roomi-fields/notebooklm-skill/notebooklm

NotebookLM

Overview

NotebookLM answers questions only from the sources uploaded to a notebook, with inline citations to the exact passages used — no open-web knowledge, so answers are hallucination-resistant and fully traceable. This skill drives the @roomi-fields/notebooklm-mcp engine to query notebooks, manage sources, and generate Studio content, and encodes the patterns that make NotebookLM usable at research scale (citation formats, the ~50-queries/day quota, batch-to-cache).

Choosing the transport

Two ways reach the same engine — pick per what the session already has:

1. notebooklm MCP tools — if tools such as notebook_ask / source_add /

server_health (or mcp__notebooklm__*) are available in the session, call them directly. This is the preferred path and needs no server.

1. HTTP REST API — otherwise, use the bundled scripts/nblm.sh, which talks

to a running NotebookLM MCP server (default http://localhost:3000, override with NOTEBOOKLM_SERVER_URL). If no server is reachable, ask the user to start one (npm run start:http from a clone) or to install the MCP.

Both are backed by the same account and session, so the choice is purely about which is already wired up.

Prerequisite: one Google login

NotebookLM needs a signed-in Google session (saved once, reused across runs). Verify with nblm.sh health (or the server_health tool) — look for authenticated: true. If not authenticated, run the interactive login in a terminal (a visible Chrome window opens):

notebooklm-mcp-setup-auth          # global install
# or:  scripts/nblm.sh auth

Run the login in a terminal rather than through an in-client tool: interactive Google login can take minutes and a stdio client's tool-call timeout may cut it off.

Core tasks

Use scripts/nblm.sh for the REST path (or the equivalent MCP tool):

scripts/nblm.sh health                       # reachability + auth status
scripts/nblm.sh notebooks                     # list notebooks (id + name)
scripts/nblm.sh ask "<question>" <notebook_id>   # citation-backed answer (JSON citations)
scripts/nblm.sh search "<query>" <notebook_id> [limit]  # ranked passages, no answer, no quota
scripts/nblm.sh generate <notebook_id> report   # audio|report|video|infographic|presentation|data_table|flashcards|quiz|mind_map

• Search before you ask. search returns the indexed passages matching a

query, ranked best first, and spends no chat quota because nothing is generated. When you intend to reason over the evidence yourself, it is both cheaper and faster than asking. Rank is global across sources and lower is better; start/end are offsets into that source's full text, and are null when the server gave none.

• Ask: the script requests source_format: json, so the answer carries

source names + cited excerpts. For a human-facing answer, prefer expanded (see references/rest-api.md to vary the format).

• Generate: flashcards/quiz route to the study-aid endpoint and

mind_map to the mind-map endpoint automatically.

Working effectively (read before large runs)

For anything beyond a few questions, load references/research-workflows.md. Key points:

• Quota: free accounts cap at ~50 chat queries/day. Rotate accounts

(/re-auth) or, better, ingest once and retrieve offline.

• Batch → cache: for literature reviews / SOTA surveys, run an exhaustive

question set through /batch-to-vault (writes markdown + nblm-answer-v1 JSON sidecars with citations), then answer repeated questions from the cache (e.g. with RTFM) — unlimited, offline.

• Fresh vs. follow-up: omit session_id for independent questions (fastest);

pass a stable one to continue a conversation.

References

• references/rest-api.md — endpoint + body reference for the HTTP path.
• references/research-workflows.md — citation formats, quota strategy, the

batch/ingestion pattern, source discovery.

Installing the engine

If neither the MCP tools nor a server are present, the engine is the npm package `@roomi-fields/notebooklm-mcp` (also a Claude Code plugin via the roomi-fields/claude-plugins marketplace). Point the user there, then run the one-time login above.