NotebookLM Custom Instructions: Reuse a Clear Research Brief

To set NotebookLM custom instructions, open a notebook, choose Configure Chat, select Custom, describe the behavior you want, and save. Google’s current help calls the product Gemini Notebook; its chat settings guide↗ (opens in a new tab) documents the controls. Start with a narrow research job and a test whose answer you already know.

A reusable instruction should explain how to work, while the notebook’s sources supply the material to work from. fenbs can keep the approved instruction in Skills and prompts so you can copy it into another notebook without rebuilding it from memory. This workflow uses manual copying; fenbs does not connect to NotebookLM.

How do you add custom instructions to a notebook?

1. Open the notebook you intend to use in the web app. Confirm that it contains the research material for this job.
2. In the Chat panel, open Configure Chat. Choose Custom as the conversational style and enter a short instruction describing the task and expected response.
3. Choose the response length and save. Start with enough space for the evidence and caveats your answer needs.
4. Ask one small question about a selected source. Compare the result with the source before expanding to the whole collection.

Google’s explanation of custom chat goals↗ (opens in a new tab) describes setting a goal, voice or role. Use that space for repeat behavior, such as explaining research to a new employee. Put the particular question, such as which findings concern onboarding, in your chat message.

Keep the first test modest. Asking for a complete strategy makes it hard to tell whether a failure came from unclear instructions, missing material or an overly broad question. A focused question gives you a visible result to improve.

What should a useful research instruction include?

Describe the outcome before the personality. “Help me compare customer interview findings” is easier to check than “be an expert researcher.” Add who will read the answer, what evidence should support it, and how to handle an unanswered question. Tone comes after those requirements.

Goal: Help me understand the selected research sources.
Audience: A project lead who has not read the documents.
Evidence: Base factual findings on the selected sources and attach supporting citations.
Uncertainty: Say when the sources do not establish an answer. Separate findings from your suggested follow-up questions.
Format: Start with a short answer, then list the supporting findings and unresolved questions.
Style: Plain US English. Explain unfamiliar terms. Do not invent quotes or missing figures.

This is a starting point, not a promise of perfect compliance. Replace “project lead” with your actual audience and remove requirements you do not need. If you want a study partner, ask for explanations and questions; if you want an evidence review, ask for findings and unresolved points.

Avoid putting an entire project history into the instruction. A sentence about the research purpose helps; several pages of changing facts make the reusable portion harder to maintain. The distinction also appears in ChatGPT custom instructions: stable behavior and changing project material need different treatment.

How do you keep answers focused on the right sources?

Check the source selection before changing the instruction. Google’s source guide↗ (opens in a new tab) explains selecting individual sources and naming them in a question. For a first test, select the interview or report that contains the answer you expect to find.

Say a design studio has interview transcripts and a proposed website plan. To learn what customers actually requested, begin with the transcripts. Including the proposed plan too soon could blur what a customer said with what the studio suggested. Your review should keep those two categories distinct.

Write a question that makes the distinction explicit: “Which interview passages describe difficulty finding contact information? Keep proposed solutions separate.” Open the supporting passages and check whether they say what the answer says. A citation is a route back to evidence, not proof that the interpretation is correct.

If the sources disagree, ask for the disagreement rather than a blended conclusion. Record a human-approved evidence rule in fenbs’s Decisions and rules if other assistants working on the same project must follow it too.

How can you test whether the instructions work?

Use three checks: a supported answer, an absent answer and a conflicting answer. For the supported case, choose a detail you have located yourself. For the absent case, ask about something the selected material does not establish. For the conflict, choose two passages that describe different experiences or conclusions.

• Evidence check: Can you locate the claim in the cited passage, with the same scope and qualifications?
• Uncertainty check: Does the response acknowledge the missing information instead of filling the gap?
• Format check: Are findings, suggestions and open questions separated as requested?
• Audience check: Could the intended reader understand the answer without a second explanation?

Change one instruction at a time and repeat the same question. If an answer becomes clearer, save the exact wording that helped. Keep the test question and expected behavior beside it so a colleague can assess a later revision. Copilot instruction examples offer another perspective on making reusable requirements concrete.

Why might a good instruction still produce a weak answer?

An instruction cannot replace missing evidence. If the material does not state why a customer canceled, asking for a confident explanation will not make that reason available. Request a list of supported observations and the questions you would need to ask next.

Conflicting requirements also make review harder. “Cover every detail” and “answer in one sentence” pull in different directions. Decide which matters for this task. If a short summary needs a detailed evidence appendix, ask for those as separate outputs rather than hiding the trade-off.

Keep tool-specific setup separate: use the guides to Claude custom instructions, Perplexity custom instructions and a custom GPT instruction template when adapting the same research approach elsewhere.

How to do this with fenbs

1. Create a free account, then add a project for the research you are doing. One person can start free, with no credit card.
2. Save the reusable research instruction in Skills and prompts. Include its purpose, intended audience and a sample question with the behavior you expect.
3. Record the evidence rules a person approved in Decisions and rules. Keep proposed changes distinct from settled requirements.
4. Copy the approved instruction into the notebook’s Custom chat setting. Check the selected sources, save, and run the same small test.
5. Save what worked in the project’s context notes and record where to continue in Where we left off. Revise the stored prompt when the tested wording changes.

fenbs is useful when several notebooks or assistants should follow the same approved approach. If you only need one notebook and can maintain its instruction there, the notebook alone may be enough. Keep source documents in the research tool; fenbs holds the reusable work instructions and the project record.

Finish by asking the original test question again and reading the supporting passage. The useful result is an instruction you can check and reuse, not simply a longer prompt. For more patterns that make instructions easier to maintain, browse the AI prompts hub.