Notebooklm

NotebookLM - Expert Knowledge to Action

Turn any expert's content into a personalized protocol with experiments you actually run. Load 300 YouTube episodes into NotebookLM from terminal, run a cited interview about your goal, create experiments in your Obsidian daily note.

Video walkthrough: https://youtu.be/KRpZSvtMiTI

What This Does

1. Load sources from terminal. You can't just tell NotebookLM to add a YouTube channel. This skill does it. One command. 300 episodes.
2. Cited answers traced to exact transcript lines. Every recommendation links back to the exact episode and passage. Verifiable.
3. Expert-informed interviews. Claude queries NotebookLM with YOUR goal. Generates questions informed by the expert's research on your specific topic.
4. Experiments in Obsidian. Protocol becomes experiments in your daily note. Morning routine skill asks every day: how is this going?
5. Any expert, any domain. Huberman for health. Lenny for product. Onboarding docs for a new job. Same pattern.

Step 0: Preflight (run this first, every time)

Before any workflow below, check the tools are installed and logged in:

bash ~/.claude/skills/notebooklm/scripts/bootstrap.sh --check

Exit 0 = ready. Exit 2 = something missing; the output names exactly what. To install whatever is missing, re-run without --check:

bash ~/.claude/skills/notebooklm/scripts/bootstrap.sh

The two login commands are the only manual steps — they open a browser for Google sign-in, so never run them from a tool call. Print them for the user instead.

Prerequisites

Handled automatically by scripts/bootstrap.sh above. This section documents what it installs, for anyone who prefers to do it by hand.

1. Install nlm CLI

uv tool install notebooklm-mcp-cli

Gives you the nlm command. See notebooklm-mcp-cli for details.

2. Install notebooklm-py (for notebook creation and channel loading)

uv tool install "notebooklm-py[browser]"
uvx --from "notebooklm-py[browser]" playwright install chromium

This installs the notebooklm command in its own isolated environment (no system-Python pollution, works even with a Homebrew/PEP-668 Python). The matching Chromium build is installed for the browser automation.

If you prefer a plain pip install: pip install "notebooklm-py[browser]" && playwright install chromium. On a Homebrew-managed Python you may need --break-system-packages.

3. Authenticate

You need two logins — the tools are independent projects with separate sessions:

# nlm: read side (queries + source listing). Borrows cookies from a
# Chromium-family browser you're already signed into. Session ~20 min.
nlm login

# notebooklm-py: write side (notebook creation + channel loading).
# Opens its own Chromium for a fresh Google login. Session lasts weeks.
notebooklm login

notebooklm-py saves cookies to ~/.notebooklm/profiles/default/storage_state.json. Check nlm with nlm notebook list (empty [] = authenticated, just no notebooks yet).

4. Obsidian Plugins

• Dataview (required) - for dashboard queries and citation tables

Quick Start

# List your notebooks
nlm notebook list

# Ask a question with citations
nlm notebook query <notebook-id> "What does Huberman say about deep focus?" --json

# List sources
nlm source list <notebook-id> --json

Workflow Routing

The Full Pipeline

This is the workflow shown in the video:

1. Pick your expert and goal

Goal: "I want to improve my health and focus"
Expert: Andrew Huberman (@hubermanlab on YouTube)

2. Load their content

# Scrape channel videos (no auth/deps needed — pure stdlib)
python3 scripts/load_channel.py scrape \
  --channel "https://www.youtube.com/@hubermanlab" \
  --output /tmp/huberman-videos.json

# Create notebook (grab the printed <notebook-id>)
notebooklm create "Andrew Huberman - Health"

# Load the 300 most recent episodes.
# Run via `uv run --with` so the script can import notebooklm-py from its
# isolated tool env. Videos are newest-first, so this keeps the latest 300.
uv run --with "notebooklm-py[browser]" python3 scripts/load_channel.py load \
  --videos /tmp/huberman-videos.json \
  --notebook <notebook-id> \
  --count 300 \
  --concurrency 6

Source cap depends on tier: free = 50, NotebookLM Plus/Pro = 300 per notebook. --count 300 assumes Plus/Pro. On a free account only ~50 ingest and the rest become red, empty rows — use --count 50 there. Confirm both nlm and notebooklm are logged into the same account with that tier (notebooklm auth check).

Keep --concurrency low. Higher concurrency causes some adds to fail (red rows whose title is the raw URL), even on Plus. At --concurrency 1 they essentially disappear. After loading, check the real split:

nlm source list <notebook-id> --json | python3 -c "import json,sys;d=json.load(sys.stdin);r=[s for s in d if s['title'].strip().startswith('http')];print(f'good {len(d)-len(r)}  red {len(r)}')"

To repair reds: delete the red rows (nlm source delete <id> --confirm) and re-add those videos at --concurrency 1.

3. Ask expert-informed questions

nlm notebook query <notebook-id> \
  "What does Huberman recommend for sustaining deep focus for 4+ hours daily?" --json

Each answer comes with [N] citations back to the exact source and passage.

4. Run a cited interview

Claude uses the notebook to generate interview questions specific to YOUR goal. You answer honestly. Claude builds a personalized protocol where each recommendation is tied to an exact episode.

5. Create experiments

The protocol becomes experiments in your Obsidian vault:

• Each experiment has a hypothesis, protocol, success criteria, and timeframe
• They appear in your daily note every morning
• Your morning routine skill asks: "How is this experiment going? Any observations?"

6. Turn it into a reusable skill

Package the workflow as a /huberman or /lenny skill. Same pattern, different expert.

Vault Structure

Your Vault/
├── Notes/NotebookLM/
│   ├── Huberman Health.md              # type: notebook (index)
│   └── huberman-health/
│       ├── Sources/                     # type: notebook-source (transcripts)
│       │   └── Episode Title.md
│       └── QA/                          # type: nlm-query (cited answers)
│           └── 2026-04-05 Focus Protocol.md
├── Notes/Experiments/
│   └── Morning Sunlight Protocol.md     # type: experiment
└── Notes/Dashboards/
    └── Health.md                        # Dashboard with embedded experiments

Scripts

All scripts use Path.cwd() as vault root. Run them from your vault directory.

Citation Resolution

The resolver turns [N] markers in NotebookLM answers into clickable [[Source#^c-XXXXXXXX|[N]]] wikilinks. Click to jump to the exact cited passage in the source transcript.

• Anchor IDs are stable (MD5 of cited text)
• Idempotent: re-running same question skips existing anchors
• Cross-source citation remap: handles collapsed source_ids
• ~96% resolution rate across tested queries

Examples

• Health: 300 Huberman episodes -> personalized health protocol with sleep, supplements, exercise experiments
• Product: 200 Lenny's Podcast episodes -> product strategy playbook with cited frameworks
• New job: Onboarding docs + team wikis + architecture decisions -> ramp-up plan with daily experiments
• Business: Hormozi content -> offer audit with value equation scoring

License

MIT