded answersnotebook_query(notebook_id="...", query="What auth approach did we choose and why?")
Key rules:
One notebook per project or domain -- avoid mixing unrelated topics
Add decision records, design docs, and meeting notes as sources
Use note tool to capture in-session insights for future retrieval
Use notebook_query for grounded answers backed by actual sources
Periodically prune outdated sources to keep answers relevant
Control notebook access with sharing tools. Always check current sharing status before modifying access -- public links expose all notebook content.
Incorrect -- sharing publicly without reviewing content:
# Dangerous: makes everything in the notebook publicly accessiblenotebook_share_public(notebook_id="...", enabled=true)# Notebook contains internal security advisories -- now exposed
Correct -- check status, review, then share deliberately:
# 1. Check current sharing settingsstatus = notebook_share_status(notebook_id="...")# Shows: public link (on/off), list of collaborators, permission levels# 2. Review notebook content for sensitive materialnotebook_query(notebook_id="...", query="Does this notebook contain credentials or internal secrets?")# 3a. Share with specific collaborators (preferred)notebook_share_invite( notebook_id="...", email="colleague@company.com", role="reader" # or "editor")# 3b. Or enable public link (use with caution)notebook_share_public(notebook_id="...", enabled=true)
Key rules:
Always call notebook_share_status before modifying sharing settings
Prefer notebook_share_invite with specific collaborators over public links
Review notebook content for sensitive material before enabling public access
Use role="reader" by default -- only grant "editor" when collaboration is needed
Disable public links when no longer needed: notebook_share_public(enabled=false)
NotebookLM Studio generates 9 artifact types from notebook sources. All generation is async -- create, poll status, then download.
Incorrect -- calling studio_create and waiting synchronously:
# Blocks for 2-5 minutes with no feedbackresult = studio_create(notebook_id="...", type="audio_overview")# User sees nothing until completion or timeout
Correct -- create, poll, download:
# 1. Create the artifact (returns immediately with artifact ID)artifact = studio_create(notebook_id="...", type="audio_overview")# 2. Poll for completionstatus = studio_status(artifact_id=artifact.id)# status: "pending" | "processing" | "completed" | "failed"# 3. Download when completeddownload_artifact(artifact_id=artifact.id, path="./output/podcast.mp3")
All 9 studio artifact types:
Type
Output
Use case
audio_overview
MP3 podcast
Summarize sources as conversational audio
video_overview
MP4 video
Visual summary with narration
mind_map
SVG/PNG
Visual topic relationships
quiz
JSON
Test comprehension of sources
flashcards
JSON
Study aid from source material
slide_deck
PDF/PPTX
Presentation from sources
infographic
PNG
Visual data summary
data_table
CSV/JSON
Structured data extraction
report
PDF/Markdown
Comprehensive written summary
Key rules:
Always use the poll pattern: studio_create -> studio_status -> download_artifact
Generation takes 2-5 minutes -- inform the user and poll periodically
Check studio_status before attempting download -- downloading a pending artifact fails
Use audio_overview for quick summaries, report for comprehensive analysis
All artifact types require at least one source in the notebook
Create a dedicated NotebookLM notebook for each OrchestKit release to preserve release context, changelog details, and key skill diffs as a queryable knowledge base.
# Dump everything into one shared notebooksource_add(notebook_id="shared", type="text", title="v7 + v6 + v5 notes", content="<all changelogs mixed together>")
Correct:
# One notebook per minor version with focused sourcesnotebook_create(title="OrchestKit v7.0 Release Notes")source_add(notebook_id="v7", type="text", title="CHANGELOG v7.0.0", content="<v7.0.0 changelog section only>")
# Query specific release contextnotebook_query(notebook_id="...", query="What changed in the implement skill for v7.0?")# Compare across releases (query the relevant notebook)notebook_query(notebook_id="...", query="What breaking changes were introduced?")