How to use NotebookLM to do real product research (with the prompts)

NotebookLM can do in an afternoon what used to take a research team a week, if you know how to prompt it.

In this episode of Human × Intelligent, we walk through a complete 8-step AI-assisted research workflow using a real Spotify UX interview study as our working example. Eight participants, 60-minute sessions and a set of raw transcripts, turned into personas, empathy maps, Jobs to Be Done analysis, How Might We questions, an opportunity matrix and a full synthesis report.

Every step includes the exact prompt to paste into NotebookLM. No vague instructions. No, just ask AI to help you. Real prompts, real frameworks, real output.

What you'll learn:

• How to orient NotebookLM before any analysis begins (and why this matters)
• How to build 3 grounded user personas, including a tension map that shows where their needs conflict
• How to create empathy maps per persona using actual participant language
• How to identify functional, emotional and social Jobs to Be Done and rank those that are most underserved
• How to generate and prioritise How Might We questions that open up real solution space
• How to build a feature opportunity matrix and effort vs impact quadrant
• How to affinity cluster raw insights into a 3-level observation → insight → opportunity hierarchy
• How to generate an executive summary, full research report and stakeholder presentation outline

All 16 prompts are included in the show notes as a ready-to-use guide.

This workflow applies to any qualitative research: user interviews, usability test notes, support tickets, survey responses. If you can put it in a document, NotebookLM can help you make sense of it.

Show Notes:

• Full prompt guide PDF
• NotebookLM
• Follow Human × Intelligent

Transcript

SPEAKER_00 0:08

Hello, welcome back to Human X Intelligent, the show where we explore what it actually looks like to work alongside AI in the real world. Not the hype, not the fear, just the practical, honest. Here's what I actually did and here's what happened version of it. I am your host, Madeline Costa, and today's episode is one I've been wanting to make for a while actually. And I've been presenting it at Yad University for two times now, and I'm very excited to bring to the podcast today. We are talking about Notebook LM. Specifically, how to use it to do real structured product research. The kind of research that used to take teams a week to synthesize, and now you can actually get it a serious first pass in the afternoon. Yeah. For but always with our judgment in mind, of course. And I'm not going to be here like sharing something vague about it. I'm going to give you the actual prompts, word for word. It's always going it's also gonna be on the description, so feel free to copy and paste it and also tweak a little bit to match what you're searching for. So you can open Notebook LM right now. And let's follow along. So I will open mine now. Okay. So I've opened our notebook LM and I'm going to share with you and walk you through a way with something useful. And by the end of this episode, you'll have it and you can do it as well. So without further ado, to make concrete, I'm using a real example throughout. A set of user interviews conducted for Spotify. Eight participants, 16-minute sessions covering how people discover music, um, how they feel about recommendation, and what's frustrating them. Basically a classic UX research. So without further ado, let's add the information. Like I had it here. Okay. Okay, let's get into it. So, with the first section of this is what is Notebook LM and why does it actually matter for research? Because I think as designers, we really want to know the why behind it. First, a quick framing moment for anyone who hasn't spent much time with Notebook LM. So, Notebook LM is Google's AI research assistant. What makes it different from just asking ChatGPT or Clove a question is basically this. It works from your document. It creates an environment around what you need and what you want for that scenario. You upload sources, PDF, transcripts, notes, articles, basically whatever is necessary for you to do this kind of research. And it only draws from this, from those. This doesn't hallucinate facts from the internet. It's grounded in what you gave it. Yes, you can search on the internet, but you can also deselect the one that doesn't make sense for what you're doing. So please always read or at least give a walkthrough of what you're actually adding to this environment you're creating. And for product research, this is actually very massive. So let's get into it. Here's how this used to work. You used to do like 10 interviews, each one around an hour, and you've got 10 transcripts on you. Someone usually like a research, sometimes a junior PM, has to read all of them. Pull the teams, write the synthesis document, build personas, map a persona, build personas, map insights to opportunities. Basically, it's slow. It's really expensive, and honestly, it's inconsistent. Different people find different things depending on what they are or what they were primed to look for. Even a better way of doing this, it's also to give it rules on the same place that you added, like here. On the same place that you added the Spotify PDF in this case, you could also add rules for this environment to follow to make sure that it matches how you do your work as a UX researcher or junior PM or PM, whoever you're whoever you're doing with this information. Notebook LM doesn't really replace the judgment, or at least it shouldn't. So please keep that in mind. But it actually does the heavy lifting of pattern recognition so you can focus on interpretation, which is pretty cool. So the workflow I'm walking you through today works for any kind of user research, interview, survey responses, support tickets, usability test notes, anything you can put in a document. And I've structured it in eight steps that build on each other, basically. So step one rounds the data. Steps two through seven are your core analysis frameworks, personas, empathy maps, jobs to be done, how might we opportunity mapping, affinity, clustering. And step eight pulls everything into an actual report. Pretty cool, right? You don't have to do all eight in one sitting. Actually, you shouldn't do it to make sure it works properly. Each step stands alone, but if you do them in order by the end, you'll have a research synthesis that would genuinely impress a VP of product. So let's go. We are starting with the grounded data, and before you run any analysis, you need to orient Notebook Land. That's basically what it means to ground the data. Think of it like a briefing a very capable colleague who has never seen your research before. Your first job is to make sure it has sense what it's looking at. So you're gonna upload your research document like I did. In our case, this is gonna be the Spotify interview PDF. Once it's in, run these two prompts to start. The primary prompts, so let's copy. One second. It's gonna be this one. This one. Again, I will add everything in the uh description so you can copy-paste it. But this is the one that we are doing right now, so we are summarizing the key facts of the research document. Once it's done, please make sure to read it. Right now, I'm not going to do it because I don't have enough time to do it, but please do it. And the next prompt that we're going to give is to read through everything that the parti all the participants quotes and actually understand the verb between teams, basically. So let's do it. But basically, this sounds pretty basic, but it matters. It actually forces Notebook LM to index the document correctly and gives you a clean baseline. If it gets this wrong, if it is miscounts participants or misreads it a demographic, you want to catch that. Now, not that after you've built the three personas on top of it. So that's why I said that it's very important for you to read it through and to guarantee that it's correct, basically. And it also gives an ID because it will help us to identify in the next steps also everything. Like here, mention by, mention by, etc. And it connects everything to what we are working on. Then basically, after we run this one, what you're doing here is letting the AI do a first pass at a pattern recognition before you bring your own assumption in. The mention by at least two participants filter its importance. It stops you from chasing one-off opinions and keeps you anchored to signal not noise. In the Spotify research, this step surfaced four clusters almost immediately. So you have, let's see, you have the mood and contextual adoption, you have the social connection gap, you have the library organization friction, you have the content sealers, you have the external discovery reliance and the algorithm rituals. So a bit more, let's say one, two, three, four, five, and six. You have six right now. Those became the backbone basically of the entire analysis. This will be the backbone. So now that you have all this information, we are going to the step two, which are personas. And I know some people hear the word persona and they roll their eyes. Fictionalized compo accompass users, marketing fluff, uh box you tick before moving on. And I do get it, I get what you are coming from, but it is an important step for right now. And done right, personas are can be genuinely useful. They give your team a shared mental model of who you're designing for. They make disagreements in design reviews more productive. Instead of I think users want X, you can say the curator definitely wants X, but this would frustrate the passive listener. Completely different arguments, and they also mismatch different personas, different groups of people that will be using your product, and that's why it's important. But here's how to build them from your research in Notebook LM with three prompts. So the first one is this one it's the identity of personas archetypes, and basically this one we run in it to get these archetypes, and then we go deeper on the next prompt because for each of these archetypes, which are the sonic ritualist, the multimodal navigator, the external discoverist, which are three different archetypes. We are going to actually give it more details so we can actually really understand where they're coming from or who they are in the sense. So we are building the full persona profiles right here, which is very important for our work. But after we do this, and we should never skip this, we are going to do the persona tension map. Basically, this is where it gets interesting. In the Spotify data, the tension between the power user, someone who wants a richer social layer, and deeper artistic context, basically. And the passive listener, of course, who just wanted to work quietly in the background is real and it's actually very significant because a feature that excites one actively unlines the other. And knowing about this tension changes are your sequence, your roadmap. And that's another reason why it's very important. So you can see like the a big overview here, so you can just see the small details and actually see where you can prioritize or why you should prioritize something. A quick note here is once Notebook LM names your persona, like I showed you the identity persona archetypes, you should write them down. You'll use those exact names in the next step. It helps the model stay consistent. So let's go to the persona archetypes for each archetype, exactly. So let's do the sonic ritualistic first. Okay, we have it here. Let's do it. Create the empathy map. Okay, it's tanking, it's thinking. Let's go get the next one. It's here. The multimodal navigator, right? Yeah. And then let's just do the external discoveries as well. Okay, so let's do it. Now we have this one, which is the size, like I asked, the quadrants, says, thinks, does, feels. And this goes to the pains and gains based on it. So we are already creating something very deep. So let's create for the next one. Then we have here already the third one. It's gonna be one per persona basically, or at least uh the following how we're doing. Um here an empty map is a simple framework, as we know. We go from four quadrants, what does this persona say? What do they think? What do they do, and what do they feel? Plus a pain section and a gain section below it. Because the goal is to get underneath the surface of your data and understand the emotional reality of your user and not just their stated preferences, basically. So let's do the last one. Again, please make sure to verify everything to see if it connects, connects within every scope. This is also very useful if you are the only designer in a team and need to make sure to have everything detailed and present something that is justifiable based on the interviews of the users, based on your users. So it's also a good way, a good friend brainstorming kind of related research uh pal. But basically, the prompts here are pretty much straightforward. But the delivery here matters. You're going to run one per persona, like I was saying. For persona one, you're going to swap this name like we did on the note, like the notebook LM gave you. And I'll I'm going to add this below so you see where you need to add the persona one name, the persona two name, the persona three name, so you can do the same work with your own scope, with your own project, basically. And basically the last instruction asks for the surprising stuff, as something to add to every third persona, because by the point the model can get a bit formal-like, basically. Asking it to flag what's unexpected forces to look harder at the edges of the data, which is something that we want. In the Spotify research, the counterintuitive findings for one persona was that they didn't actually want more recommendation. They wanted better findability of things they already saved. That's a completely different design problem, and it only surfaced because, oh yes, this unexpected one, right? So, the next step of this session of this next board is session is step four, which is jobs to be jobs to be done, GTBD, for those of you who live in the product land. The framework is pretty simple. People don't buy products, they hire them to do a job. Understanding the job, the real underlying motivation, is more durable than understanding futures or preferences, which change constantly. And there are three types of jobs. Functional, what does the user literally need to get done? Emotional, how do they want to feel? Social, how do they want to be perceived by others? So let's add the next prompt. And as you see here, we have here a format each job as when certain situation, I want to certain motivation, so I can expect an outcome. And this is really gonna help us. Basically, it's where we are going to also go to the connection with directly with the product decision. So now we have the functional jobs, emotional jobs, social jobs, but we actually want to connect it to Spotify or to the jobs that we want to do. So which jobs are currently undeserved by Spotify based on participant feedback, and then by frequency of mention and intensity of frustration expressed. So, what you end up with is a prioritization list of user needs that aren't being met right now, ranked by how badly and how commonly they are happening. That's the input your product team needs to start making decisions about what to build next. So, from Spotify interviews, the most undeserved emotional job was something like when I open Spotify in a specific mood, I want to immediately understand my headspace so I can skip the friction of finding the right music and just be in it, you know? So six out of eight participants express some version of it, different ways of writing, different ways of saying, but basically that's the overview, the goal, what they are saying. And that's not a feature request, right? That's a fundamental unmet need. And it's very important we understand how to do it. And step five is how might we questions, which MW. This is a design thinking staple. It's one of those frameworks that sounds almost too simple until you see what it produces. The idea is to take a user problem and reframe it as an open-end opportunity question. Not users are frustrated that recommendations don't match their mood, but how might we make Spotify feel like it reads the room? The reframe matters because it opens up solution space. Fix the algorithm is one answer to the first framing. The second framing invites a hundred different answers. Basically, gesture controls, time of day awareness, manual mood selector, biometric input from a wearable, contextual playlist triggered by calendar events, for example. And all of those are valid explorations that the problem statement would have closed off. So let's do the how my twee, shall we? Let's do it. This is the first part of the how my twee. We are generating the questions right now. And then we are going to a fun one. I believe it's a fun one, which is the prioritization pass. So we have all the how my twis, like how my tweet create a seamless bridge for users to discover music on external platforms like TikTok or Shazam to save those tracks instantly to their Spotify library. But let's see what our system is building and what's the environment we are creating, and which ones are actually the prioritization one, the priority ones, or the highest potential impact for our users. Basically, that three criteria filter like a frequency, intensity, strategic fit is something I recommend for any prioritization exercise. Frequency alone gives you what's common, basically. I think it's like um, and it gives you intensity alone gives you what hurts most. Strategic fit asks whether solving it actually moves the needle for the business. You need all three to see this all environment. And we have it. We have here the first one, which is how might we allow users to signal their current activity to dynamically adjust recommendation energy in real time? And we have the frequency, the intensity, and the alignment for the three of them. No, we have five, sorry, for the five of them. Now we are going to something a bit different. I'm sure you know of it, which is step six and seven. Basically, I'm going to move through together because they're both about making sense of everything you've generated so far. Step six is the opportunity map, and we have two prompts for it. The first one is feature opportunity matrix. So let's do it. Let's create this feature opportunity matrix. While we wait, I can tell you the next one, which is after effort versus impact within this opportunity map that we are creating and based on the participant quotes. So we can mean we can match both of them and continue creating this prioritization scale that makes sense for the opportunity maps that we are creating. And step seven is the next one, which is the affinity mapping. This is where you cluster everything into a hierarchy of insights. So we have all of this that we are building and now and even the low priority, as you can see, collaborative playlist promotion. So while users like James, the P02, like the idea I was telling you about, find collaborative playlists fun and underrated, the document classifies increasing their promotion as low priority. This is because basically the underlying feature already exists and works well. They're usually simply one of the discoveries for casual users rather than missing functional or emotional need. So it actually gives us a cool justification as to why, and you can also click here to see where it's coming from and where's the environment you're creating and the connections you are creating. So let's do it. Let's create the affinity clusters of the affinity mapping. Then perfect. We have the cluster label, participant counts, and synthesis statement. Amazing. And then we have the key cluster insights, real-time context, social preserv presence, library friction, cross-content discovery. We are going to the insights hierarchy based on the affinity clusters that we just created. Tree-level hierarchy like observation, insight, and opportunity is something I now use as a quality check basically on all research synthesis. If you can't make the leap from what you observe to what it means and what to do about it, your insight isn't done yet. And you need to make sure you have this. Perfect. Now we have everything. Again, don't forget to read everything once you do it and make sure that it connects correctly. But we are getting to the final step, which is the final synthesis, the report, basically. And we are going to start with the executive summary before we go to the full research. So let's do it. This is where everything comes together. And I'll be honest, if you run all the previous steps, Notebook LM has built a lot of context. Now it has a lot of context and analyze it a lot and how we want it to be analyzed. This prompt is going to produce something genuinely good. And this is based on three different prompts, the one that I'm going to give you now, and depends on what you need. So for the first one, it's going to be the executive, like I said, the personalization discovery study. So the Q1 2025 research study evaluated Spotify's personalization and discovery engine through 8 60-minute qualitative interviews, qualitative interviews with a global mix of free and premium users. The objective was to access the efficiency of personalized recommendation rituals and existing playlist discovery features. Then we have the critical findings, which are the algorithm stagnation, the social deficity, and the fragmented content ecosystem. Then we have the strategic recommendations, which are implement a session mode selector, revive the social layer, deploy a cross-content bridge. Also, what you are going to do right now is go here, generate report based on your sources. We are going to create our own in English. Let's do it. And we are going to paste here our prompt. Everything is here that we want to have it. The structure, the steps, and then the professional. So let's generate. It's being generated here as you can see. While it's being generated, I'm going to share with you a stakeholder presentation outline prompt that you can use. Create a side by side outline for a 15 minute stakeholder presentation based on this research. For each slide, include slide title, two, three bullet points of content, and a speaker note suggestion. If you need more to each one, you can give it as well. And you can also Connect the report search with another tool, like for example, Claude Shat GPT Gemini, to actually use this prompt there with the report that you are generating, like this. It will have images and other things, and you can just change it everywhere with the slides with the slides Google Slides app, for example. But for today, we are going to do something simple and just create it here. So let's do it. Basically, here the goal is for you to run whichever matches your audience, or run all three. They each serve a different moment as we can see a different purpose. Here, the summary can be a different thing, the slides can be a different thing, and the actual Spotify reports that you can see here. Study overview, participant profiles. We have the key findings, user personas, empathy map highlights, the jobs to be done like we did, perfect, how might we? Feature opportunity matrix, recommendations that we can use strategically, uh, or quick queens that we can use, and the next steps for us. Also, and that is confidential no test, basically. Here we have the structure of the slides, the content, the speaker note, basically everything. Of course, review it and give it your own flow, your own uh way of thinking. But yes, basically, this is what we we do, what we can do with UX research. This is really a really simple way of doing something quick for you to see it and start to use it. But please make sure to use it in a way that works for your product and it's focused on your product. Also, if you have any doubts, if you want to talk about it, if you want my help to even personalize more to your product, your type of research, because each different UX researcher different ways of working, different ways of creating or using frameworks. So feel free to message me. I'm more than happy to help. But yeah, this is pretty simple. You can also add a lot of more information here. What I like to do is besides this one, I like to add a set of a certain type of guides, branding, and everything I actually need to make sure that while it responds, it gives me exactly what I want. Also, for example, the Spotify description, wherever it is, Spotify application, how it works, what it is, depending on the context of what we are doing. So while this is generating, I just want to say and hope that this lands for you because eight steps, grounded data built by personas, the map empathy, identity, identify jobs, the reframe of how my tweet, map opportunities, cluster your insights, and synthesize the report. All of it on Notebook LM. All of it grounded in your actual research documents. A few things I want to leave you with is one, this doesn't replace the judgment. I cannot say this enough. Notebook LM is very good at pattern recognition. It is not good at understanding why a pattern matters specifically or strategically, or whether your company is actually positioned to do anything about it. That is your job. Use the AI to surface then. Use your brain to actually interpret this data. That's why Human X Intelligent exists. That's why we created this for you to know how you can use these tools in a way that still makes you matter and shows your strengths, basically. The number two that I would like to leave you with is save your outputs as notes inside Notebook LM as you go. The model doesn't carry memory between sessions the same way a human would. Pinning strong outputs keeps them in context for later. So you can also pin everything, like save to note right here, and it will save to note and you can see it wherever so you don't lose exactly what you want. It's amazing. You can convert to a source even, you can convert all notes into sources, you can export to docs, and you can export to sheets for you to do and continue doing your job. So please do this as well. The number three is if a response feels generic, push back, tell it. Then follow-up prompt I keep in mind, basically on my back pocket, is make this more specific to the contest and ground it in participant quotes from this document. I will also leave it below, but basically this is like my safe to go to make sure it's working. How like I want. So nine times out of ten, it sharpens the output immediately. Everything in this episode, all eight steps, all the prompts, word for word, is in the show notes. Copy them, paste them, use them. That's why they are there. If you try this on your own research, I'd love to hear what came out of it basically. Share with me, message me, email me, reply on the YouTube comments. But please do um say something about it because I think it's interesting and we should help each other and grow with each other and actually share what we are learning. And find us basically wherever you want to listen and reach out. I am your host, this is Human X Intelligence, and we won't be able to see the slide deck because it's taking along. But if you were able to, please also message me. Let's talk. This is what it makes us human in this new AR world. So, see you next episode. Thank you for watching and listening.