NotebookLM Prompts: How to Write Them, and Why Yours Get Refused

Most guides to NotebookLM prompts hand you a numbered list and leave. That is why the prompts in them so often land flat, or come back with a sentence Google itself documents: "Gemini Notebook can't answer this question." The tool is not being difficult. It is doing exactly what it was built to do, and almost nobody writing prompt lists explains what that is.

Google's own help documentation is blunt about the mechanism. Chat responses draw only on the material you have uploaded. A prompt that works beautifully in a general chatbot can therefore be rejected outright here, while a prompt that would be useless anywhere else becomes the most valuable thing in your week. What follows is the reason that happens, the limits that really apply to what you type, the three-part shape that survives, and fourteen prompts worth copying.

The Key Takeaways

• It is source-grounded by default: Google states that chat responses "only use data from your sources", which is why a creative request gets refused rather than answered badly.
• Google publishes no character limit for the chat box: the widely quoted 10,000 characters belongs to a different field, and the limits Google does publish are 500,000 words per source and 200MB per upload.
• The refusal is a feature, and it has a fix: rephrase toward what your documents contain instead of away from them.
• The best prompts are the ones that would fail elsewhere: contradiction hunting, cross-source comparison and citation-backed study guides all depend on grounding.
• The name changed: Google's help centre now calls it Gemini Notebook, while almost everyone still searches for NotebookLM.

Why NotebookLM Prompts Are Different From Chatbot Prompts

Every general assistant you have used answers from two places at once: what it absorbed in training, and whatever you just pasted in. Gemini Notebook deliberately closes the first door. Google's documentation for chat in Gemini Notebook states it plainly. "Chat responses in Gemini Notebook only use data from your sources."

That single design decision is responsible for almost every frustration people have with the tool, and for everything it does better than the alternatives. It is also why copying a prompt out of a ChatGPT prompt pack so often disappoints here.

The refusal you are getting is documented behaviour

Google gives the example itself. Ask it to "rewrite the end of my short story" and the help page says you may receive "Gemini Notebook can't answer this question" as a response, with the advice to rephrase or ask something different. Nothing has gone wrong. You asked for invention, and invention cannot be traced back to a source.

Once you read the refusal that way it stops being an error message and starts being a signal. It means the request had no anchor in your material. The fix is almost always to point the prompt back at the documents: not write a conclusion for this report but summarise what my sources conclude, and note where they stop short of a conclusion.

Citations change what a good prompt is worth

Answers arrive with citations drawn from the uploaded material. Google notes you can hover over any citation to see the full quoted text, or select it to jump straight to that passage in context. Each source also carries a checkbox, so you can include or exclude it from a given question.

This is the part that should reshape how you prompt. Because every claim is traceable, it is worth asking questions whose answers you intend to check: where sources disagree, which claim rests on the thinnest evidence, what nobody addresses. In a normal chatbot those questions invite invention. Here they are verifiable. If you want the wider picture of what the tool does, our guide to how to use NotebookLM covers the interface itself.

How Long Can a NotebookLM Prompt Be?

This is one of the most common questions about NotebookLM prompts, and most answers on the web get it wrong by mixing up two different boxes. Here is what is actually published, and what is not.

Google's Gemini Notebook FAQ publishes the source limits and says nothing at all about how many characters your question may run to. That is worth stating carefully: Google does not publish a limit for the chat box, which is not the same as promising there is none.

The 10,000-character figure people quote belongs to the custom chat style field, reached through Configure Chat, where you set the model's goal, style and role. Google announced that change itself on its own account in December 2025: "We expanded the character limit for Chat customization from 500 to 10,000 characters, so now you can create much more detailed personas." That figure is official, but it is not documented in the help centre and it does not apply to the question you are typing.

In practice, length is rarely the constraint. A short prompt that names its source, its task and its output beats a long one that rambles, for the same reason a precise database query beats a vague one.

How to Structure a NotebookLM Prompt

Three components do nearly all the work. Miss one and you get something generic; include all three and the output usually needs no second attempt.

Name the sources, or narrow them

A notebook with thirty documents answers differently from one with three. Either say which material to use in the prompt itself, or use the source checkboxes to exclude everything irrelevant before you ask. Using only the two lecture transcripts, not the textbook chapters is a meaningful instruction here in a way it simply is not elsewhere.

Name the task, not the topic

Tell me about mitochondria gives you a summary nobody needed. List every function of mitochondria my sources mention, and flag the ones only one source mentions gives you something you can act on. The verbs that work best are the ones that require the model to move across your material: compare, contrast, count, trace, reconcile, find the gap.

Name the output you want back

Chat can now generate real files, not just text: charts and images, documents including PDF, Word and Markdown, structured data as CSV or JSON, spreadsheets, and slide decks. Asking for a table with one row per study, columns for sample size, method and conclusion costs you nothing and saves the reformatting step entirely.

NotebookLM Prompt Best Practices That Change the Output

Beyond structure, a handful of habits separate people who find the tool transformative from people who find it underwhelming.

Set the role once, in the right box

Rather than opening every prompt with act as a PhD supervisor, set it once through Configure Chat. Google's documentation describes the Custom conversational style as the place to choose a specific style such as "Respond like a PhD student" or suggest a role, alongside a Learning Guide style aimed at grasping new concepts, and a response length setting. Set it once and every answer inherits it.

Ask for disagreement, not agreement

The single highest-value habit. Any model will happily synthesise your sources into a smooth consensus that none of them actually supports. Asking where they conflict forces the opposite, and because every answer carries a citation you can check each side in two clicks.

Ask what is missing

A grounded model knows the edges of its material in a way a general one does not. What question do my sources fail to answer? and which claim here has the weakest support? are close to useless in an ordinary chatbot and sharply useful here. For prompting technique that carries across tools, our general guide to AI prompts goes wider.

NotebookLM Prompts by Job

Google publishes prompts of its own, including a set of seven it posted in August 2026 covering learning, business and personal use. Those lean on the agentic side, asking the model to plot revenue trends or build a spreadsheet from receipts. The fourteen below go the other way and lean on grounding. Paste any of them into a general chatbot with no documents attached and you get invention or a shrug.

Studying and exam preparation

The densest use case by search demand, and the one the tool suits best.

Prompt 1. Write 10 exam questions that can only be answered from the material I uploaded. After each question, cite the source it came from.

Prompt 2. Find the five concepts my sources explain most briefly, and expand each one using only what the sources actually say.

Prompt 3. Build a study guide ordered by how often each topic appears across all my sources, not by the order the lectures used.

If you are studying rather than researching, Gemini Study Notebooks covers the study-specific surface in more detail.

Research and literature review

Prompt 4. Compare all my sources on this topic and list every point where two of them disagree. Quote both sides.

Prompt 5. List every claim in the first source that no other source supports.

Prompt 6. What question do my sources collectively fail to answer, and which one comes closest?

Summarising a stack of documents

Prompt 7. Summarise each source in three sentences, then write one paragraph on what changes when you read them together.

Prompt 8. Give me the argument of each document in one line, then group the documents by which position they take.

Slides and presentations

Prompt 9. Turn my sources into a 10-slide outline. One claim per slide, with the citation for each.

Prompt 10. Draft a slide deck that presents only the findings at least two of my sources agree on.

Infographics and data extraction

Prompt 11. Extract every number from my sources into a table with the metric, the value, and which source it came from.

Prompt 12. Chart how the figures for this metric change across my sources by publication date.

Software engineering and documentation

Prompt 13. Using only the docs I uploaded, list every function that is mentioned but never defined.

Prompt 14. Where does this documentation contradict itself about default behaviour?

What Changed: Chat That Works With or Without Sources

There is a complication worth knowing, because it undercuts the tidy version of the story above. Alongside the grounded chat, Google has added agentic capabilities that can search the web, run code, and produce downloadable files, charts and images. The help documentation states you can use that chat experience "with or without sources."

So the honest position is not that the tool can never look beyond your documents. It is that grounded answering remains the default and the reason to use it, while the agentic layer is something you invoke deliberately. If your prompt needs outside information, say so explicitly rather than assuming it will reach for it.

Google attaches a warning to that layer which is worth repeating before you trust it with anything load-bearing. The documentation calls these functions "experimental and in early development" and says your supervision "is important to help prevent unintended and potentially harmful actions". A prompt that asks the model to go and act is a different risk from one that asks it to read.

Where it still refuses, and what to use instead

None of this changes the creative refusal. A notebook is built to answer from evidence, which makes it excellent for research and a poor choice for drafting, rewriting or anything that needs a voice. The practical workflow most people land on is two tools: the notebook for grounded answers you can cite, and a general model for the writing that follows.

That is the gap Fello AI fills on a Mac. It puts GPT, Claude, Gemini, Perplexity, Grok, DeepSeek and others behind one keystroke, so the draft your notebook refuses to write is one shortcut away rather than another subscription and another tab. It also does its own document work, which our NotebookLM alternatives comparison puts in context against what a notebook is for.

One More Thing: the Name Changed

Google renamed the product in July 2026, and its help centre has followed: the support pages for chat, for upgrading and for the FAQ are each titled Gemini Notebook. Search behaviour has not caught up, which is why nearly everyone still types NotebookLM and why this article does too. Nothing about the prompts changes, but if you are hunting official documentation, the new name is the one that finds it. We covered the transition in NotebookLM is now Gemini Notebook, and what you get on each tier is in our NotebookLM pricing breakdown.

The Verdict

The reason most NotebookLM prompt lists underdeliver is that they were written for a different kind of tool. Prompts that ask a notebook to invent will keep getting refused, and no amount of rephrasing changes that, because the refusal is the product working correctly.

Write prompts that make grounding an advantage instead. Ask where your sources disagree, what they fail to cover, which claim rests on one paper. Those are the questions you cannot safely ask anything else, and they are the ones that come back with a citation you can click. Get that right and the notebook stops feeling like a chatbot with its hands tied and starts being the only tool in the stack you can actually check.

Frequently Asked Questions