It only knows what you give it. That is the whole point.

Use it when all three are true

This filter saves the most time. If any one is false, a general assistant serves you better.

When to use something else

• Coding and open-ended reasoning. Not what this is. It is a librarian.
• Genuine web research. Deep Research exists here but is weaker than the dedicated deep-research modes in Gemini, ChatGPT and Claude. Do the research there, then bring the report in as a source.
• Anything beyond your sources. Grounding is a wall, not a preference. Ask it something your material does not cover and it declines rather than filling the gap. That is the feature working correctly.
• A first pass on a subject you know nothing about. You cannot curate sources for a field you cannot yet judge. Get oriented in a general assistant first, then build the notebook.

Three panels, worked left to right

The whole product is a pipeline and the layout tells you the order.

The mistake almost everyone makes is living in Chat. Chat is for questions you already know how to ask. Studio is for material whose shape you do not yet know.

Feeding it well

What it accepts

• PDF, Word, plain text and Markdown
• Google Docs, Slides and Sheets from Drive
• CSV files, and images with OCR — text inside diagrams and scans is readable
• Web pages by URL
• Public YouTube videos — it imports the transcript, not the video, and cites timestamped passages
• Audio files, including your own recordings

Size ceiling

Each source caps at roughly 500,000 words or 200 MB for local uploads, on every tier. A failed upload is almost always over that line, not a bug. Split the file.

Finding sources you do not have

Fast Research behaves like a search bar — over the Web, or over your own Drive. It returns candidates and you pick. Deep Research is an agent that reads across many sources and writes a cited report in the background, which you can then add as a source. Discover proposes around ten curated web sources on a topic.

For studying, Fast Research and Discover are the ones you want: they build the reading list and leave the judgement to you. Deep Research earns its place only on a genuine cold start in an unfamiliar field.

Curation rules that actually matter

• One notebook, one subject. Grounding is a strength only while sources are coherent. A notebook holding a distributed-systems paper and your insurance policy answers both worse.
• Prune, do not hoard. Nothing forces you to fill the source slots. A near-duplicate drags every summary toward whatever repeats itself.
• Deselect to scope. Individual sources can be toggled off. Ask the same question against source A alone, then B alone — the fastest way to find where two explanations of the same concept diverge.
• Add the difficult source too. A notebook of three friendly blog posts teaches you the friendly version. Include the original paper or specification even if you cannot read it yet; that is what you are here to fix.
• Style guides are sources. Upload a visual reference or style guideline and reference it in the prompt. Infographics, slides and reports will follow it.

Configure Chat is the highest-leverage control in the product

Behind the gear icon is a per-notebook instruction — a role and a goal, up to 5,000 characters. It frames every answer in that notebook and it survives the conversation. Almost nobody sets it. It is the difference between a summarizer and a tutor.

You are a demanding tutor. I am an experienced professional in an
adjacent technical field, so skip introductory framing and skip
analogies to things I already know well.

For every answer:
- Give me the mechanism, not the description of the mechanism.
- Where the sources present a trade-off, state both sides before
  any synthesis.
- Distinguish what the sources state from what they imply. Label
  implications as such.
- Name the gap explicitly when the sources do not cover something.
  Never fill it from general knowledge.
- End with one question that tests whether I actually understood
  the answer, not whether I can repeat it.
- If I state something incorrect, correct it directly before
  continuing. Do not soften it.

Never write an introduction, a recap of my question, or a closing
summary.

Your job is to make difficult material in these sources
comprehensible and, where possible, drawable.

For every concept I ask about:
- State the one thing that must be understood first, before
  anything else makes sense.
- Describe the concept as a structure: what the parts are, what
  flows between them, and what changes over time.
- Propose one diagram that would carry the explanation, and
  describe it precisely enough that someone could draw it: what
  is on each axis or in each box, what the arrows mean, what is
  deliberately left out.
- Name the most common misunderstanding and say what causes it.

Use plain language. No metaphors unless the sources use one.

How to ask

• Ask for the location, not just the answer. "Which source and section covers X" beats "tell me about X" when you intend to read the original anyway — which, when studying, you should.
• Ask across sources. Synthesis is its real advantage. "Where do these two explanations of the same idea actually differ" is a question no single document can answer.
• Ask for absences. "What does this treatment not cover that a complete one would" is a grounded question and a genuinely strong use.
• Ask it to grade you. Write your own explanation of a concept, paste it in, and ask which parts the sources do not support. This is the highest-value thing you can do in the Chat panel and almost nobody does it.
• Click the citations. Every claim carries a marker back to the exact passage. When you are learning, an unchecked answer is how a misunderstanding gets set in concrete.

Turning a hard concept into something you can look at

This is the workflow to actually learn. It uses four Studio outputs in a deliberate order, because each one fails differently and the failures are informative.

1. Mind Map — find the shape. Always first, before you read anything. It shows every topic in the sources and how they connect, and branches are clickable: click one and a grounded chat opens on that sub-topic. You learn what the material contains and, more usefully, which part everything else depends on.
2. Chat — force a diagram out of it. Using the explainer instruction above, ask for a described diagram of the concept. You are not asking it to draw. You are asking it to commit to a structure. A concept that cannot be described as boxes and arrows usually is not understood yet — by it or by you.
3. Infographic — make the artifact. Feed the diagram description back in as a steering prompt. One polished visual, grounded in the sources. Upload a style reference as a source if you want it to look like yours.
4. Video Overview — make it move. For anything with a sequence, a flow, or a state change, a static image lies. A narrated Explainer walks through it step by step. Slower to generate, but this is the format for protocols, algorithms, pipelines and processes.

Take [concept] as the sources describe it.

Give me a diagram specification, not a description:
- The elements: what each box or node represents, in the sources'
  own vocabulary.
- The relationships: what each arrow means. Be specific about
  direction, and about whether it is data, control, causation,
  or containment.
- The dimension: if something changes over time or scale, say
  which axis carries it.
- What to leave out: name the details that are true but would
  make the diagram unreadable, and say why they can be dropped
  without making the picture wrong.
- The one thing the diagram will inevitably misrepresent.

Build a single explanatory visual for [concept], aimed at someone
technically literate who has never met this specific idea.

Requirements:
- One central structure that carries the explanation. Not a list
  of facts arranged prettily.
- Label every element with the term the sources actually use, so
  the visual maps onto the reading.
- Include the numbers and limits the sources state, exactly.
- Include one small panel showing the common misunderstanding and
  why it is wrong.
- No decorative icons that do not encode information.

Format: Explainer. Audience: a technically experienced person who
is new to this specific subject. Skip the definitions of adjacent
concepts I already know.

Structure the video as a walkthrough of the mechanism in order:
what happens first, what state that leaves things in, what happens
next, and where it can fail. Spend the most time on the step the
sources treat as the hardest.

Do not open with why the topic matters and do not close with an
optimistic summary.

The study loop

Visuals get you comprehension. This gets you retention. Run it after the visual pass, not instead of it.

1. Explain it back, in writing. Write your own explanation of the concept without looking at the sources. Two paragraphs is enough.
2. Have it marked. Paste your explanation into Chat and ask which parts the sources do not support, which are technically true but misleading, and what you omitted. Grounding means it marks against the material, not against a general impression of the subject.
3. Generate a quiz on what you got wrong. Multiple choice, from your sources. Do not generate a quiz on the whole notebook — generate one on the parts you just failed.
4. Build a flashcard deck for the parts that are pure memory. Terminology, limits, sequences, API surfaces. Set the difficulty, add a custom prompt to target the hardest material, and use the mobile app. Progress is saved.
5. Come back and ask the same first question a week later. Ask it cold, before re-reading. The gap between your answer and the cited one is your actual retention.

Below is my own explanation of [concept], written from memory.
Assess it strictly against the sources only.

Tell me:
1. Which claims the sources directly support.
2. Which claims are unsupported by the sources — separate the
   ones that are merely absent from the ones that are contradicted.
3. Which claims are technically true but would mislead someone
   reading them.
4. What I left out that the sources treat as essential.
5. The single most important thing I appear to have misunderstood,
   and the passage that corrects it.

Do not rewrite my explanation. Do not soften anything.

MY EXPLANATION:
[paste]

Generate cards only for material that has to be recalled exactly:
terms and their precise definitions, stated limits and figures,
ordered sequences, and distinctions between concepts that are
easily confused.

Do not create cards for anything that can be reasoned out from
first principles. Where two concepts are commonly mixed up, make
one card that asks for the difference rather than two separate
definition cards.

Which output is worth generating

Nine outputs, wildly unequal value. Ranked for learning and explaining, which is not the same ranking you would get for office work.

Software engineering, without touching your codebase

You said no coding and no work material. Both still leave the largest category of value on the table: understanding the field. All of the below run on public documents.

Learning a technology from primary sources

Sources: the official documentation, the original design document or RFC, one deep conference talk transcript, one critical blog post that disagrees with the official line. Mind map first, then work the branches with the tutor instruction. This is how you learn a thing properly instead of accumulating tutorial-shaped knowledge.

Reading papers you keep bouncing off

Distributed systems, consensus, database internals, memory models — the canonical papers are public and famously hard. Put one paper plus two explanations of it in a notebook. Ask which parts of the explanations the paper does not actually support. The gap between a paper and its popular summary is where the real understanding is.

Understanding a specification properly

A protocol spec, a language memory model, a serialization format. Ask for the state machine, the failure modes, and the parts that are explicitly implementation-defined. Then generate a Video Overview of the message flow. Specifications are exactly the material this tool handles best: precise, self-contained, and painful to read linearly.

Certification and exam preparation

Sources: the official syllabus, the vendor's own study material, your notes. The syllabus is the highest-value source because it tells the notebook what is actually examinable. Generate flashcards scoped to it, quiz yourself, and re-quiz on failures. Progress saves, so the mobile app becomes usable dead time.

Interview preparation

Sources: the job description, the company's public engineering blog and talks, your own CV. Ask which stated requirements your CV does not evidence at all — the useful direction, not the flattering one. Ask what the engineering blog reveals about how they actually build things, and generate questions to ask them. One notebook per search, not per application.

Keeping up without drowning

A rolling notebook of release notes, RFCs, and long-form posts you saved and never read. Once a month: what changed materially, what is genuinely new versus renamed, and which of these do I actually need to care about. Then an Audio Overview of the residue for the commute.

Teaching someone else

The fastest way to find the holes in your own understanding. Build the notebook, generate the deck for its proposed narrative, then rebuild the explanation yourself. The proposed structure saves the blank-page phase; the rebuilding is where you learn.

Health — comprehension, not diagnosis

This is a legitimate and genuinely useful category, with one hard boundary: the tool helps you understand documents and ask better questions. It does not tell you what is wrong with you, and it has no context beyond what you upload.

Understanding your own results over time

Upload the same kind of report across several years. The strongest question is comparative: what changed materially between them, what stayed flat, and which values the reports themselves flag as outside their stated reference ranges. It is a reading aid for documents you already own and struggle to interpret.

These sources are the same kind of report from different dates.

1. Build a table of every measured value, one column per date.
2. Mark which values moved meaningfully between dates, using each
   report's own reference range as the definition of meaningful.
3. Separate values the reports themselves flag from values that
   merely changed.
4. List anything measured in one report and not in the others.
5. Write the questions these differences raise — phrased as
   questions to ask a doctor, not as conclusions.

Quote figures and units exactly. Do not interpret, do not
reassure, and do not suggest what anything means.

Preparing for an appointment

Appointments are short and you forget things. Upload your documents, ask what a well-prepared patient would want clarified, and take the list with you. This is the single best health use of the tool: it converts a pile of paper into a short list of questions.

From these documents, produce a one-page brief I can take to an
appointment:

- A timeline of what happened and when, from the documents only.
- Terms used in these documents that I should ask to have
  explained, with the sentence each appears in.
- Points where two documents say different things.
- Anything a document says should be followed up, checked again,
  or repeated after a period — with the stated timeframe.
- Questions worth asking, ordered by importance.

No advice, no interpretation, no reassurance. Questions only.

Reading the material you were handed

Medication leaflets, procedure information sheets, insurance coverage documents. Ask what a document states plainly, what it states in conditional or hedged language, and what it does not address at all. Good use of grounding: you want to know what this paper says, not what the internet says.

Understanding a subject properly, in general

Sleep, training physiology, nutrition science, how a class of medication works. Build the notebook from proper sources — review papers, official health service material, textbook chapters — and run the same visual workflow you would use for a technical concept. The mechanism is a mechanism; boxes and arrows work fine. This category is about the science in general, kept separate from your own records.

Fitness and training material

Upload the programme you are following, the source it came from, and one critique of the approach. Ask what it claims, on what evidence, and what it deliberately trades away. Useful precisely because training material is written persuasively and grounding strips the persuasion back to what is actually stated.

Life admin and curiosity

Documents you are obliged to understand

Insurance policies, rental and purchase contracts, warranty terms, terms of service. Ask what obligations exist, on whom, by when, and what the document says happens if they are missed — quoting the governing sentence. Then ask what a document of this type would normally address that this one does not. Verify before acting; you are liable, not the notebook.

A subject you want to learn for its own sake

History, a language, music theory, cooking technique, anything. The workflow does not change. Build sources, mind map, tutor instruction, visual output, flashcards for what must be memorised. The novelty here is that the material can be a lecture series or a book you own rather than documentation.

Long-form content you never get to

Saved articles, newsletters, podcast transcripts, conference talks. Batch them into a notebook and ask what is worth your time and what can be summarised safely. Audio Overview the residue. This is the honest use case for the podcast feature.

Prompt library

Chat panel unless noted. All of them lean on grounding — they ask for things only a source-bound model can honestly answer.

Orientation — a notebook you have not read

List every distinct concept covered across these sources. For each:
which source covers it, roughly how much space it gets, and whether
another source contradicts it. Then tell me which concept everything
else depends on, so I know where to start. No summaries yet.

Build a reading order through these sources for someone learning
this subject from scratch. For each step: what to read, what it
assumes you already know, and what question you should be able to
answer before moving on. Mark anything that can be skipped on a
first pass.

Understanding

Explain [concept] three times, using only these sources:

1. The one-sentence version that is still true.
2. The working version — enough to reason about it correctly,
   including the parts that are usually simplified away.
3. The precise version, in the sources' own terminology, with
   the conditions and limits they state.

After the three, name what level 1 sacrifices to stay short.

I keep failing to understand [concept]. Based on these sources,
what makes it hard? Specifically:
- Prerequisite ideas people are assumed to have and usually don't.
- Terminology that means something different here than elsewhere.
- The common wrong mental model, and the point at which it breaks.
- The one thing that, once it clicks, makes the rest follow.

Identify every point where two or more sources make incompatible
claims. For each: state both claims with citations, name the
source of each, and say what evidence would settle it. Ignore
differences that are only wording.

Given the subject of these sources, list what a complete treatment
would normally cover that is absent here. Separate: (a) topics
mentioned but not developed, (b) topics never mentioned at all,
(c) claims made with no supporting evidence in any source.

Testing yourself

Ask me one question at a time about [topic], grounded in these
sources. Start easy and escalate based on my answers. After each
answer: tell me what was right, what was wrong, and cite the
passage that settles it — then ask the next question. Do not move
on while I am still wrong about something. Do not give me the
answer unless I ask for it.

Based on the questions I have asked in this notebook so far, what
parts of this material have I not engaged with at all? Which of
those does the rest depend on? Give me one question on each that
would show whether I actually understand it.

Extraction

Extract every quantitative claim in these sources — figures,
limits, thresholds, dates, versions, percentages. Present as a
table: value, what it measures, source, and whether the source
gives a method or basis for it. Mark any figure appearing in only
one source.

Find every passage in these transcripts where the speaker discusses
[topic]. Give the timestamp, a one-line paraphrase, and whether the
statement is a firm claim, a hedge, or speculation. Do not merge
separate mentions.

Comparison over time

These sources are the same kind of document from different periods.
List everything that changed materially between them, ordered by
how much it matters. Ignore formatting and boilerplate. Flag
anything present earlier and now removed.

Stress-testing

Using only these sources, build the strongest case against the
central claim of [source name]. Cite the passages you rely on.
Then list which of its claims the other sources actually support,
so I can tell disagreement from noise.

I have thirty minutes and cannot read all of this. Tell me which
specific sections of which sources I must read myself, and why
each one cannot be summarised safely. Summarise everything else
in one paragraph per source.

Plans and limits

Paid tiers mostly raise meters rather than unlock different core features. The exceptions are Cinematic Video Overviews, native code execution and the newest agentic features, which arrive on higher tiers first. For the study workflow above, the free tier's three video and three audio generations a day is the constraint you will hit first.

Gotchas

• Grounding is not accuracy. It faithfully reports what your sources say. If a source is wrong, so is the answer — with a citation attached, which makes it more convincing, not less wrong.
• YouTube means transcript. No captions, no source. Nothing visual in the video is available to it — a whiteboard talk loses the whiteboard.
• Chat history colours later answers. Long threads drift. Clear them deliberately.
• Generated visuals can be confidently wrong. An infographic is a compression, and a compression makes editorial choices you did not approve. Check the labels against the citations before you commit a diagram to memory.
• Slide export is not editable. Images in a PPTX shell. Plan around it.
• Two products, similar names. Gemini Notebook is the standalone app. "Notebooks in Gemini" is a section inside the Gemini chat app. They sync, but the standalone app has the full feature set.
• Browser extensions broke on the address move. If a NotebookLM extension stopped working in late July 2026, update it — the app moved to notebook.google.com.

What changed, briefly