The first month I used NotebookLM, I did what most people do. Upload the document, type “summarize,” read the output, and feel mildly disappointed without quite knowing why.
The results were accurate. Technically fine. But I kept sensing the most useful parts of the source were somewhere in there, and the summary had sailed right past them. Key context. Subtle connections between sections. The specific data point buried on page 14 of a transcript.
The problem turned out to be one word. “Summarize” is the wrong instruction. Swap it out, add a handful of steps before and after, and NotebookLM produces something completely different.
This guide walks through the exact five-step workflow I use whenever I upload new sources. You’ll get the specific prompts to type, where to put them, and why each step works. No vague advice about “prompting better.” Actual prompts you can copy right now.
NotebookLM is free to use at notebooklm.google.com and recently added ePub support, making it one of the most capable free research tools available right now.
Most people are using it at maybe 20% of its potential because of how they phrase their first question.
What Happens When You Ask NotebookLM to Summarize
When you type “summarize,” you’re telling NotebookLM to compress. The model’s job becomes finding the most prominent themes, cutting anything that looks marginal, and packaging the result into a readable block of text.
The problem is that marginal-looking information is often the most valuable part. A summary of a business call transcript will catch the headline decisions and miss the hesitations, the caveats, the offhand remark in minute 47 that changes how you read everything else.
A summary of three research papers will blend them together and lose the contradictions between them.
From what I’ve seen, this matters most with messy inputs. Audio transcripts, meeting notes, and collections of PDFs that overlap on similar subjects are the worst case. The model flattens them. What you get back reads like a Wikipedia article on the general topic, not a synthesis of what’s actually in your specific documents.
The word that changes everything
“Summarize” signals: be brief, be broad, leave out the edges. “Explain” signals: build a structure, go through the reasoning, connect the parts. Those signals produce different outputs from the same underlying model. Test both instructions on the same source, and the gap becomes obvious. The next four steps build on that insight.
How to Build a Topic Index Before Asking Questions
The first thing to do when you upload sources is to stop yourself from immediately typing a question. Before you ask NotebookLM anything, ask it to organize itself.
The technique is to request a topic index using only titles, nothing else. Ask it to scan all uploaded sources and output a list of the main subjects it finds, without explaining any of them. You get a clean skeleton of what is inside the documents.
This works particularly well with messy data. If you’re uploading overlapping PDFs, audio transcripts, or notes from multiple sessions, the index reveals structure you didn’t know was there.
What seemed like one big topic turns out to be seven distinct sub-topics, some of which only appear in one source.
When to skip the index step
If the source already has a table of contents, skip this. Books, formal reports, and structured whitepapers already provide the skeleton.
The index trick is for sources that lack structure: raw transcripts, research notes, exported chat logs, and mixed PDF collections.
Vague prompt (avoid this):
Summarize my uploaded documents.
Index prompt (use this instead):
Scan all uploaded sources and generate a list of the main topics and sub-topics they contain. Output the topic titles only, with no explanations. Use numbered formatting.
Why “Explain” Gets Better Results Than “Summarize”
Once you have the index, the next step is asking NotebookLM to work through it. And the word you choose here matters more than anything else in the workflow.
“Explain” tells the model to build, not compress. When you use it, NotebookLM constructs a logical structure around each topic, walks through the reasoning, and pulls in detail from across your sources.
You get paragraphs that follow a line of thought rather than a compressed highlight reel.
In my testing, “explain” responses run two to three times longer than summaries and retain far more of the source material.
The explanation format forces the model to commit to an argument or a process rather than just noting that something exists.
How to frame the ‘explain’ request
The most effective version pairs the ‘explain’ instruction with the index you already generated. Ask NotebookLM to explain each topic from the index, drawing from all uploaded sources.
This keeps the model working with your specific documents rather than defaulting to its general training knowledge.
Before:
Summarize the documents.
After:
Using the topic list from the index, explain each topic. For each one, draw specifically from the uploaded sources and build a structured explanation rather than a brief overview.
| Instruction | Typical Output | What Gets Missed |
|---|---|---|
| Summarize | 3-5 paragraphs, general themes | Nuance, contradictions, specific data |
| Explain (no index) | Longer but still broad | Structure, source-level detail |
| Explain with index | Deep, structured, source-specific | Very little |
| Explain one-by-one | Maximum depth per topic | Nothing when done right |
The One-by-One Technique for Deep Research Results
If you need professional-grade analysis, the next level is asking NotebookLM to explain each topic from the index individually rather than all at once.
This is the slowest approach, and it produces the best output.
When you ask for all topics in a single response, the model balances effort across all of them. Some topics get more space than they deserve; others get less.
Asking for one topic at a time removes that constraint entirely.
The individual approach works best for research projects, due diligence reviews, and preparing for presentations or interviews where you need the full picture on every sub-topic.
For everyday queries, the explain-with-index method is enough.
How to run it
Start with the first item from your index. Paste it and ask NotebookLM to explain it fully, drawing from every uploaded source. When that response is complete, move to the second item and repeat.
Prompt template for each topic:
Explain [TOPIC TITLE FROM INDEX] in full detail. Search across all uploaded sources for every relevant piece of information on this specific topic. Do not rush. Build a complete, structured explanation.
How to Set Up Custom Instructions and the Patience Prompt
NotebookLM has a Custom Instructions field in the notebook settings. Most people leave it blank. Filling it in changes the behavior of every response in that notebook, not just the one you’re currently typing.
Two things belong in your Custom Instructions for any research notebook. First, paste the index you generated. This keeps the model oriented around the structure of your specific documents rather than general knowledge. Second, add a patience instruction.
The patience prompt sounds odd but it works. Telling the model to “take your time and don’t rush” gives it conceptual permission to generate longer, more carefully constructed responses. You get fewer clipped answers and more thorough explanations without having to ask for elaboration every time.
What to paste into Custom Instructions
Custom Instructions template: “` Research focus for this notebook: [paste your topic index here].
Important: Take your time when researching. Go deep into the uploaded sources. Do not rush your analysis. Build thorough, detailed responses by reading carefully across all documents before answering. “`
Paste this once, and every subsequent conversation in that notebook starts with this context already loaded.
The Full Five-Step NotebookLM Workflow
Putting all of this together, here is the sequence for every new notebook:
- Upload your sources (PDFs, transcripts, audio files, ePub files, notes)
- Generate the index: ask NotebookLM to output topic titles only, no explanations
- Paste the index and the patience prompt into Custom Instructions
- Ask NotebookLM to explain each topic from the index, drawing from all sources
- For high-stakes research, go topic-by-topic through the index for maximum depth per subject
The sequence adds maybe five minutes to setup compared to typing “summarize.” What you get back is output that reads like a briefing prepared by someone who read every page.
If you need similar multi-source synthesis across different tools and file types, Sider AI applies a comparable approach and works with content types NotebookLM doesn’t currently support, including live web content and browser-based research.
