NotebookLM Prompt Templates for Source-Based Research

TL;DR: NotebookLM prompts work backwards from normal prompting. The tool answers only from sources you upload, so the best prompts ask it to extract, compare, trace and audit what is already there. Google's own help pages confirm chat responses use no outside knowledge. Below: 28 copy-paste templates.

What are NotebookLM prompts, and why don't ChatGPT prompts work here?

A NotebookLM prompt is an instruction aimed at a closed corpus. You upload the sources, and the model answers from those and nothing else. Google states it directly in the chat documentation: "Chat responses in Gemini Notebook only use data from your sources."

That single sentence inverts almost everything you know about prompt engineering. In ChatGPT you are trying to pull the right slice out of an enormous, invisible training set, which is why persona framing, worked examples and chain-of-thought instructions help so much. In NotebookLM the corpus is small, known and sitting in the left-hand panel. You are not trying to summon knowledge. You are trying to interrogate a specific pile of documents.

So the prompts that earn their keep here are the ones a good research assistant would run: pull every claim, line up what three papers say side by side, find where they disagree, trace a number back to the sentence it came from, and tell me what my sources do not cover.

Prompts that ask for outside knowledge, or for invention, either fail or produce something worse than failure. Google's FAQ is blunt about it: "Gemini Notebook answers questions based on the information provided in your uploaded sources. If the answer isn't in the source material, it won't provide a response."

Is NotebookLM still called NotebookLM?

Officially, no. Google renamed NotebookLM to Gemini Notebook on 16 July 2026, announced by Josh Woodward, VP of Google Labs, Gemini app and AI Studio. Google's help centre now says Gemini Notebook on every page.

Almost nothing else moved. Google described it as "the same standalone product, now doing more across the Google ecosystem and updated with a secure cloud computer" that lets it "write and execute code natively" for data analysis grounded in your sources. The Workspace announcement confirmed that "automatic redirects will ensure existing shared notebooks and user links continue to work," and that admins had nothing to do.

Old URLs still work. Checked on 26 August 2026, notebooklm.google.com returns a 301 redirect to notebook.google.com.

Throughout this post I use "NotebookLM" because that is still what people type into search and say out loud. Every fact is sourced to the Gemini Notebook help centre, accessed 26 August 2026.

What can NotebookLM actually see, and what are the limits?

Your prompts are only as good as what you fed the notebook, so the ceilings matter. These come from Google's upgrade page and FAQ, both accessed 26 August 2026, and Google labels them "subject to change".

Per source, the cap is 500,000 words or 200MB for local uploads, and Google says there is no page limit. Imports fail on three things: exceeding the word limit, exceeding the file-size limit, or a copy-protected PDF.

Supported source types, per Google's source documentation: Google Docs, Google Slides (up to 100 slides), Google Sheets (limited to 100k tokens), PDF, TXT, Markdown, DOCX, PPTX, CSV, ePub, web URLs, YouTube URLs for public videos, pasted text, Gemini Chats, images in a long list of formats including PNG, JPG, WEBP, HEIC and TIFF, and audio files including MP3 and WAV.

That last group matters more than it looks. An interview recording, a lecture, a screenshot of a table: all of them become quotable, citable sources. Half the value of a good notebook is deciding what goes in it.

Where do you type a prompt in NotebookLM?

Four places, and most people only ever use the first.

1. The chat box. The main surface. Everything in the template sections below goes here.

2. Configure Chat. In the chat panel there is a Configure Chat control. It sets a conversational style, either Default, Learning Guide, or Custom, plus a response length of Shorter, Default or Longer. Google's own examples of a Custom style are "Respond like a PhD student" and "Pretend to be a role-playing game host." This is the nearest thing NotebookLM has to a system prompt, and it persists across the notebook, so it is worth setting once rather than repeating a persona in every message.

3. Studio artefact prompts. Audio Overviews take an instruction to focus on specific topics or adjust the expertise level. Video Overviews take an optional steering prompt. Reports let you write your own brief instead of picking FAQ, study guide or briefing document. Useful detail from Google's docs: you can select "View custom prompt" in the three-dot menu next to a generated artefact to see the prompt that produced it, and flashcards, quizzes and videos show a "Show prompt" option next to the title.

4. The source checkboxes. Not a prompt, but functionally part of one. Google's docs: "you can use the checkbox on each source to include or exclude certain sources the model should use to answer your question." Google also spells out how the pieces combine. Sources are "always used in either the entire set or the subset you select", conversation history is used to generate responses, and notes are used "only when you specifically select it."

Deselecting sources is the highest-leverage move in the whole tool. Half the prompts below get dramatically better when you narrow the corpus first.

Extraction prompts: what exactly do these sources say?

Extraction is the job NotebookLM does best, because it is a retrieval problem with a citation attached. Keep the output shape explicit and ask for a citation on every row.

1. Claim inventory

List every distinct claim made across the selected sources.

For each claim, give me:
- The claim in one sentence, in the source's own framing
- The source it came from
- Whether it is presented as a finding, an assumption, or a citation of other work

Cite every row. If two sources make the same claim, list it once and
name both sources. Do not add claims that are not in the sources.

2. Definition extraction

Find every place the selected sources define or characterise "[TERM]".

For each, quote the defining sentence verbatim, name the source, and note
whether the definition is formal, operational, or informal.

If the sources use the term without ever defining it, say so explicitly.

3. Method and sample extraction

For each source, extract the following into a table:
Source | Study design | Sample or dataset | Size | Time period | Primary outcome measure

Leave any cell blank and write "not stated" if the source does not report it.
Do not estimate or infer any value.

4. Every number, with its context

Extract every quantitative figure in the selected sources.

Table columns: Figure | Unit | What it measures | Population or scope | Source | Page or section

Include figures that appear only in tables, footnotes or figure captions.
Do not convert units. Do not round.

5. Stated limitations

List every limitation, caveat, or threat to validity the authors state
about their own work.

Quote each one verbatim, attribute it, and group them by type:
sample, measurement, scope, generalisability, conflict of interest, other.

Only include limitations the authors state themselves.

6. Thematic quote pull

Pull verbatim quotes about "[THEME]" from the selected sources.

Give me 8 to 12 quotes. For each: the quote, the speaker or author,
and one sentence of surrounding context so I can tell whether the
author endorses the position or is describing someone else's.

Verbatim only. Do not paraphrase or tidy the wording.

Cross-source synthesis prompts: what do these sources say together?

This is where a notebook beats a folder of PDFs. The instruction that makes or breaks these prompts is per-source attribution. Without it you get a blended summary that reads well and cannot be checked.

7. Comparison matrix

Build a comparison matrix of the selected sources.

Rows: [DIMENSION 1], [DIMENSION 2], [DIMENSION 3], [DIMENSION 4]
Columns: one per source

Fill each cell with the source's actual position, cited. Write
"not addressed" where a source is silent on a dimension. Do not
infer a position from a source that does not state one.

8. Consensus and minority positions

On the question "[QUESTION]", split the selected sources into:

1. Sources that agree, and the exact position they share
2. Sources that hold a minority or dissenting position
3. Sources that do not address the question

For 1 and 2, cite the specific passage that places each source in
that group. Do not weight by how many sources are in each group.

9. Chronology of a position

Order the selected sources by publication date.

Then trace how the treatment of "[TOPIC]" changes across them:
what was assumed early, what was revised, what was abandoned,
and what stayed constant.

Cite the passage that marks each shift. If dates are not stated
in the sources, say so rather than guessing.

10. Terminology reconciliation

Identify terms that the selected sources use differently from one another.

For each: the term, each source's usage quoted verbatim, and whether the
difference is a genuine conceptual disagreement or just wording.

Flag any case where two sources appear to agree but are using the same
word to mean different things.

11. Attributed synthesis

Write a synthesis of what the selected sources establish about "[TOPIC]".

Rules:
- Every sentence carries a citation
- Where sources differ, say so in the sentence rather than smoothing it over
- Anything supported by only one source is labelled "single-source"
- No sentence that draws on knowledge outside these sources

Length: [N] paragraphs.

Contradiction-finding prompts: where do the sources disagree?

Contradiction hunting is the prompt category most people never try, and it is the one that most justifies loading 30 documents into a notebook. Models are agreeable by default and will paper over conflict unless you ask for it directly.

12. Direct contradiction hunt

Find every place where two or more of the selected sources make
incompatible claims.

For each conflict:
- State both claims in their own words, quoted
- Name both sources
- State precisely what the disagreement is about

Report conflicts even where one source looks obviously more credible.
If you find no contradictions, say so rather than inventing a mild one.

13. Numerical disagreement

Compare every figure that appears in more than one source and
measures the same thing.

Table: What is measured | Value in each source | Difference | Are the
definitions or populations the same?

Flag any case where the figures look comparable but the underlying
definitions differ.

14. Why they disagree

Take the disagreement about "[CLAIM]" between [SOURCE A] and [SOURCE B].

Explain, using only what the sources state about their own methods,
what could account for the difference: population, time period,
measurement, analysis choice, or definition.

If the sources do not give enough methodological detail to explain
the gap, say that explicitly instead of speculating.

15. Self-contradiction check

Read [SOURCE] on its own.

Find any place where it contradicts itself: abstract versus results,
a claim in the discussion not supported by its own data, a limitation
that undercuts a stated conclusion, or a figure that disagrees with
the text.

Quote both halves of each inconsistency.

Citation-tracing prompts: can I prove where this came from?

NotebookLM's citation behaviour is the reason to use it over a general chatbot for research work. Google's docs describe it plainly: the tool "uses direct quotes, text, and images straight from your sources as citations", you can hover a citation to see the full quoted text, and selecting one jumps to the passage in context.

One documented gotcha: "If your source content is too short, Gemini Notebook references the entire document without a cited individual text from your source." Short sources get document-level citations rather than passage-level ones.

16. Trace a claim to source

Find the origin of this claim in the selected sources: "[CLAIM]"

Give me:
- The exact passage that supports it, quoted
- The source and location
- Whether that source is making the claim itself or citing someone else
- If it cites someone else, who

If no passage supports the claim, say "not supported by these sources".

17. Verify a pasted claim

Here is a statement from a draft I am editing:

"[PASTE STATEMENT]"

Check it against the selected sources only. Return one of:
SUPPORTED (quote the passage), PARTIALLY SUPPORTED (quote it and state
what the draft adds), CONTRADICTED (quote the contradicting passage),
or NOT ADDRESSED.

Do not use any knowledge outside these sources to judge it.

18. Provenance chain

For the finding "[FINDING]", map the chain of attribution across the
selected sources.

Which source reports it as original work? Which cite it secondhand?
Does any source cite it to a document that is not in this notebook?

Show the chain as a list. Flag anything that looks like a citation
loop or a claim that has drifted from its original wording.

19. Quote table with locators

Build a quote table for "[TOPIC]".

Columns: Verbatim quote | Author | Source title | Section or page | What
it is evidence for

Only include quotes you can cite to a specific passage. Exclude anything
you would have to paraphrase or reconstruct.

20. Audit my own draft

I am pasting a section of my draft below. Every factual statement in it
is supposed to be supported by the sources in this notebook.

Flag every sentence that is NOT supported, and for each, say whether it
is unsupported, overstated relative to the source, or contradicted.

Ignore sentences that are clearly my own argument rather than a
factual claim.

DRAFT:
[PASTE]

Gap-identification prompts: what is missing from these sources?

This is the highest-value category and the least used. A tool that can only see your sources is unusually well placed to tell you where your sources run out, as long as you ask for absence rather than presence.

21. Unanswerable questions

Given only the sources in this notebook, list the questions about
"[TOPIC]" that these sources cannot answer.

For each, say what kind of evidence would be needed and whether any
source gestures at it without providing it.

Be specific. "More research is needed" is not an acceptable answer.

22. Missing evidence for a specific claim

The claim "[CLAIM]" appears in these sources.

What evidence would be needed to establish it properly, and how much
of that evidence is actually present here?

Separate: evidence present and cited, evidence asserted but not shown,
and evidence absent entirely.

23. Coverage gaps

Describe the coverage of the selected sources along these axes:
population studied, geography, time period, methods used, and
perspectives represented.

Then state what is systematically missing or under-represented,
based only on what the sources themselves report about their scope.

24. What would change the conclusion

Based only on these sources, what single finding would most undermine
the dominant conclusion in this notebook?

Then check: do any of the sources already contain partial evidence
pointing that way? Quote it if so.

Steering prompts for Audio Overviews, Video Overviews and Reports

The Studio panel produces Notes, Audio Overviews, Video Overviews, Mind Maps, Reports, Data Tables, Flashcards, Quizzes, Slide Decks and Infographics. Three of those take free-text steering, and the defaults are generic without it.

Audio Overviews come in four formats per Google's docs: Deep Dive (the default, two hosts in conversation), The Brief (a single speaker covering key takeaways in under two minutes), The Critique (two hosts giving a constructive evaluation of something like an essay or design doc), and The Debate (two hosts arguing formally). Length can be set to Shorter, Default or Longer, with Longer marked English-only, and Google says Audio Overviews are available in 80+ languages.

25. Audio Overview steering prompt

Audience: [WHO], who already knows [BASELINE] and does not need
definitions of [OBVIOUS TERMS].

Focus only on: [TOPIC 1], [TOPIC 2].
Skip entirely: [TOPIC TO SKIP].

Where the sources disagree, say so out loud rather than picking a side.
Name the source when you state a finding.

26. Video Overview steering prompt

Video Overviews come as Cinematic, Explainer, or Short (around 60 seconds), with visual styles including Classic, Whiteboard, Watercolor, Retro Print, Heritage, Paper-craft, Kawaii and Anime. Google notes Cinematic and Short support English only and require users aged 18 or over, and that generation can take a while.

Focus the video on [SPECIFIC QUESTION] rather than a general overview.

Structure: the problem, what the sources found, what remains contested,
what it means for [AUDIENCE].

Use the actual figures from the sources. Do not use stock analogies.

27. Report custom prompt

Write a briefing document for [AUDIENCE] who has 10 minutes.

Sections:
1. What these sources establish (cited)
2. What they disagree about (cited both ways)
3. What they do not cover
4. Three decisions this evidence is good enough to support
5. Three decisions it is not

Every factual line carries a citation. Where sources conflict,
present both rather than resolving it.

28. Configure Chat custom style

Respond like a research assistant preparing a memo for a sceptical
reviewer. Lead with the evidence, keep hedging minimal but honest,
and never state something as established when only one source says it.

Why does NotebookLM refuse to answer, and is that a bug?

It is the feature. It is the only reason to prefer this tool over a general chatbot for evidence work.

Google's FAQ names three causes for a non-answer. Safety flags, where "the content of your source contains language that triggers a safety flag", including in historical contexts. Unclear phrasing, where retrieval cannot match your question to a passage. And the important one: "Information not in sources: Gemini Notebook answers questions based on the information provided in your uploaded sources. If the answer isn't in the source material, it won't provide a response."

The chat documentation adds the refusal string. Ask for something creative, Google's example being "rewrite the end of my short story", and "you may receive, 'Gemini Notebook can't answer this question' as a chat response."

There is one thing worth doing before you accept a refusal. Google's fix for vague questions is precision: "A specific question helps Gemini Notebook locate the most relevant pieces of information." Narrow the question, name the source, name the section, and try once more. If it still refuses, believe it.

What NotebookLM prompts still can't do

Being honest about this is more useful than another ten templates.

It will not bring in outside knowledge in chat, by default. No industry benchmarks, no definitions it was not given, no "compare this to standard practice". If you want that comparison, the standard practice has to become a source.

There are exceptions, and you should know them. Source discovery searches the web to find sources, and Deep Research, which Google describes as able to "automatically browse up to hundreds of websites", produces a report you can import. Both add sources to the notebook rather than answering from the open web. Separately, Google's chat documentation says Google AI Ultra and Pro subscribers on desktop get agentic chat that can take actions including "searching the web, running code, or creating downloadable files, charts, and images", and that "you can use this chat experience with or without sources." If you are on a paid tier, the closed-corpus guarantee is no longer automatic. Google labels these functions "experimental and in early development."

Citations are not always passage-level. Short sources get referenced as whole documents. Check before you rely on a locator.

It is not a writing tool. Creative rewriting gets refused. Drafting still belongs in ChatGPT or Claude, with NotebookLM as the fact-check pass. The workflow in how to summarize 50 papers without losing citations covers that handoff, and 30 AI prompts for literature review covers the reading stage that comes before you build the notebook.

Free-tier caps bite fast. Fifty chats a day and three Audio Overviews. Iterating on a prompt in the chat box burns the same budget as running it.

How do you keep source-based prompts reusable?

The 28 prompts above have a property that makes them worth saving: they are structural. "Claim inventory with per-source attribution" is the same prompt whether the notebook holds twelve papers on soil chemistry or twelve customer interviews. Only the bracketed slots change.

That is exactly the case for a prompt template library rather than a scratch file. Prompt Architects stores prompts with named variables you fill at use time, so [TOPIC] and [SOURCE A] become fields rather than find-and-replace. NotebookLM is on our extension's list of natively supported platforms, alongside ChatGPT, Gemini, Claude, Perplexity and the rest, so a saved prompt is a click away rather than a tab switch. The free plan includes 5 prompt enhancements per day, forever, per our FAQ; current paid pricing is on the pricing page.

Two habits matter more than the tool. Version the prompt when you change it, so a rerun three months later is comparable. And keep the corpus definition attached to the prompt, because a source-based prompt without its source list is not reproducible. If you want the general version of that argument, how to build a personal AI prompt library covers the structure, and what is prompt engineering covers the fundamentals these templates assume.

One last translation table, for the prompts you will be tempted to paste over from ChatGPT.

Sources. All Google product facts in this post come from Google's Gemini Notebook help centre and Google's own blog, accessed 26 August 2026: the chat documentation, FAQ, source types and limits, plan limits, Audio Overviews, Video Overviews, notebook creation and Studio outputs, privacy, and the rename announcement. Google marks its usage limits "subject to change", so check the page before quoting a number.