Gemini Notebook Prompts That Work Under the New Usage Limits

Gemini Notebook only answers from the sources you upload, and since 2 September 2026 the complexity of your prompt draws down a compute budget. Both facts make the long, rule-stacked prompts written for older models a bad trade: you pay for text that either restates a built-in guardrail or actively degrades the answer. What follows is two settings, five prompts, and three jobs that are buttons rather than typing.

New usage limits

Google replaced Gemini Notebook’s fixed daily caps with compute-based usage limits. The old system counted generations: a set number of chat queries and audio overviews per day. The new one works like a budget, and Google is explicit about what draws it down — prompt complexity, chat length, number of sources, and which features you use.

The budget refreshes every five hours rather than at midnight, with a weekly ceiling. Your remaining allowance appears below the chat and under Settings. Studio generations carry a bar estimating what they will cost, and a “Generate later” option defers expensive outputs like Video Overviews until the limit resets. Allowances scale with your Google AI tier, roughly doubling at each step.

Two consequences follow. A prompt that stacks four jobs and eleven prohibitions costs more than one that states a job. A long back-and-forth in a single chat gets progressively more expensive as the transcript grows, which matters for the one prompt below designed to run for many turns.

Prompt length is now a cost, not a signal of rigour.

Two settings that do the work of half a prompt

Configure Chat

In the Chat panel, Configure Chat sets a conversational style and a response length that persist across turns. The styles are Default, Learning Guide, and Custom. Custom takes a style or a role in a sentence — Google’s own examples are “Respond like a PhD student” and “Pretend to be a role-playing game host.” Response length is Default, Longer, or Shorter.

This is where a role belongs. Typing “act as a lead research analyst” at the top of eight separate prompts pays for the same instruction eight times, and it puts the persona in the wrong place: Gemini 3 treats an assigned persona seriously enough that it will sometimes ignore other instructions to stay in character. Set it once, keep it short, and leave the prompt for the job.

Learning Guide handles the layered-explanation use case on its own — the “explain this at three levels of depth” prompt that circulates for older models is mostly a hand-rolled version of this setting.

Source checkboxes

Every source in the panel has a checkbox that includes or excludes it from the model’s answer. This is a sharper instrument than describing sources inside a prompt. If you want to know whether two papers actually disagree, check those two and nothing else, rather than writing a paragraph asking the model to compare “Source A” with “Source B” while eighteen other documents sit in context.

Under compute limits, this is also the cheapest lever you have, since the number of sources in play is one of the inputs to the cost.

Four rules from Google’s current prompting guidance

Google’s Gemini 3 prompting guide documents specific failure modes, and the widely circulated Notebook prompt templates walk into most of them.

• Be concise. Gemini 3 responds best to direct instructions and may over-analyze the verbose prompt-engineering techniques that helped 2.x models. Prompts built in 2024 and 2025 read as overkill now. Delete first, add detail back only if the output is wrong.
• Put constraints last. On complex requests, the model may drop negative, formatting, and quantitative constraints if they appear early in the prompt. Put context and task first, and restrictions as the final line.
• Avoid blanket negatives. An open-ended “do not infer” or “do not guess” makes the model over-index on the instruction and fail at basic arithmetic and at connecting information across parts of a document. Google’s recommended fix is to say what the model should use rather than what it must not do.
• One prompt, one job. Google’s own troubleshooting checklist flags prompts that ask for several distinct cognitive actions in a single pass — summarise, extract, translate, draft — and tells you to split them. Numbered templates that map the terrain, cite the evidence, sequence the learning path and name the gaps are doing exactly this.

Five prompts worth the compute

Each does one job. Replace [TOPIC] with your subject. Scope the sources with the checkboxes before you run any of them.

1. Map the terrain

Map [TOPIC] across the selected sources.

Give me the core pillars, the subtopics under each, and which
source carries the primary evidence for each subtopic.

Order the subtopics as a reading sequence: foundations first,
implications last.

Run this on day one, before you have read the sources properly. It produces structure rather than answers, which you need to ask better questions later. The reading-sequence instruction sits at the end because it is the constraint most likely to be dropped.

2. Separate claims from evidence

Build a table of every substantive claim about [TOPIC] in the
selected sources.

Columns: claim, the evidence given for it, source, and the limits
on that evidence — sample size, assumptions, context.

Where a claim has no data behind it, put "asserted, not evidenced"
in the evidence column.

The last line earns its place. Without it, the model fills the evidence column with restated claims, and the table looks rigorous while telling you nothing. Ask for this as a Data Table in Studio rather than in chat, and the result exports to Sheets.

3. Find the disagreements

Where do the selected sources disagree about [TOPIC]?

For each point of friction: the competing positions, the evidence
each side offers, and whether the disagreement is about data,
method, or vocabulary.

Rank the disagreements by how much each one would change a
decision.

The data–method–vocabulary split is the useful part. A large share of apparent contradictions in a research pile are two authors using different words for the same thing, and you want those separated from the disputes that actually matter. The ranking line turns the output from a catalogue into something you can act on.

4. Name what is missing

List what the selected sources do not cover about [TOPIC].

Include: questions they raise but never answer, steps missing
between their stated evidence and their conclusions, and the kind
of data that would settle the open questions.

Build the gap list from what is present in the sources rather than
from outside knowledge.

This is the one prompt where a grounding instruction still earns its keep, because gap-finding is the request most likely to pull the model toward general knowledge about the field. Note the phrasing: it tells the model what to build from, not what to avoid.

5. Pressure-test yourself

Test my understanding of the selected sources one question at a
time.

Ask a scenario question that requires applying something from the
sources. Wait for my answer. Then tell me what I got wrong or
imprecise, and cite the passage that settles it.

Keep the scenarios operational rather than definitional. Start with
question one.

The only prompt here that tests you rather than the sources. It is also the most expensive one to run, because chat length feeds the compute cost and this is designed to go many turns. Start a fresh chat for it, keep the session to a handful of scenarios, and use Save to note for anything you want to keep.

Three jobs that are buttons, not prompts

Some of the most-shared Notebook prompts reproduce a Studio output in prose. The button version is cheaper, more consistent, and exportable.

Studio also produces Audio and Video Overviews, flashcards, quizzes, infographics and slide decks from the same sources. Each generation shows an estimated compute cost before you run it.

What code execution changes

Alongside the July rename, Google gave notebooks a secure cloud computer that can write and run code over your sources rather than only reading them. Availability differs by plan, account type, and rollout stage, so check what your account actually has before building a workflow on it.

Where it is live, the evidence prompts change character. “Extract the metrics into a table” becomes computation rather than quotation: recalculating a figure from underlying numbers in a CSV source, checking whether percentages in two documents were derived the same way, or aggregating across a set of tables that report at different granularities. Ask for the working alongside the answer — a result you cannot inspect is a result you cannot verify, and the citation trail that makes Notebook trustworthy for text does not cover arithmetic the model performed itself.

When it goes wrong

• “Gemini Notebook can’t answer this question.” The request went outside the sources. Google’s guidance is to rephrase or ask something different — the refusal is the grounding working, not a bug. Creative rewriting and speculation are the usual triggers.
• The answer pulls in documents you did not want. Every checked source is in play. Uncheck the ones outside the question rather than telling the model to ignore them.
• The table comes back generic. Usually too many jobs in one prompt. Split it: one table, one lens, one run. Confirm that the constraint shaping the output is the last line.
• The model ignores an instruction to stay in character. A persona set in Custom style is being followed at the expense of the prompt. Shorten the persona or switch back to Default.
• You hit the limit mid-session. The budget refills every five hours, not at midnight. Check Settings for the reset time, and use Generate later to queue heavy Studio outputs so they run automatically when it resets.

Common questions

Set the style and length once, scope the sources with the checkboxes, then run the five prompts in a fresh chat each. If an output disappoints, cut a job out of the prompt before adding a rule to it.