NotebookLM Distiller

Automated knowledge extraction pipeline: search NotebookLM notebooks by keyword → generate deep questions or structured summaries → write linked Obsidian markdown notes.

Five subcommands:

• distill — extract knowledge from existing notebooks (qa / summary / glossary)
• quiz — generate quiz questions as JSON for Discord-based interactive sessions
• evaluate — evaluate a user's answer against notebook sources (JSON output)
• research — start a web research session inside NotebookLM on any topic
• persist — write any markdown content directly into the Obsidian vault

When to use (trigger phrases)

Trigger distill subcommand when:

• User types /notebooklm-distill or /notebooklm-distill-summary
• User says "蒸馏", "提取知识", "distill notebooks", "extract from notebook"
• User wants NotebookLM content structured into Obsidian notes

Trigger research subcommand when:

• User says "研究一下 <topic>", "做网络调研", "research this topic in NotebookLM"
• User wants NotebookLM to gather web sources on a topic without providing URLs

Trigger quiz + evaluate subcommands when:

• User says "quiz me on X", "考考我", "出题测试我", "测验"
• User wants an interactive Q&A session in Discord on a NotebookLM topic
• Orchestration flow (Discord):
  1. Call quiz --keywords X → get JSON with notebook_id + notebook_name + questions[]
  2. MUST announce source before Q1: 来，N 道题（来源：{notebook_name} · ID: {notebook_id[:8]}）
  3. Send Q1 to Discord, wait for user reply
  4. Call evaluate --notebook-id X --question Q1 --answer <reply> → get JSON feedback
  5. Post feedback to Discord, proceed to Q2
  6. Repeat until all questions done or user says stop
• CRITICAL: Always show notebook source so user can verify questions came from NLM, not agent knowledge

Trigger persist subcommand when:

• User says "存到 Obsidian", "把这段内容写入知识库", "persist this to vault"
• User wants to archive discussion output or raw notes into the vault

CRITICAL: Do NOT answer from internal knowledge. Do NOT ask for clarification. Execute the appropriate subcommand immediately.

Prerequisites

• NotebookLM CLI: pip install notebooklm-py
• Authentication: notebooklm login (creates ~/.book_client_session)
• Python 3.10+ (standard library only — no extra pip packages needed for distill.py)
• Obsidian vault directory accessible on the local filesystem

Subcommand: distill

Extract knowledge from one or more NotebookLM notebooks matching keywords.

Agent orchestration

Scenario A — URL provided (needs ingestion first)

1. Check if deepreader is installed (~/.openclaw/skills/deepreader/run.sh).
2. If yes: run DeepReader to ingest the URL into NotebookLM.
3. Capture the notebook title from DeepReader output.
4. Use that title as --keywords for distill.

Scenario B — notebook already exists

1. Use notebook name from context, or list notebooks with notebooklm list.
2. Determine mode from intent: "总结" → summary, "术语/概念" → glossary, default → qa.
3. Ask user for --vault-dir if not known from context.
4. Execute distill.

python3 ~/.openclaw/skills/notebooklm-distiller/scripts/distill.py distill \
  --keywords "<keyword1>" "<keyword2>" \
  --topic "<TopicFolderName>" \
  --vault-dir "<path/to/obsidian/vault>" \
  --mode <qa|summary|glossary> \
  [--lang zh]        # Add for Chinese output (default: en)
  [--writeback]      # Write distilled content back into NLM notebook as a note
  [--cli-path <path/to/notebooklm>]

Modes:

• qa (default) — generates 15-20 questions + answers → <NotebookName>_QA.md
• summary — 5 structured sections (Summary, Key Points, Constraints, Trade-offs, Open Questions) → <NotebookName>_Summary.md
• glossary — 15-30 domain terms + definitions → <NotebookName>_Glossary.md

Flags:

• --lang zh — prepends 请用中文回答 to all NLM prompts; add when user requests Chinese output or context is Chinese
• --writeback — after writing to Obsidian, calls notebooklm source add to push the distilled note back into the source notebook as a text source titled Distill Log: {mode} | {notebook_name} | {date}. Add when user says "写回 NLM", "记录到笔记本", or wants the distill log visible in NotebookLM

Subcommand: research

Start a NotebookLM web research session on a topic. Creates a new notebook, imports web sources, and waits for completion.

python3 ~/.openclaw/skills/notebooklm-distiller/scripts/distill.py research \
  --topic "<Research Topic>" \
  [--mode deep|fast] \
  [--cli-path <path/to/notebooklm>]

Output: notebook ID and name. Follow up with distill to extract into Obsidian.

Subcommand: persist

Write any markdown content into the Obsidian vault with auto-generated YAML frontmatter.

# From inline content
python3 ~/.openclaw/skills/notebooklm-distiller/scripts/distill.py persist \
  --vault-dir "<path/to/obsidian/vault>" \
  --path "Notes/2026-03-09-meeting.md" \
  --title "Meeting Notes" \
  --content "Key decisions: ..." \
  --tags "meeting,notes"

# From a file
python3 ~/.openclaw/skills/notebooklm-distiller/scripts/distill.py persist \
  --vault-dir "<path/to/obsidian/vault>" \
  --path "Notes/draft.md" \
  --file ~/Desktop/draft.md

Output format (distill)

Each notebook produces one file at <vault-dir>/<topic>/<NotebookName>_<Mode>.md:

---
title: "<NotebookName> | Deep Q&A"
date: YYYY-MM-DD
type: knowledge-note
author: notebooklm-distiller
tags: ["distillation", "qa", "<topic-slug>"]
source: "NotebookLM/<NotebookName>"
project: "<topic>"
status: draft
---

# <NotebookName> — Deep Q&A

## Q01

> [!question]
> <question text>

**Answer:**

<answer from notebook sources>

---

Output language

Add --lang zh to distill, quiz, or evaluate to get Chinese output. Default is English.

NLM CLI session behaviour

notebooklm ask --new creates ephemeral sessions that are not visible in the NotebookLM web UI. This is by design — the CLI and web interface use separate conversation spaces. Answers are still scoped to the specified notebook's sources.

Error handling

• No notebooks found: verify keywords match notebook titles (use notebooklm list).
• Timeout / rate limit: built-in retry logic and delays. Monitor with ps aux | grep notebooklm.
• Auth failure: run notebooklm login to refresh ~/.book_client_session.