Notebooklm

---
name: notebooklm
description: >-
  Drive NotebookLM over a programmatic CLI with no browser: turn a vault slice into an audio
  overview, study guide or mind map, and fold the artifacts back into the vault. Browser
  automation is the fallback if CLI auth breaks. Triggers: "/notebooklm", "make an audio digest
  of <topic>", "podcast from the vault".
license: MIT
---

# /notebooklm — audio/artifact layer over the Second Brain

> 🧒 When reporting to Anton end with a child-simple "In plain words" recap. (memory `eli5-always`)

NotebookLM = the **artifact/explanation layer** (Audio Overview, study guide, mind map, reports) over the vault — NOT memory or RAG (that stays Obsidian + e5/rerank + SQLite). Flow: `vault → Claude Code curates a NARROW packet → NotebookLM → artifacts back into vault as Derived`. Full state = memory [[notebooklm-integration]].

## Rail = CLI (verified 2026-07-04, ADOPT — H16)
Programmatic CLI `teng-lin/notebooklm-py` (17k⭐, MIT). No Chrome window, no DOM, no truncation. Works on Anton's consumer `a2@` PRO (no Enterprise API).
- **CLI binary:** `%USERPROFILE%\notebooklm-cli\venv\Scripts\notebooklm.exe` (venv, v0.7.3). Set `NB="%USERPROFILE%\notebooklm-cli\venv\Scripts\notebooklm.exe"`.
- **Auth:** browser-cookies from Firefox (Chrome/Edge blocked by App-Bound Encryption + Chrome 146 DBSC, TPM-bound — Firefox does neither, see `02-Decisions\decision-2026-07-16-browser-automation-layer.md`). Stored `%USERPROFILE%\.notebooklm\profiles\default\storage_state.json`, account `owner.work@example.com`.
- **Cookies expire ~every few weeks** → on any `auth`/`401` failure, RE-LOGIN: Anton opens NotebookLM in **Firefox** (a2@), then `"$NB" login --browser-cookies firefox --account owner.work@example.com`. Verify with `"$NB" auth check`.
- **⚠️ Fallback if `login --browser-cookies firefox` picks the WRONG profile or fails to find the session** (the CLI's cookie-reader can hit the same class of bug as generic browser_cookie3-style libs: it may grab a stale `[ProfileN] Default=1` profile instead of the one Firefox actually launches — fixed once already in our own helper, see [[deterministic-script-gotchas]]): build `storage_state.json` directly with the fleet-shared, /tt-tested helper instead of fighting the CLI's own extractor:
  ```bash
  python "$env:USERPROFILE\.claude\scripts\_shared\firefox_cookies.py" --domain google.com --export "%USERPROFILE%\.notebooklm\profiles\default\storage_state.json"
  ```
  Then `"$NB" auth check` to confirm. Canon: `02-Decisions\decision-2026-07-16-browser-automation-layer.md`, DR26-07-16-HUB-01. (Note: fb-post/x-post/fb-reply do NOT use this path — they drive Anton's live logged-in Chrome tab via Claude-in-Chrome MCP, a deliberate anti-ban choice; `firefox_cookies.py`/`firefox_login.py` are for API-style clients like this one, not for social posting.)
- **Anti-abuse:** run from the hub (residential IP), **no VPN**.

## Access reality check (verified 2026-10-06)
- **There is no public consumer API.** You cannot pick up a NotebookLM API key from a settings page and script your own notebooks.
- **The official API is Enterprise-only** — now branded *Gemini Notebook Enterprise*, shipped inside Gemini Enterprise on Google Cloud. It does expose REST endpoints for notebooks, sources and audio overviews, with regional endpoints, VPC Service Controls and CMEK, but it is sold to organisations, not to individuals. The older standalone Podcast API is deprecated and is not allowlisting new customers.
- **There is no official MCP server** as of October 2026. Community ones exist (for example `julianoczkowski/notebooklm-mcp-2026`), and they are worth knowing about if you want *other* agents to reach NotebookLM — but every one we looked at wraps the same unofficial rail this skill already uses: cookies plus internal endpoints, or a driven browser. One more layer, same terms-of-service surface. Check the specific server before you trust it; we did not audit all of them.
- So on a consumer account the CLI below is the rail. Budget for the cookie expiry, not for an API key.
- Check these claims yourself before you build on them, because Google moves this product often: the Enterprise
  surface and its deprecations are documented in the Gemini Enterprise release notes
  (`docs.cloud.google.com/gemini/enterprise/docs/release-notes`). Dates above are when we last read them.

## Three generation defaults (these move output quality more than the prompt does)
1. **Always maximum length** — `--length long`. Go shorter only when explicitly asked, and say why in the report.
2. **Two languages, one after the other** — run `--language ru`, download it, then re-run `--language en` on the SAME notebook. As of October 2026 parallel generations on one notebook break the CLI poller, so this is sequential by necessity, not by taste. It is an unofficial client, so re-test rather than assume.
3. **Put clarifying questions in the prompt itself.** Two blocks earn their keep: "ask each other a smart beginner's questions and answer them immediately", and "ground every claim in the sources and name where the fact came from". Without them the hosts stay pleasant and shallow.
4. **Feed 10-15+ sources.** Your own curated packet counts as ONE. Give it a single source and NotebookLM just paraphrases you back at you in a nicer voice. Verify each URL returns 200 with live code before `source add`; top up the pool with `source add-research "<query>" --import-all --cited-only --timeout 1800`.

## digest (default) — vault → audio
1. **RECALL + curate a NARROW packet** ([[vault-data-architecture]]): `brain_ask.py`/`/ask` + grep to pull ONLY the relevant slice. Write it to a clean `.txt`/`.md` packet.
2. **Create + add source + generate (CLI):**
   ```bash
   NB="%USERPROFILE%\\notebooklm-cli\\venv\\Scripts\\notebooklm.exe"
   "$NB" create "<Title>" --use                       # --use sets current context (don't parse --json 'id' — key differs)
   "$NB" source add "<packet.md>" --type file          # uses current context; or --type text "<inline>", or a URL. Wait for Status=ready.
   "$NB" generate audio "<style hint>" --length short --wait   # uses current; or generate mind-map / report / flashcards…
   ```
   Other artifacts: `generate mind-map | report | flashcards | quiz | slide-deck | video | data-table | infographic`.
3. **Pull artifacts back:** `"$NB" download audio "05-Resources\NotebookLM\artifacts\audio\<name>.mp3"` (current context; or `--all <dir>`, `-n <id>`). Non-audio: `artifact export`. Get the notebook id for the index: `NBID=$("$NB" list --json | python -c "import sys,json;print(json.load(sys.stdin)[0]['id'])")` (newest first).
4. **Index + link:** add the notebook/artifact to `05-Resources\NotebookLM\_NotebookLM-Index.md`; link the derived note back to its source concept(s). Backup before vault write ([[vault-backup-rule]]).

## pull — Phase B
Pull existing artifacts from the top notebooks into the vault as Derived notes: `"$NB" list` → `download ... --all` per notebook → index + link.

## Fallback — Chrome-autonomy (only if CLI breaks)
If Google rotates endpoints and the CLI errors (and a re-login doesn't fix it), drive the logged-in UI via Claude-in-Chrome ([[chrome-autonomy-self-drive]]) on `a2@`: Create notebook → Add sources → Studio → generate → **Download** menu (NOT DOM-scrape, truncates ~1000 chars). This was the old primary rail; kept as backstop.

## Safety
- CLI uses undocumented Google APIs (community project) — same ToS surface as driving the logged-in UI of the same account; low-med risk for personal volumes.
- Audio/video generation has a **daily quota** — don't burn it on test runs; add `--retry` + delays for batches.
- Stay on a2 PRO (free, included — [[prefer-included-limits-before-paid-api]]).

## Output
Notebook created/used · artifact type + where saved · index updated. Then 🧒 recap. Routine candidate: weekly "audio digest of my brain" — see /skill-gap and [[evaluate-recurring-into-routine]]. Decision: `02-Decisions\decision-notebooklm-claude-code-artifact-layer`.

<!--kit-footer-->

---

**Like this skill?** It is one of 100 in [second-brain-starter-kit](https://github.com/tonydzi/second-brain-starter-kit): the second brain we built for ourselves and run every day at Palo Alto AI Research Lab. Install the whole set with `npx skills add tonydzi/second-brain-starter-kit`. Everything is open source and free, so take what you need.

Flagships worth a look on their own: [secondop-panel](https://github.com/tonydzi/secondop-panel) (a second opinion from a panel of external models), [claude-memory-tidy](https://github.com/tonydzi/claude-memory-tidy) (stop your agent's memory from rotting), [telegram-mcp-kit](https://github.com/tonydzi/telegram-mcp-kit) (your own Telegram over MCP in about 15 minutes).

Author: **Anton Dziatkovskii**, Palo Alto AI Research Lab. Telegram [@tonydzi](https://t.me/tonydzi) - WhatsApp [+1 341 222 9178](https://wa.me/13412229178) - X [@Tony_Stef_](https://x.com/Tony_Stef_)

**Engineers: want to test-drive this setup?** Message me. I hand out free starter seeds to engineers who test and report back, and custom skill requests are welcome.