Gemini Notebook only answers from documents you give it, and it shows you the line it pulled each answer from. That makes it the most trustworthy AI tool most people are not using properly. This guide gives you the method we use with clients: fix what the tool can see, then build the prompt in seven blocks, then verify. Twenty-one prompts included, plus the skill that writes new ones for you.
Jump to the promptsRead it in order the first time. After that, treat it as a reference and come back for the prompts.
Understand the difference. Ten minutes on why this tool needs different prompts than ChatGPT.
Fix retrieval first. Almost every weak answer traces back to what the tool could see.
Build in seven blocks. The structure that makes a prompt work here.
Steal the prompts. Copy, fill the brackets, paste. Or install the skill and have it write them.
Gemini Notebook, renamed from NotebookLM in July 2026, is a free research workspace from Google. The difference between it and a normal chatbot is one word: grounding. A chatbot answers from everything it has ever read. Gemini Notebook answers only from the materials you give it, and shows you the exact passage each claim came from.
Under the hood this is called retrieval-augmented generation, or RAG. Before it writes a word, the system goes and finds the relevant passages in your documents. Then it writes only from what it found. Two separate steps, and you can influence both.
That changes what good prompting means. With ChatGPT or Claude you are optimizing the wording of a request. Here you are optimizing two things: what gets retrieved (which sources, what scope) and how it gets synthesized (comparison, citations, format). Get the first one wrong and no amount of clever wording saves you.
The single sentence worth remembering: most weak answers are retrieval problems, not wording problems. That is why the method in the next section fixes what the tool can see before it touches the prompt.
Your raw material. PDFs, Google Docs, Sheets and Slides, websites, YouTube links, audio, images, EPUB.
Watch thisGoogle Docs, Sheets, and Slides stay synced automatically. A PDF is a frozen snapshot from the day you uploaded it. If a living document changed last week, your notebook does not know.
Where you ask. Every answer comes back with clickable citations that jump to the exact passage.
CapacityRoughly a million tokens of context on every plan, free included, so it can reason across a large collection at once. History is saved between sessions and stays private to you even in a shared notebook.
The transformation engine. Audio overviews, video overviews, mind maps, reports, flashcards, quizzes, infographics, slide decks, data tables.
NoteThis is where most of the leverage is, and where most people never look. Section 4 covers how to direct it.
Five steps, same order every time, whether you are comparing two vendor contracts or turning a year of meeting notes into a decision brief. Step 2 is the one people skip and the one that causes the most disappointment.
One notebook, one topic. The grounding advantage disappears the moment you mix unrelated subjects, because retrieval can no longer tell which body of material your question belongs to.
Before you upload anything, finish this sentence: "This notebook exists to help me answer questions about ______." If you cannot finish it in one clause, you are building two notebooks.
These are actions you take in the interface, not words you type into the chat. They matter more than the prompt does.
Prompt 1 in the library is the standing chat goal we set on client notebooks. Paste it into Configure Chat once and every answer in that notebook inherits it.
Not every prompt needs all seven, but three of them are never optional: source scope, the citation rule, and the gap rule. Section 3 breaks down all seven with examples.
The shortcut version: name a role, use one precise verb, say exactly which documents to use, say how to combine them, demand citations, forbid guessing, and specify the shape of the output.
Click the citations. Every one that matters. Inline citations are the most valuable feature in this tool and the one people ignore, and checking them is the entire difference between research and confident guessing.
Then ask yourself the question the gap rule is designed to surface: did it tell me what it could not find? An answer with no acknowledged gaps in a messy source set is a warning sign, not a good result.
Once you have an answer you trust, Studio turns it into whatever format the work needs: a briefing doc, a podcast, a deck, a table, a quiz. Section 5 covers the pipeline that makes those outputs sharp instead of rambling.
One rule before you generate anything: select only the relevant sources first. Studio uses whatever is checked.
Assemble in this order. Three of them are marked never skip, because leaving them out is what produces confident, uncited, subtly invented answers.
A persona tied to how evidence gets used, not decoration. Tell the role to stay inside the sources.
Act as a procurement analyst who works only from the documents provided.
One precise verb for the synthesis you want. Compare, reconcile, extract, map, rank, audit, distill, test. Never "tell me about."
Reconcile the delivery dates across these three documents.
Name exactly which documents to use or exclude, mirroring what you checked in the left panel.
Use only the two sources named "Vendor A Proposal" and "Vendor B Proposal." Ignore anything dated before 2024.
How to combine evidence across documents. This is the block that turns a summary into an analysis.
Surface where they agree first, then where they contradict each other, then what only one of them covers.
Make every factual claim traceable. For anything client-facing, ask for the section or page too.
Support every factual statement with a citation to the source it came from, and note the section it appears in. Keep what the sources say separate from your interpretation.
The guardrail. Without it the tool will quietly reach for general knowledge to fill a hole.
If the sources do not contain the answer, say exactly that. Do not fill gaps with general knowledge.
Structure, length, and audience level. Tables for comparisons, numbered findings for audits, plain paragraphs for narrative.
Table first, then a summary under 200 words written for an owner, not a lawyer.
A real prompt for comparing two vendor proposals. Every block is in here.
Act as a procurement analyst who works only from the documents provided.
Compare only the two sources named "Vendor A Proposal" and "Vendor B Proposal."
Build a table with rows for price, implementation timeline, support terms, and
contract exit clauses. For each row, cite the exact section of each proposal the
information came from. After the table, list every point where the proposals
contradict each other, and every point where one is silent on something the other
covers. If either proposal does not address a row, write "not addressed" rather
than guessing. Keep the closing summary under 200 words, written for a business
owner, not a lawyer.Notice what is not in there: no "please," no "you are an expert AI assistant," no flattery. Every sentence is doing a job.
Studio takes custom instructions before it generates. Most people click the button and accept whatever comes out, then conclude the feature is mediocre. It is not. It is undirected.
A podcast-style conversation between two AI hosts, natural enough that most listeners would not immediately clock it as AI.
Pick the formatDeep Dive (two hosts explore), Brief (short summary), Critique (expert review of your material), or Debate (hosts take opposing sides). Critique is underrated for pressure-testing your own work.
Then direct it"Spend most of the time on [topic], address the listener as a [role], keep it practical, end with three action steps."
A narrated explainer with custom visual styles, from classic to whiteboard to watercolor. Short vertical format available for social.
Direct itSpecify audience level, the narrative arc (problem, evidence, recommendation), what belongs on screen, and pacing.
For client workAsk for citations to appear on the slides. Traceability is the reason you are using this tool.
Two options. A detailed deck carries full text and reads on its own, so it emails well. Presenter slides stay visual with talking points for a live room.
Direct itDeck length, which format, and the slide-by-slide flow if you have one in mind. You can revise slide by slide afterward and export to PPTX.
Brand itUpload your brand guidelines as a source, then instruct it to follow that document for colors, fonts, and style.
Briefing docs, study guides, blog posts, or a custom format you describe.
Direct itGive it the same seven-block treatment as a chat prompt: scope, synthesis directive, citation rule, audience.
Pro moveGenerate the report from a specific chat thread rather than the whole notebook, after you have already worked the subtopic properly in conversation.
A structured table pulled straight from your sources, for comparisons, inventories, and specs.
Direct itDescribe the columns explicitly. "One row per [unit], columns for [fields], every cell traceable to a source."
LimitExports to Sheets, but edits in the Sheet do not sync back. Treat the export as a one-way snapshot.
Active recall built from your material, and both work on a phone.
Direct itSet difficulty and quantity, and scope which sources the questions may draw from.
Best useOnboarding someone onto a policy or a product. You already have the documents. This turns them into training in about a minute.
A visual summary generated from your sources, with adjustable orientation and detail level.
Direct itSay the one message it must communicate. Without that it will try to show everything and land nothing.
LimitDenser settings on complex topics can produce small text errors. Read it before it goes anywhere public.
Shows how concepts connect across your sources, which surfaces relationships you would miss reading in a straight line.
Best forUnderstanding the shape of a body of material before you try to write about it.
LimitDesktop only.
On paid tiers, on desktop, notebooks include a sandboxed environment that can write and run code, producing spreadsheets, decks, PDFs, charts, and CSVs.
Best forCross-source calculations, reconciling messy spreadsheets, and finished analytical deliverables without leaving the tool.
CaveatGoogle labels this experimental and recommends human supervision. Check the numbers before anything high-stakes leaves your desk.
Plan limits on notebooks, sources, and daily generations vary by tier and change often. We have deliberately left specific numbers out of this guide rather than publish figures that go stale. Check the current limits in the product.
This is the most powerful workflow in the tool and the least obvious. It is the difference between a rambling twenty-minute podcast and a sharp eight-minute one, and it works the same way for decks, reports, and videos.
Use the prompt library to work the material until you have the exact angle you need. Compare, find the contradictions, map the evidence. This is the thinking part, and it is worth doing properly.
Anything you would be comfortable putting your name on, save. Everything else, let it scroll away.
Your notebook now contains a distilled, already-synthesized version of the material alongside the raw pile.
Uncheck the raw documents. The podcast, deck, or video gets built from the refined view rather than from four hundred pages, so it stays on message instead of wandering.
Your notebook gets smarter every time you use it this way. Chat responses become notes. Notes become sources. The next question starts from a better place than the last one did.
Work through these in order. In our experience the first two account for most of it, and almost nobody checks them before rewriting the prompt for the fourth time.
Nobody needs five AI subscriptions. But these tools are good at genuinely different jobs, and reaching for the wrong one is why people decide AI does not work for their business.
| Tool | What it is actually good at | Reach for it when |
|---|---|---|
| Gemini Notebook | Answering from a body of material you control, with citations you can click | You already have the documents, and being able to prove where an answer came from matters |
| ChatGPT | General-purpose work, broad tasks, connecting to other tools | The answer lives outside your documents, or you need it to go do something |
| Claude | Reasoning, writing, and holding a long complicated thread without losing the plot | You want a thinking partner on something hard, or you are building something |
| Perplexity | Fast, sourced search of the live web | You are assembling the corpus, not analyzing it yet |
| Notion AI | AI sitting inside the docs and databases your team already lives in | Notion is your operating system and you want AI woven through it |
A useful sequence rather than a competition: Perplexity to find it, Gemini Notebook to ground it, Claude to build something with it. Most heavy users keep two or three in rotation and stop pretending one tool does everything.
Most people are not short on tools. They are short on a clear read of where their hours are going. The free Quick Wins™ Assessment takes three minutes and shows you 5 to 10 small changes that could give you hours back every week.
Take the free 3-minute assessment Book the AI Quick Wins™ Briefing CallTwenty-one prompts built on the seven blocks, grouped by what you are trying to do. Every one of them carries a source scope, a citation rule, and a gap rule, because those are the three that keep the answers honest. Copy one, replace anything in [brackets], paste it into the chat panel. Start with number 1: it goes in Configure Chat, not the chat box, and it improves every answer that follows.
Twenty-one prompts cover a lot, but not your specific notebook on your specific Tuesday. This is the skill we built to close that gap. You tell Claude what you are trying to get out of your notebook, it asks you a handful of questions, and it hands back a setup checklist plus a ready-to-paste prompt built on the same seven blocks. It also diagnoses answers that came back weak, if you paste in what you got.
What it will do on a first run: ask you about the job, what is in the notebook, how wide the scope should be, who the audience is, and whether you need every claim traceable. Then it delivers the checklist and the prompt, explains in one line why it built it that way, and offers to tune it after you have seen the answer.
Scroll inside the box, or hit copy and paste the whole thing.
---
name: gemini-notebook-prompt-architect
description: Guided prompt builder for Google Gemini Notebook (formerly NotebookLM). Interviews the user about their notebook and goal, then generates an optimized source-grounded prompt plus retrieval setup steps. Use when the user mentions Gemini Notebook or NotebookLM prompts, wants better answers or citations from their notebook, or wants to create a Studio output (Audio Overview, Video Overview, slide deck, report, quiz, flashcards, mind map, infographic, data table). Triggers include "help me prompt Gemini Notebook", "NotebookLM prompt", "optimize my notebook question", "my notebook keeps giving weak answers".
---
# Gemini Notebook Prompt Architect
You are a prompt architect for Google Gemini Notebook (renamed from NotebookLM in July 2026). Your job: interview the user briefly, then hand them a ready-to-paste prompt plus a short setup checklist that makes the prompt actually work.
Brought to you by aha! Talent Experts. Always write aha! in lowercase. Never use em dashes in anything you produce with this skill.
## Why this tool needs different prompts
Gemini Notebook is not a general-purpose chatbot. It is a source-grounded system: it answers only from the documents the user has uploaded, using a technique called Retrieval-Augmented Generation (RAG). In plain terms, before writing an answer it first retrieves relevant passages from the user's sources, then writes only from what it retrieved.
That changes everything about prompting:
- With ChatGPT or Claude, you optimize the wording of the request.
- With Gemini Notebook, you optimize two things: what gets retrieved (which sources, which scope) and how it gets synthesized (comparison, citations, format).
Most weak answers are retrieval problems, not wording problems. So this skill always fixes retrieval setup first, then writes the prompt.
## What you deliver
Every run of this skill ends with:
1. A setup checklist (2 to 5 steps): which sources to select, what to put in Configure Chat, whether to start a fresh chat.
2. The optimized prompt in a copy-paste code block.
3. When relevant: a reusable "chat goal" persona block and/or Studio generation instructions in their own code blocks.
4. One sentence explaining why the prompt is built the way it is, so the user learns the pattern.
Then offer one refinement pass: "Paste the answer you get back and I'll tune the prompt."
## Step 1: Interview the user
Ask questions one at a time. Never more than two in a single message. Skip anything the user already told you. You need answers to these five things:
1. **The job.** What do they want out of the notebook? Common jobs: answer a specific question, summarize or synthesize, compare documents, find conflicts or gaps, extract every mention of something, make a decision, create study materials, or generate a Studio output (podcast, video, slides, report, quiz, flashcards, mind map, infographic, data table).
2. **What is in the notebook.** Roughly how many sources, what kinds (PDFs, Google Docs, websites, YouTube videos, meeting notes), and whether they trust all of them equally.
3. **Scope.** Should the answer draw on the whole notebook or only certain documents? If certain ones, get their names.
4. **Audience and voice.** Who reads or hears the result, and at what expertise level?
5. **Citation needs.** Do they just want the answer, or do they need every claim traceable to a source (client deliverables, compliance, research)? Default to requiring citations. Traceability is the whole point of this tool.
If the user gives you a vague ask like "make my notebook answers better," ask for a real example of a prompt they ran and the answer that disappointed them. Diagnose from the example.
## Step 2: Fix retrieval before wording
Based on the interview, include the relevant items in the setup checklist. These are actions the user takes in the Gemini Notebook interface, not prompt text:
- **Select only the sources that matter.** Every source has a checkbox in the left panel. Deselecting irrelevant sources is the single highest-impact move. Fewer, tighter sources means more precise retrieval.
- **Set a chat goal.** In the chat panel, the Configure (gear) icon opens chat settings. The Goals field accepts up to 5,000 characters and applies to every message in that notebook. Standing instructions belong here (persona, citation rules, audience), so each individual prompt stays short. Also set response length here if the user wants consistently brief or detailed answers.
- **Start clean for important runs.** Chat history is saved and prior exchanges can color new answers. For a high-stakes prompt, save anything valuable as a note, then clear or start a fresh chat.
- **Fill gaps with Deep Research.** If the notebook is missing coverage on a topic, the built-in research tools can find sources on the web or in Google Drive and add them. Fast Research returns a list of candidate sources to review; Deep Research reads them and writes a sourced report into the notebook.
- **Promote good answers to sources.** A great chat response saved as a note can be converted into a source. This is the key move for focused Studio outputs (see the pipeline section below).
- **Mind the source mix.** Google Docs, Slides, and Sheets stay synced automatically; PDFs are frozen snapshots from the day of upload. If a "living" document changed recently, confirm the notebook has the current version.
Teach as you go: give a one-line reason for each checklist item you include.
## Step 3: Build the prompt from the seven blocks
Assemble the prompt from these blocks, in this order. Not every prompt needs all seven, but never skip Source scope, Citation rule, or Gap rule.
1. **Role.** A persona tied to evidence use, not decoration. Examples: research methodology analyst, strategic product manager, devil's advocate, middle school teacher, hiring manager coach. The role must be told to stay grounded in the sources.
2. **Task verb.** One precise verb for the synthesis you want: compare, reconcile, extract, map, rank, audit, distill, test. Avoid "tell me about."
3. **Source scope.** Name exactly which documents to use or exclude, mirroring the checkbox selection. Examples: "Use only the two documents named X and Y." "Ignore anything published before 2024." "Exclude the draft versions."
4. **Synthesis directive.** How to combine evidence across documents: compare positions, surface agreements and contradictions, map every piece of evidence for and against, trace how a number or claim changes across documents.
5. **Citation rule.** "Support every factual statement with a citation to the source it came from." For client-facing work, add: "Note the section or page where each claim appears" and "Clearly separate what the sources say from your interpretation."
6. **Gap rule.** "If the sources do not contain the answer, say exactly that. Do not fill gaps with general knowledge." This is the guardrail that keeps the answer honest.
7. **Output format.** Structure, length, and audience level. Tables for comparisons, numbered findings for audits, plain paragraphs for narratives.
### Worked example
User job: compare two vendor proposals for a franchise client and flag risks.
```
Act as a procurement analyst who works only from the documents provided.
Compare only the two sources named "Vendor A Proposal" and "Vendor B Proposal."
Build a table with rows for price, implementation timeline, support terms,
and contract exit clauses. For each row, cite the exact section of each
proposal the information comes from. After the table, list every point where
the proposals contradict each other or where one is silent on something the
other covers. If either proposal does not address a row, write "not addressed"
rather than guessing. Keep the summary under 200 words, written for a
franchise owner, not a lawyer.
```
## Step 4: Deliver
Present the output in this order: setup checklist, prompt block, optional goal block, optional Studio block, one teaching line, offer of a refinement pass. Keep commentary minimal. The deliverable is the prompt, not an essay about the prompt.
## Prompt pattern library
Reach for these when the user's job matches. Adapt names and details from the interview.
**Cross-document comparison.** "Compare how [Doc A] and [Doc B] each handle [topic]. Show agreements first, then disagreements, then topics only one covers. Cite each point."
**Contradiction finder.** "Identify every place the selected sources contradict each other about [topic]. For each contradiction, quote or closely paraphrase both positions with citations, and state which source is more recent or more authoritative if that can be determined from the sources themselves."
**Evidence map.** "Map all evidence for and against the claim that [claim]. Two lists, every item cited. End with what evidence is missing that would settle the question."
**Exhaustive extraction.** "Find every mention of [term or theme] across all selected sources. List each with its source and context. Do not summarize; I need the complete inventory."
**Devil's advocate.** "Act as a devil's advocate. Using only the selected sources, build the strongest case against [plan or conclusion]. Cite the evidence behind each objection. Where the sources give you nothing to object with, say so."
**Timeline builder.** "Reconstruct the timeline of [events or decisions] across all sources. Flag any dates that conflict between sources, with citations for each version."
**Decision brief.** "I need to decide [decision]. From the selected sources only: the three strongest factors for, the three strongest against, the key unknowns the sources do not resolve, and a one-paragraph recommendation clearly labeled as interpretation. Cite everything factual."
**Translate for an audience.** "Explain [topic] from the sources at a level a [new franchisee / middle schooler / board member] would follow. Keep citations on factual claims. Flag anywhere you simplified enough that nuance was lost."
## Studio outputs: prompt the generator, not just the chat
Studio outputs (in the right panel) accept custom instructions before generating. The available formats as of August 2026: Audio Overview, Video Overview (including short vertical format, plus cinematic on Ultra), Mind Map, Reports, Flashcards, Quiz, Infographic, Slide Deck, and Data Table.
General rules for all Studio instructions:
- Name the audience and their level.
- Name the focus: which themes to spotlight and which to skip.
- Scope the sources first with checkboxes; Studio uses whatever is selected.
- For the tightest results, generate from a synthesis source (see pipeline below) instead of the whole raw notebook.
Format-specific guidance:
- **Audio Overview.** Choose the format first: Deep Dive (two hosts explore), Brief (short summary), Critique (expert review of the material), or Debate (hosts take opposing sides). Then direct the hosts in the instructions: "Spend most of the time on [topic], address the listener as a [role], keep a practical tone, end with three action steps."
- **Video Overview.** Specify audience level, the narrative arc (problem, evidence, recommendation), what belongs on screen, and pacing. Ask for citations to appear on slides when traceability matters. Style options range from classic to whiteboard to watercolor.
- **Slide Deck.** Specify deck length, detailed versus presenter format, and slide-by-slide flow if you have one in mind. Decks can be revised slide by slide with follow-up prompts and exported to PPTX. If the user has brand guidelines, upload them as a source and instruct: "Follow the attached brand guideline for colors, fonts, and design style."
- **Reports.** Pick the format (briefing doc, study guide, blog post, or custom) and give the same seven-block treatment as a chat prompt: scope, synthesis, citations, audience.
- **Quiz and Flashcards.** Set difficulty and quantity, and scope which sources the questions may draw from.
- **Infographic.** Specify orientation, detail level, and the one message it must communicate.
- **Data Table.** Describe columns explicitly: "One row per [unit], columns for [fields], each cell traceable to a source." Tables can export to Sheets.
## The Chat, Notes, Sources, Studio pipeline
The most powerful workflow in the tool, and the one to recommend whenever a user wants a polished Studio output from a large messy notebook:
1. **Explore in Chat.** Use the patterns above to synthesize exactly the angle they need.
2. **Save the best responses as Notes.**
3. **Convert those notes into Sources.** The notebook now contains a distilled, pre-synthesized version of the material.
4. **Generate the Studio output from only the synthesis sources selected.** The podcast, deck, or video is built from the refined view instead of hundreds of raw pages, so it stays on message.
Explain this in one or two sentences when you recommend it. It is the difference between a rambling 20-minute podcast and a sharp 8-minute one.
## Diagnosing weak answers
When a user says the notebook gave a bad answer, check causes in this order:
1. Wrong sources selected, or too many. Narrow the checkboxes.
2. The information is not actually in the sources. Add sources or run Deep Research. The gap rule in the prompt will reveal this.
3. The prompt asked for general knowledge ("what do you know about X"). Rewrite as a source-grounded task ("what do my sources say about X, cited").
4. Old chat history steering the answer. Fresh chat.
5. Missing synthesis directive, so the tool summarized one document instead of comparing several. Add block 4.
## Current product facts (August 2026)
Use these when relevant; verify with a quick search if something seems off, since the product changes fast.
- NotebookLM was renamed Gemini Notebook in July 2026. Same product, now integrated with the Gemini app, with notebooks syncing across both.
- The chat holds up to 1 million tokens of context on all plans, free included, so it can reason across a large source collection at once.
- Chat goals (Configure Chat) accept up to 5,000 characters and persist per notebook. Chat history is saved between sessions and private to each user even in shared notebooks.
- Deep Research can pull sources from the web or Google Drive. Supported source types include PDFs, Google Docs, Sheets, Slides, websites, YouTube, audio, images, and EPUB. Google file types stay synced; PDFs do not.
- Newer capabilities on paid tiers include code execution for data analysis and cinematic video overviews. Plan limits on notebooks, sources, and daily generations vary by tier and change often; do not quote specific numbers without checking.
- Citations in chat answers are clickable and jump to the exact source passage. Encourage users to keep them in the workflow even when reformatting for APA or MLA on top.
## Final check before you hand over a prompt
- Source scope named? Citation rule in? Gap rule in?
- Does the prompt ask the notebook to retrieve and synthesize, not to "know"?
- Is anything in the prompt really a standing instruction that belongs in the chat goal instead?
- No em dashes anywhere. aha! written in lowercase if mentioned.
- Did you include the setup checklist and offer the refinement pass?Normally you work inside Gemini Notebook in your browser. There is a free community-built connector that lets Claude reach into your notebooks directly, so you can ask Claude a question and have it pull the cited answer out of your sources without you switching tabs. Setup is about five minutes and involves copying four short lines into a terminal window. You do not need to know how to code. You need to be able to copy and paste.
nlm, and it does the connecting for you.Before you start: you need a Google account with Gemini Notebook set up, and either Claude Desktop or Claude Code installed. Do the steps in order. If a line fails, stop there rather than skipping ahead.
Open your terminal and paste this one line. It installs both pieces you need, the nlm command and the connector itself.
uv tool install notebooklm-mcp-cliIf your computer says it does not know what uv is, use one of these instead:
pip install notebooklm-mcp-clipipx install notebooklm-mcp-cliYou can also try it without installing anything permanently: uvx --from notebooklm-mcp-cli nlm --help
This one opens a browser window and asks you to sign into Google, the same way any login works. Once you are in, it saves the session so you do not have to do it again for a few weeks.
nlm loginTo check later whether you are still signed in, run nlm login --check. If you keep separate Google accounts for work and personal, add a profile name: nlm login --profile work
If you use Claude Desktop (the app), the easiest path skips the terminal entirely. Go to the project page linked at the bottom of this section, download the one-click .mcpb extension file, and double-click it. Done, no editing anything.
If you would rather use the command, quit Claude Desktop completely first. Not minimized, fully quit. Then run:
nlm setup add claude-desktopReopen Claude Desktop afterward. The reason for quitting first: Claude rewrites its own settings file when it closes, so anything added while it is running gets wiped. The tool will refuse to run rather than let that happen.
If you use Claude Code (the terminal version), it is one line:
nlm setup add claude-codeThis command checks the install and the login and tells you in plain language if something is off.
nlm doctorThen open Claude and ask it something harmless: "list my notebooks." If your notebooks come back, you are connected. Start with a question that only reads, never one that changes or deletes anything, until you trust it.
Claude Code users can install an expert skill that helps Claude pick the right tool for the job instead of guessing.
nlm skill install codexSkip this if you are on Claude Desktop. It will not hurt anything, it just will not apply.
nlm login again. That is usually the whole fix.@gemini-notebook-mcp when you are not actively using it, and back on when you are.notebooklm-cli or notebooklm-mcp-server as separate packages, remove them first with uv tool uninstall notebooklm-cli, then reinstall using uv tool install --force notebooklm-mcp-cli. Skipping this leaves broken shortcuts behind.Project page and full documentation: github.com/jacob-bd/gemini-notebook-mcp-cli
Scope it to a single job, ground it before you prompt, and run three questions through the seven blocks. You will know inside half an hour whether this earns a place in your week. If you would rather have someone map where AI actually fits across your business, that is the conversation we have on the AI Quick Wins™ Briefing Call.
Book the AI Quick Wins™ Briefing Call