The NotebookLM Prompt Library

33 prompts, organized by the job they do. Enter hub password to continue.

That's not right. Try again.

Working draft · 29 July 2026

The NotebookLM Prompt Library

Every other prompt collection on the internet is transcribed. Nobody shows you what came back, nobody tells you which prompts break on a small notebook, and nobody writes for a library holding fifty sources or more. This one is built to fix all three, and it publishes its own failures.

Status: draft. The prompts are written and the reasoning is done. None have been run yet, so every one below is stamped UNTESTED. Nothing goes public until it has been verified against a real notebook and the actual output saved as evidence.

33prompts written

0verified so far

8jobs they cover

7popular prompts to skip

331notebooks to test against

How to read this

Each prompt carries the job it does, what it needs to work, and the way it fails. That last field is the one worth reading. A synthesis prompt aimed at a two-source notebook will invent a pattern and state it confidently, and knowing that in advance is worth more than the prompt itself.

Prompts marked as going in a customize box belong in the Studio panel before you generate an audio overview or a deck, not in the chat.

Interrogate a single source

For when one paper, report, or transcript is in front of you and the question is "can I trust this, and what's actually in it?"

1.1 The Three-Sentence Grounded Summary UNTESTED

Triage. Deciding whether a source deserves an hour of your time.

Summarize this source in exactly three sentences: (1) the central claim, (2) the single strongest piece of evidence offered for it, (3) the implication that matters most to a practitioner. After each sentence, cite the specific section it came from. If any of the three cannot be supported from the source, say so instead of filling the gap.

Copy the text above.

Needs: 1 source.
Fails when: The source is a landing page, abstract, or paywall stub. You'll get a summary of the access barrier, not the research.
Then run: 1.3 if it passes triage.

1.2 Evidence Quality Audit UNTESTED

Deciding whether to cite something.

Audit the evidence in this source. For every substantive claim: quote the claim, state what evidence is offered (data, citation, expert assertion, or nothing), and label it SUPPORTED, PARTIALLY SUPPORTED, or ASSERTED. List the ASSERTED claims separately at the end. Do not soften the assessment. Finish with an overall credibility call of HIGH, MEDIUM, or LOW and the single biggest reason for it.

Copy the text above.

Needs: 1 full-text source.
Fails when: Run on an abstract. Abstracts are all assertion, so everything returns ASSERTED and the audit is meaningless.
Then run: 1.4, then 3.1 if you have competing sources.

1.3 Methodology Extractor with Rigor Score UNTESTED

Peer-review posture. Separating "a study found" from "a good study found."

Extract the methodology only: design, sample size and characteristics, what was measured and how, analysis method, funding source, and stated limitations. Then rate rigor 1 to 5 and justify the number in one sentence. Then list what the authors did NOT control for that a skeptical reviewer would raise.

Copy the text above.

Needs: 1 empirical source (not a review or an opinion piece).
Fails when: The source is a narrative review or a blog post. There's no methodology to extract and it will pad.
Then run: 3.1

1.4 The Load-Bearing Claim Finder UNTESTED

Finding the single point of failure in an argument before you build on it.

Identify the claims in this source that everything else depends on: if this one claim were wrong, what conclusions collapse? List up to five, ranked by how much of the argument rests on each. For each, state the evidence behind it and how fragile that evidence is.

Copy the text above.

Needs: 1 source with an actual argument.
Fails when: Purely descriptive sources (data tables, reference material). There's no load-bearing structure to find.
Then run: 4.2

1.5 Quote Mining with Provenance UNTESTED

Pulling material for writing without misquoting.

Extract the 10 most quotable statements: the ones that are surprising, unusually precise, or that a reader would repeat. For each give the exact verbatim quote, the section it appears in, and one sentence on why it lands. Do not paraphrase. If a statement is only interesting with context, include the minimum context in brackets.

Copy the text above.

Needs: 1 source.
Fails when: Nothing. This is one of the most reliable prompts in the collection.
Then run: 6.1

Synthesize a whole corpus

This is the category the rest of the internet skips, because most people never build a notebook past a handful of sources. These are written for 20 to 100+ source notebooks.

2.1 The Consensus and Contested Map UNTESTED

The first thing to run on any mature notebook. It tells you what your library actually believes.

Across ALL sources in this notebook, produce two lists. CONSENSUS: claims that three or more independent sources support, with the source names. CONTESTED: claims where sources genuinely disagree, with who is on each side. Then a third list, THIN: claims that appear repeatedly but trace back to only one original source. Be strict about the difference between many sources agreeing and many sources citing the same original.

Copy the text above.

Needs: 8+ sources. Below that the three buckets collapse into each other.
Fails when: Your notebook is topically scattered. One notebook, one topic, or the consensus list is nonsense.
Then run: 3.1 on the contested items, 4.1 on the thin ones.

2.2 The Citation Cascade Detector UNTESTED

Catching the thing that ruins health and nutrition research: forty papers citing one 1994 study, drifting a little each time.

Find claims in this notebook that appear across multiple sources but trace back to a single original study or dataset. For each: the claim, every source repeating it, and the original it traces to. Flag any claim where the repetitions have drifted from what the original actually said.

Copy the text above.

Needs: 15+ sources on one topic.
Fails when: Sources lack reference lists (transcripts, blog posts). It can spot repetition but not the trace.
Then run: 1.3 on the original study.

2.3 Corpus Coverage Map UNTESTED

Knowing what to go find next. Run it before another research session.

Map what this notebook covers. Produce a table of the major sub-topics, how many sources address each, and the depth of coverage (deep, moderate, passing mention). Then list the sub-topics a well-read expert would expect to see here that are absent or thin. Rank the gaps by how much they weaken the collection.

Copy the text above.

Needs: 10+ sources.
Fails when: Very new notebooks. Everything reads as a gap.
Then run: Use the gap list as your next search, then 2.1 again after adding sources.

2.4 The Evolution Timeline UNTESTED

Not embarrassing yourself by citing a finding that was overturned in 2019.

Order the sources in this notebook by publication date and trace how thinking on [TOPIC] changed. What was believed early that was later revised or abandoned? What has held up unchanged? Where does the most recent work point? Cite the source and year for each shift. If publication dates are unavailable for some sources, say which ones and exclude them rather than guessing.

Copy the text above.

Needs: 10+ sources spanning several years.
Fails when: Undated sources, which is common with web clips and PDFs missing metadata. The refusal clause is what keeps it honest.
Then run: 2.1

2.5 The Steelman of the Minority View UNTESTED

Guarding against a notebook that only confirms what you already believed when you built it.

Identify the minority position in this notebook: the view held by the fewest sources but not obviously wrong. Build the strongest possible case for it using only what's in my sources. Then state precisely what evidence would be needed to make it the majority view, and whether anything in my sources already points that way.

Copy the text above.

Needs: 10+ sources with genuine variety.
Fails when: You built the notebook from one side of a debate. It will steelman a position no source holds, and drift off-grounding.
Then run: 4.3

Find the conflict

Summaries hide disagreement. These prompts are built to drag it into the open.

3.1 The Irreconcilable Conflict Hunt UNTESTED

The flagship corpus prompt. Run it on any notebook you plan to make decisions from.

Across all sources, find places where two or more make claims that cannot both be true. For each conflict: (1) state the disagreement in one sentence, (2) name the sources on each side, (3) quote the conflicting line from each, (4) say which side has the stronger study design and why. Rank by how much the conflict would change what a reader should actually do. Skip anything where sources merely differ in emphasis rather than genuinely contradicting.

Copy the text above.

Needs: 8+ sources.
Fails when: Small or one-sided notebooks. It will manufacture conflict to satisfy the request. If output feels strained, that's the signal.
Then run: 3.2 on the top conflict.

3.2 The Resolution Path UNTESTED

The follow-up that turns "experts disagree" into "here's why, and here's what would settle it."

Take this disagreement: [PASTE THE CONFLICT]. Using only my sources, explain the most likely reason the two sides disagree: different populations, different measurement, different timeframe, different funding, or genuine unresolved science. Then state what single study design would settle it, and whether anything in my sources comes close.

Copy the text above.

Needs: Run 3.1 first.
Fails when: Nothing significant. Reliable follow-up.
Then run: 5.1

3.3 Source-to-Source Cross-Examination UNTESTED

Understanding a debate properly rather than picking the side you already liked.

Put [SOURCE A] and [SOURCE B] in direct cross-examination. What would A's authors say is the fatal flaw in B? What would B's authors say back? Where would both concede the other has a point? Use only claims actually present in each source. Do not invent positions the authors did not take.

Copy the text above.

Needs: 2 substantive sources that genuinely disagree.
Fails when: The sources don't actually conflict. You'll get a polite non-debate.
Then run: 3.2

Audit yourself

The most valuable and least used category. These point the tool at your own thinking rather than the material.

4.1 The Single-Source Dependency Audit UNTESTED

Finding where your confidence is unearned. Genuinely uncomfortable, which is why it's useful.

List every significant claim in this notebook that rests on exactly one source. For each: the claim, the lone source, and what would break if that source turned out to be wrong or retracted. Rank by how much of my overall understanding depends on it.

Copy the text above.

Needs: 10+ sources.
Fails when: Nothing. One of the strongest prompts here.
Then run: Go find corroborating sources for the top three, then re-run.

4.2 Bias and Funding Sweep UNTESTED

Running before you build anything on a body of research, especially in supplements, nutrition, and pharma.

Sweep every source for things that should affect how much I trust it: funding source, author affiliation, publication venue, declared conflicts, and whether the framing favors a particular conclusion. Produce a table: source, potential bias, severity (low/medium/high), and whether its conclusions align suspiciously well with its funder's interest. Say explicitly where this information is not disclosed in the source.

Copy the text above.

Needs: 5+ sources.
Fails when: Sources are clipped web pages missing front matter. Funding info usually lives in material that gets stripped.
Then run: Re-run 2.1 weighting the low-bias sources.

4.3 What Would Change My Mind UNTESTED

Before publishing, before a decision, before you commit publicly to a view.

I currently believe [YOUR POSITION] based on these sources. Using only my sources, identify: (1) the evidence that most strongly cuts against my position, (2) the weakest link in my supporting evidence, (3) what I would need to observe to abandon this position, and (4) whether anything already in my sources should be worrying me more than it is. Be blunt.

Copy the text above.

Needs: 8+ sources, ideally not all agreeing.
Fails when: One-sided notebooks return a weak answer. That weakness is itself a finding: your corpus is an echo chamber.
Then run: 2.5

4.4 The Overclaim Check UNTESTED

The last pass before publishing anything that cites research. This is the highest-value prompt in the entire collection for anyone who writes publicly.

Here is something I wrote: [PASTE YOUR DRAFT]. Check every factual claim in it against my sources. For each: SUPPORTED (with the citation), OVERSTATED (source says something weaker, quote it), or UNSUPPORTED (not in my sources at all). List the OVERSTATED and UNSUPPORTED items first. Do not comment on style.

Copy the text above.

Needs: A notebook whose sources actually cover the draft's claims.
Fails when: Your draft ranges wider than your notebook. Everything off-corpus returns UNSUPPORTED, which reads as alarming but only means the source isn't in the notebook.
Then run: Fix the draft, re-run until clean.

Turn research into decisions

5.1 The Evidence-Weighted Decision Framework UNTESTED

Turning a pile of reading into an actual choice.

Build a decision framework for [DECISION]. Produce: the criteria that actually matter, a weight for each and why, the realistic options, and a score per option per criterion. For every score, cite the source behind it and note whether the evidence is strong, moderate, or thin. Where my sources cannot support a score, mark it UNKNOWN rather than estimating.

Copy the text above.

Needs: 10+ sources relevant to the decision.
Fails when: The decision depends on personal factors your sources know nothing about. Expect a lot of UNKNOWN, which is the honest answer.
Then run: 5.2

5.2 Scenario Planning with Signals UNTESTED

Planning under uncertainty without pretending to certainty.

Generate three scenarios for [TOPIC]: most likely, best case, worst case. For each: the key drivers, what my sources say about how plausible it is, the earliest observable signals that it's happening, and what I should do now to prepare. Ground every driver in a source. Mark any scenario element that is your inference rather than something my sources support.

Copy the text above.

Needs: 10+ sources.
Fails when: Purely historical corpora with nothing forward-looking.
Then run: 5.3

5.3 The Protocol Builder UNTESTED

The end of the line for health, training, and process research: converting reading into action.

Turn what's in this notebook into a specific protocol for [PERSON OR SITUATION]: what to do, how much, how often, in what order, and how to tell if it's working. Cite the source and the strength of evidence for every element. Separate the elements with strong support from the ones that are plausible but thin. Explicitly list what my sources say NOT to do, and anything requiring professional supervision.

Copy the text above.

Needs: 15+ sources for anything you'd actually follow.
Fails when: Thin corpora produce confident protocols from weak evidence, which is the most dangerous failure in this whole collection. Always run 4.1 first.
Then run: 4.1, then 4.3 against the finished protocol.

Turn research into writing

Most repurposing prompts online are general-purpose content prompts wearing a NotebookLM badge. These are the ones that genuinely need grounding.

6.1 The Citation-Locked Draft UNTESTED

Writing that will be scrutinized, where every claim has to survive a reader who checks.

Draft [FORMAT] about [ANGLE] for [AUDIENCE], using only what's in my sources. Every factual claim must carry an inline citation to the source it came from. Where the piece needs a claim my sources don't support, insert [NEEDS SOURCE] rather than writing it. Do not smooth over the gaps.

Copy the text above.

Needs: A corpus that actually covers the piece.
Fails when: Nothing. Genuinely strong, and the gap markers are more useful than the prose.
Then run: 4.4 on the finished draft.

6.2 The Surprise Extractor UNTESTED

Finding the angle. The reason to read your piece instead of the twenty others.

I'm writing about [TOPIC]. From my sources, what would genuinely surprise a well-informed reader who already knows the basics? Not the standard talking points. Give me the findings that cut against conventional wisdom, the numbers that are bigger or smaller than expected, and the caveats that usually get dropped. Quote each one and cite it.

Copy the text above.

Needs: 10+ sources.
Fails when: Corpora built from popular articles, which all repeat the same talking points.
Then run: 6.1

6.3 The Objection Pre-empter UNTESTED

Publishing without getting taken apart in the comments.

Here is my draft: [PASTE]. Using my sources, list the strongest objections an informed critic would raise, ranked by how damaging each is. For each: the objection, whether my sources can answer it, and the specific citation that does. Flag any objection my sources cannot answer.

Copy the text above.

Needs: 8+ sources.
Fails when: Nothing significant.
Then run: 4.4

Steering Studio artifacts

Studio outputs are generated from a customization box, not the chat. These are written for that box. Coverage here goes well past the slide-deck-only collections.

7.1 Audio Overview: the Practitioner Brief UNTESTED

The default Audio Overview wastes 90 seconds on context-setting and stays general. This fixes both.

Audience: [WHO], who already knows the basics. Skip all background and definitions. Open with the single most actionable finding in the sources, then spend the bulk of the time on what to actually do and the evidence behind it. Cover the caveats honestly rather than reassuringly. Do not summarize the sources one by one. End on the most important open question, not a recap.

Copy the text above into the Audio Overview customize box.

Needs: 2+ substantive sources.
Fails when: Sources are paywall stubs. You get a podcast about the paywall. This is documented and real.
Then run: 7.2

7.2 Audio Overview: the Disagreement Episode UNTESTED

Far more interesting listening than a consensus summary, and it teaches the shape of the debate.

Build the conversation around where the sources disagree, not where they agree. Have one host present the strongest version of each position and the other push back using the actual evidence in the sources. Do not resolve disagreements that the sources leave open. Name the specific studies. Audience: [WHO].

Copy the text above customize box.

Needs: 8+ sources with genuine disagreement. Run 3.1 first to confirm the disagreement exists.
Fails when: One-sided corpora produce a fake debate.
Then run: 3.1 first.

7.3 Slide Deck: the Evidence Deck UNTESTED

Presenting research without the audience having to take your word for anything.

Generate a deck for [AUDIENCE] making the case for [POSITION]. Use the Detailed format. One claim per slide, with the supporting evidence and its source named on the same slide. Include one slide near the end covering the strongest counter-evidence in the sources. Maximum [N] slides. No slide without a citation.

Copy the text above.

Needs: 8+ sources.
Fails when: Unstructured sources (raw transcripts) produce weak slide breaks. Structured sources with headings work far better.
Then run: 7.4

7.4 Infographic: the One-Number Story UNTESTED

Infographics default to busy and decorative. This forces one message.

Build the infographic around the single most important number in these sources. Lead with that number at the largest size. Support it with at most four secondary data points, each cited. Include one element showing the uncertainty or range rather than presenting the number as exact. No decorative elements that carry no information.

Copy the text above.

Needs: Sources containing actual quantitative data.
Fails when: Qualitative corpora. There's no number to lead with.
Then run: none

7.5 Flashcards: the Application Set UNTESTED

Flashcards default to definitions, which is the least useful thing to memorise.

Generate flashcards that test application rather than recall. Instead of "what is X", ask what would happen in a specific situation, or which approach fits a described case, or what the evidence would predict. Each card's answer must cite the source. Include a handful of cards on the limits of the evidence, not just the findings.

Copy the text above.

Needs: 5+ sources.
Fails when: Nothing significant.
Then run: none

7.6 Data Table: the Comparison Grid UNTESTED

The single fastest way to see the shape of a literature. This is what a systematic review's evidence table costs a researcher days to build.

Build a table comparing every study in these sources on: study design, sample size, population studied, intervention or exposure, primary outcome, effect size, and stated limitations. One row per study. Where a source does not report a field, write "not reported" rather than leaving it blank or inferring it.

Copy the text above.

Needs: 8+ empirical sources.
Fails when: Mixed corpora with reviews and opinion pieces produce mostly "not reported" rows.
Then run: 2.1

7.7 Mind Map: the Argument Structure UNTESTED

Mind maps default to a topic taxonomy, which tells you nothing you didn't know. Structuring by argument shows you where the support is thin.

Structure the mind map by argument rather than by topic: the central claim at the center, the main supporting lines of evidence as primary branches, and the specific studies as leaves. Add a separate branch for unresolved questions and one for counter-evidence.

Copy the text above.

Needs: 8+ sources.
Fails when: Corpora with no central argument.
Then run: 4.1

Persistent setup

Configured once per notebook, then every later answer improves. Almost nobody does this, and it has the highest return of anything here.

8.1 The Configure Chat Standing Instruction UNTESTED

Set once per notebook. Every subsequent answer inherits it.

You are helping me [GOAL]. Always cite the specific source for every factual claim. When my sources disagree, say so explicitly rather than presenting one view. When you cannot support something from my sources, say "not in your sources" instead of answering from general knowledge. Prioritise [WHAT MATTERS] over [WHAT DOESN'T]. My audience is [WHO]. Prefer specifics, numbers, and quotes over general statements.

Copy the text above into Configure Chat, not the message box.

Needs: Any notebook.
Fails when: Nothing. Highest return per minute in the collection.
Then run: Everything else gets better afterwards.

8.2 The Voice Source UNTESTED

Getting drafts that don't need a full rewrite.

Create a document describing how I write: sentence length, vocabulary, what I avoid, my typical structure, and two or three paragraphs of my actual writing as examples. Add it as a source. Then in chat: "Match the voice described in my voice document when drafting. Substance from the research sources, style from the voice document."

Copy the text above as a text source added to the notebook, not a chat message.

Needs: A written sample of your own voice.
Fails when: Thin voice documents. Two paragraphs of real writing beat a page of adjectives.
Then run: 6.1

8.3 The Note-to-Source Compounding Loop UNTESTED

Turning a notebook into something that gets smarter with use.

Run any high-value synthesis prompt. Save the answer as a Note. Convert the Note into a Source. Every future question can now build on that synthesis instead of re-deriving it.

Copy the text above workflow, not a single prompt.

Needs: Any notebook.
Fails when: Promoting weak answers. Anything promoted becomes permanent influence, so promote verified output only.
Then run: Re-run 2.1 afterwards and watch it improve.

The anti-library

Popular prompts worth skipping here, and what to run instead. This exists because no other collection is willing to tell you a prompt is bad.

Popular prompt	Why it wastes NotebookLM	Use instead
"Generate 10 social media hooks"	Pure generation. Grounding contributes nothing, and NotebookLM's writing is blander than a general model's.	Write hooks elsewhere. Use 6.2 to find the surprising fact worth hooking on.
"Summarize this document"	The single weakest prompt in circulation. Produces generic overview and wastes the citation architecture.	1.1
"Act as a world-class expert in X"	Persona prompts work by pulling on a model's general knowledge, which is exactly what grounding suppresses.	Skip the persona, specify the output.
"Write a 1,500-word blog post"	Long-form generation from a grounded model produces stiff prose.	6.1 for the citation-locked skeleton, then write it yourself.
"Generate a content calendar"	Nothing to do with your sources. A scheduling task.	Anything else.
"Explain like I'm 5"	Fine, but it works identically in any chatbot. Not a reason to be here.	Use it, but don't credit the tool.
"Give me everything about X"	Maximum-length prompts return padding. Grounded models pad by restating.	Two or three narrow prompts chained.

What works at what size

Most guides are written by people whose notebooks hold four PDFs. Prompts behave differently as a corpus grows, and the most common mistake is asking a broad question of a sixty-source notebook, getting a vague answer, and concluding the tool is weak. Narrow the source selection first.

Notebook size	What works	What breaks
1 to 3 sources	Category 1 only	Every synthesis prompt. They will manufacture patterns from nothing.
4 to 10 sources	Category 1, light Category 3	Corpus mapping, citation cascades, timelines
10 to 30 sources	Almost everything	Nothing much
30 to 100+ sources	The whole collection, and this is where 2.x and 7.6 become genuinely irreplaceable	Answers get vaguer as the corpus broadens. Use the source checkbox selector to narrow to the relevant 5 to 10 sources for specific questions.

Generated from the canonical library so the two cannot drift. Testing evidence will appear against each prompt as it is verified.