NotebookLM Coach — a Claude Code skill

Turn any expert's podcast/video library into a personal coach that cites its sources. Bulk-load hundreds of YouTube episodes into NotebookLM from your terminal, ask questions and get answers traced to the exact transcript passage, then turn the advice into a protocol you actually follow.

Give it the Huberman Lab channel and it becomes a health coach. Give it Lenny's Podcast and it's a product coach. Same pipeline, any expert.

Credit: This is an adapted, fixed fork of the notebooklm skill from ArtemXTech/personal-os-skills. All original design credit to that author. See What's different for the changes.

What it does

1. Load sources from the terminal. NotebookLM's UI won't let you add a whole YouTube channel. This does — one command, up to 300 episodes.
2. Cited answers. Every recommendation links back to the exact episode and spoken passage. Verifiable, not vibes.
3. Expert-informed interview. Claude queries the notebook with your goal and builds a personalized, source-backed protocol.
4. Any expert, any domain. Health, product, business, onboarding — same pattern.

Prerequisites

You need three tools: Claude Code, NotebookLM (a free Google account), and optionally Obsidian (for the vault/notes side).

1. Install the two CLIs

# Read side: query notebooks + list sources
uv tool install notebooklm-mcp-cli

# Write side: create notebooks + bulk-load channels (+ its browser)
uv tool install "notebooklm-py[browser]"
uvx --from "notebooklm-py[browser]" playwright install chromium

Using uv tool keeps each CLI in its own isolated environment — no system-Python pollution, and it works even on a Homebrew/PEP-668 Python. (Don't have uv? See astral.sh/uv.)

2. Authenticate (two separate logins)

These are two independent projects with separate sessions — you need both:

nlm login          # read side. Borrows cookies from a Chromium-family
                   # browser you're already signed into. Session ~20 min.
notebooklm login   # write side. Opens its own Chromium for a fresh
                   # Google login. Session lasts weeks.

Verify: nlm notebook list — an empty [] means you're authenticated with no notebooks yet.

3. (Optional) Obsidian

Enable the Dataview community plugin if you want the cited answers, sources, and experiments written into an Obsidian vault.

Install the skill

Clone this repo straight into your Claude Code skills directory:

git clone https://github.com/artemnovitckii/notebooklm-coach.git ~/.claude/skills/notebooklm

Restart Claude Code (or start a new session) and the notebooklm skill is available. Just say "notebooklm", "load channel", or "health protocol".

Quick pipeline (Huberman example)

# 1. Scrape the channel (pure stdlib, no auth needed)
python3 ~/.claude/skills/notebooklm/scripts/load_channel.py scrape \
  --channel "https://www.youtube.com/@hubermanlab" \
  --output /tmp/huberman-videos.json

# 2. Create a notebook (note the printed <notebook-id>)
notebooklm create "Andrew Huberman - Health"

# 3. Load the 300 most recent episodes (run via `uv run --with` so the
#    script can import notebooklm-py from its isolated tool env)
uv run --with "notebooklm-py[browser]" python3 \
  ~/.claude/skills/notebooklm/scripts/load_channel.py load \
  --videos /tmp/huberman-videos.json \
  --notebook <notebook-id> \
  --count 300 --concurrency 6

# 4. Verify how many sources actually landed
nlm source list <notebook-id> --json | grep -c '"id"'

# 5. Ask a cited question
nlm notebook query <notebook-id> \
  "What does Huberman recommend for sustaining deep focus?" --json

From here, let Claude run the interview and build your protocol — that's what the skill orchestrates.

Gotchas (read these — they'll save you an hour)

• Source cap is per tier: free = 50, NotebookLM Plus/Pro = 300 sources per notebook. This is the single biggest trap. If you load 300 on a free account, the first ~50 ingest and the rest show up as red, empty rows — it looks like a bug but it's the tier limit. The loader defaults to --count 300 (the Plus number); on free, use --count 50. For full coverage on free, split a big channel across multiple notebooks and query them together with nlm cross.

• Make sure BOTH logins are the same, correct account. nlm login (read) and notebooklm login (write) authenticate independently and each defaults to whatever Google account your browser is already signed into — so they can silently land on different accounts, or on a free account when your Plus is elsewhere. Verify: nlm login profile list and notebooklm auth check. To switch: nlm login --clear (fresh browser → pick the right account) and notebooklm auth logout && rm -rf ~/.notebooklm/browser_profile && notebooklm login.

• Reds = failed adds; you can spot and repair them. A failed source keeps the raw URL as its title (NotebookLM never fetched the video). Higher --concurrency causes more of them (an ingest race), even on Plus — at --concurrency 1 they essentially disappear. To repair: delete the red rows and re-add at low concurrency.

  copy

  NB=<notebook-id>
  # delete every red row (title is a URL)
  nlm source list $NB --json | python3 -c "import json,sys;[print(s['id']) for s in json.load(sys.stdin) if s['title'].strip().startswith('http')]" \
    | while read id; do nlm source delete "$id" --confirm; done
  # then re-add the missing videos at --concurrency 1
  

  Verify the real good/red split anytime:

  copy

  nlm source list $NB --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)}')"
  

• nlm sessions are short (~20 min). If queries start failing with auth errors, just re-run nlm login. The notebooklm (write) session lasts weeks.

• Processing lag. After upload, NotebookLM indexes each transcript server-side (a few minutes). Early queries may be thin until processing finishes.

What's different in this fork

Fixes that make it work as of mid-2026:

• YouTube scraper fixed. YouTube migrated channel listings from videoRenderer → lockupViewModel; the original parser silently returned 0 videos. Added a lockupViewModel parser (scripts/load_channel.py).
• 300-source cap enforced in the loader, with a clear message when a channel exceeds it.
• Lower default concurrency (20 → 6) so success/failure reporting is honest, plus a documented warning about the false-failure behavior.
• Corrected auth docs: the current nlm CLI uses nlm login, not nlm auth login.
• Install via uv tool (isolated, PEP-668-safe) and run the loader via uv run --with instead of assuming a ~/projects/notebooklm-loader virtualenv.
• Documented the real-world traps that the original glossed over: the per-tier source cap (free 50 / Plus 300), the two-logins-must-match-account pitfall, and how to detect & repair red (failed) sources. See Gotchas.

License

MIT — see LICENSE. Adapted from ArtemXTech/personal-os-skills.