Strategic Note-Taking for UX Research & Better AI Prompts

1. Introduction - Welcome to the Course: Hey, I'm Pascal. I've spent over a decade helping companies understand their customers and make better product decisions using human centered design methods. Over the last few years, I've been deep in a question that sits at the intersection of tech and human development. How do we use technology to support more conscious ways of living and working instead of outsourcing our judgment? This class is one practical answer to that question built for real research and product work. You work in research or design right now, you've probably noticed something weird happening. We have more data than ever, recordings, transcripts, AI summaries, and yet it's getting easier to feel less connected to our customer. The risk isn't that AI misses things. The risk is that we disengage. We do a customer interview. We click summarize, we read the output, and without noticing, our perception gets outsourced. And that matters because in UX research, your job isn't to produce a transcript or a summary. Your job is to make sense of messy human reality. Take responsibility for what you conclude. This course gives you a simple and powerful method to stay cognitively engaged while you interview and to use AI without seeding your brain. It's a workflow I call human first, machine second. Human first means you capture your signals before the AI sets the frame. Machine second means you use AI to test your thinking, expand your view, and retrieve evidence, not to replace your judgment. You'll learn a fast five minute post session ritual that helps you walk out of every interview with real hypotheses, not just pages of notes. You'll learn a small toolkit of notation techniques that make your note taking more strategic Metacognition markers, emotional arc tracking, question cascades, context anchors, observations versus interpretations, and contradiction mapping. You'll learn how to turn those human signals into better AI prompts and a clear evidence trail. So your insights are easier to stand behind. This works because AI is literal. It's great at compression, but it doesn't care. It doesn't notice like you notice, and it can't take responsibility. Strategic note taking is how you keep the quality of attention that makes research valuable in the first place. This is for UX researchers, designers, service designers, and product teams doing discovery, especially if you're often working solo and you want to use AI to move faster without lowering the quality bar. If you're ready to stop being an operator of outputs and start being a steward of impact again, you're in the right place. Let's get into it. 2. Lesson 2 - Awareness as a research instrument: One's worried about AI becoming self aware. But here's the real question. Are you self aware? I mean that genuinely because awareness, the ability to notice what's happening inside you while you're paying attention to someone else is one of the most underrated skills in research. And it's the one skill that AI simply cannot do for you. Let me explain what I mean. When you sit down with a participant, you're doing two things at once. You're capturing what they say and do. That's the behavior, the stories, the reactions, and that's the outer data. Most of us are pretty good at that part, but there's a second strain running underneath, and it's what you are thinking and feeling while they are talking. Your surprise, your boredom, your urge to jump in and help, that nagging feeling that something doesn't add up, and that's inner data. Most of us let that float right past. And here's the thing. Inner data is not true. If you feel bored, that doesn't mean the participant is boring. If you feel defensive, it doesn't mean they're wrong. These reactions are signals, and they're worth capturing, not because they're automatically right, but because they give you something concrete to test later. And this is what researchers call meta cognition. Thinking about your own thinking. And I know that word can sound a bit academic or even a bit woo woo. But in practice, it's incredibly simple. It just means pausing long enough to notice, what am I reacting to right now? I noticed I wanted to rescue them. That's a signal. I felt bored when they talked about onboarding, that's a signal. I felt defensive when they challenged the design. That's a signal, too. None of those are conclusions. They are starting points. And once you write them down, you can do something with them, including asking AI to help you test whether the signal holds out. So how do you actually build this muscle? I want to walk you through a short practice. It takes about 3 minutes, and you can do it before any interview. It's completely optional, and some people love it, and some people find their own version of it. But give it a try at least once because the principle behind it matters more than the specific steps. Ready? So here's how it goes. So just close your eyes for a moment if that feels comfortable and take a deep breath in through the nose and out through the mouth. Just one breath. Just notice how that feels. And now rub two fingers together slowly with just enough pressure that you can feel the rides on your fingertips. And this is a small track that anchors you into the present moment. Impose your attention into your body and out of whatever was spinning in your head a second ago. Now, open your eyes and look around and notice one detail that you hadn't noticed before a color, a texture, a sound, just something small. And now ask yourself, right now, what am I feeling? Energized, skeptical, overwhelmed, curious? There's no wrong answer, notice. And think about your day so far. What surprised you? What confused you? What confirmed something you already expected? And one more, think about AI in your work for a moment. What feeling comes up? Relief, anxiety, skepticism, excitement. See if you can notice where that feeling sits in your body, put it in your chest, your shoulder, your stomach. You don't need to do anything with it, notice. And now notice what it felt like to notice. That's the muscle. That's metacognition in action. You don't need to do this full practice before every interview, but even a 32nd version, just one breath, just one check in, can shift the quality of your attention. Because when you walk into an interview already aware of your own state, then you're far less likely to confuse your reactions with your participants reality. And here's why this matters for the rest of the course. Later, when we get into prompting AI, the quality of what you ask depends entirely on the quality of what you noticed. If you walked out of an interview with a vague sense that it went well, then you'll write vague prompts and get vague outputs. But if you captured a specific signal, I noticed I felt uncomfortable when they described their workaround. Now you have something real to work with. You can then ask AI to pressure test it to find counter evidence or surface patterns you might have missed. So awareness isn't a nice to have. It's the foundation that makes everything else in this course more useful. And you don't need perfect notes. You just need to pause and ask, What am I thinking right now? And that's where we start. And in the next lesson, we'll turn that into a repeatable method. 3. Lesson 3 - The method: Human-first → Machine-second: The last lesson, we talked about awareness, noticing your own reactions during a session, surprise, confusion, defensiveness or boom. And we called those signals. But here's the honest truth. Noticing a signal in the moment is only half the job. If you don't do something with it quickly, it fades. By the time you open your laptop an hour later, it's gone. Or worse, it's been quietly rewritten by your memory into something tidier and less useful than what actually happened. So in this lesson, I want to give you a simple repeatable method you can use after every interview. It takes about 5 minutes, and it's the bridge between what you sensed as a human and what you'll later ask AI to help you with. Method has four steps. Tune into the body, name the signal, then capture it, and then prompt AI. Step one, tune into the body. The interview just ended, and you've said goodbye. Before you do anything else, before you check slack, before you tidy your notes, before you debrief with a colleague, pause for about 30 seconds. Take a breath, maybe rub two fingers together with such a pressure that you can feel the ridges on your fingertips. This is a mindfulness hack that gets you to tune into your body and then ask yourself, how do I feel right now? Not how did the interview. That's an evaluation. I'm asking something more basic. Are you energized or drained, tense or relaxed? Do you feel like something important happened or like the whole thing was flat? You're just checking in. The body often registers things before the mind catches up. If your shoulders are tight or your energy suddenly dropped or you feel a weird restlessness, that's data. It's worth paying attention to. Step two, name the signal. Now get a bit more specific. What stood out? Not what the participants said, we'll get to that, but what did you notice? Try to name it in plain language. I wanted to rescue them when they described the workaround. I felt skeptical when they said they loved the onboarding. Something felt off about the way they talked about their manager, but I can't put my finger. These don't need to be polished. They don't need to be right. They just need to be honest. You're naming what your instrument that's you picked up during the session. Step three, capture it explicitly. This is the part most people skip, and it's the part that matters most. Write it down, actually write it. Write it in your notes or on a sticky note or in a document wherever your session notes live. This simple format. What surprised me so something I didn't expect. Are there patterns forming something that echoes previous sessions? What's my curiosity? What do I want to dig into next? And to validate, what's a specific claim I want to test? Let me show you what this looks like in practice. Say I just finished an interview with a banking customer about their experience with a savings tool. Here's what my post session dump might look like. So I was surprised that they said they trust the app with their savings, but then describe checking their balance three times a day. That doesn't sound like trust to me. It was the third participant who describes the notification system as annoying but necessary. So I was starting to wonder if there's a tension between wanting control and wanting to not think about it. I was curious why they get emotional when describing the moment they hit their savings goal that felt like it meant more than just money. And they claim they never used the budgeting feature, but earlier mentioned setting a weekly limit. Worth checking if this is a contradiction or if they don't see the limit as budgeting. Notice what just happened. In about 2 minutes, I've captured four concrete things I can work with. None of them are transcription, none of them are summaries. They're signals grounded in what I actually experienced during that conversation. And here's what makes this powerful. Each one of those is now something I can hand off to AI with a specific question attached. This is where sequencing matters. If I'd gone straight to AI after the interview and said, summarize this session, the AI would have given me a perfectly competent, perfectly generic summary. It would have flattened out all the interesting tension, the contradiction between trust and checking three times a day, AI might not flag that. The emotional moment about the savings goal, AI might note it happened, but it wouldn't know that I felt something shift in the room. But now that I've captured my signals first, I can write prompts that are genuinely useful. The participant describe trusting the savings tool, but also checks their balance multiple times daily. Find evidence in the transcript for and against the interpretation that frequent checking reflects anxiety rather than trust. Or three participants have described a notification system as annoying but necessary. What language patterns in the transcripts might help me understand whether this is genuine ambivalence or a polite way of saying the notifications are too frequent. See the difference? These aren't lazy prompts. They're investigative prompts. They start with something the human noticed and ask AI to help test it, challenge it, or find supporting evidence. That's the partnership working well, and that's why the sequence matters. Human first, machine second. You notice, you name it, you write it down, and then you bring in AI. Because if you skip straight to AI, you're letting the tool decide what's interesting. And the tool doesn't know what surprised you. The tool doesn't know what felt off. Only you know that. One more thing before we wrap up. You might have noticed that in the post session template, some of those signals are already halfway to being hypotheses, and that's the goal. A signal like they said they trust the app, but check it constantly can become a hypothesis pretty quickly. Frequent balance checking may indicate monitoring behavior driven by low confidence in the tool's accuracy rather than high trust. This is what I believe now and what I want to test. Or the savings goal may carry symbolic meaning beyond financial, possibly tied to self efficacy or a personal milestone. Again, this is a hunch I've got that I can now test. You don't need perfect academic language for this. You just need to take your signal and rewrite it as something testable, something you can look for evidence for or against. And once you have a hypothesis, your AI prompts practically write themselves because now you're asking a specific question, not just tell me what happened. Before you move on, take two to 3 minutes to do this. Write down three signals you noticed in your last interview or in any recent conversation. Pick one signal and turn it into one testable hypothesis. Just one sentence. See you in the next lesson. 4. Lesson 4 - Your Note-Taking Toolkit: Last lesson, we talked about a simple sequence, human first, machine second. You tune into your reactions, you name the signal, you capture it explicitly, and then you invite AI in to help you test what you're seeing. In this lesson, I want to give you the missing piece that makes that practical in real interviews. Don't worry. It's not a big framework and not a new checklist you have to remember. It's a tool kit. It's a small set of lightweight notations you can use to capture signals without breaking the flow of a conversation. And if you use them well, they also make your prompts later dramatically better because you're handing the AI structured intentional inputs. Keep in mind that you do not need to use the whole toolkit all the time. You need one or two notations that fit your brain and your context. So we're going to do three things. First, I'll show you the three buckets of strategic notes. Then I'll show you a handful of notations and the job each one does. And finally, I'll show you how those marks translate into better AI instructions later. When people first hear notations, they sometimes imagine learning a whole language. That's not what we're doing. Think of this like adding two or three useful sticky notes to your system. And to keep this simple, we'll organize the toolkit into three buckets. Notes to capture thinking. So what's happening in your head? Notes to help you steer the session, what to do next in the moment and notes to direct AI later, what you want the model to test. The important thing is you're not trying to do all three perfectly. You're choosing a bucket that solves your biggest problem in interviews right now. Capture thinking is for fast marks that record your internal signal without interrupting the conversation. This is for those micro reactions you have in a session, maybe a weird tightness in your chest or a sudden spike of curiosity or a moment where something doesn't quite add up. If you don't mark it, it disappears. So the first bucket is about leaving tiny breadcrumbs for your future self. Here are a few example symbols. You can copy these or invent your own. What matters is that each mark means one thing every time. I use an exclamation mark for surprise, a question mark for curiosity or confusion, a lightning bolt for an energy shift, and the not equal sign for a mismatch. These are not notes about what they said, they're tags for what lit up in you and what you want to revisit later. You can choose your own symbols. The point is consistency. If you always use the not equal sign for mismatch, your notes become searchable at a glance. Here's what this looks like in the wild mid interview. For example, when the participant says, I trust the tool but checks their balance three times a day, I might write that down like this. That one mark is enough to remind you there's a tension here worth exploring. Bucket two, steer the session. This is where notations become an in the moment steering wheel because when you notice a signal, you basically have four options. Go deeper, park it, ask for an example, or gently challenge it. You do not want to do that decision making from scratch every time. So we use a few marks that translate directly into actions. On this slide, each notation maps to a move you can make in the moment. The benefit is you stop improvising your next step from scratch. You see the signal, you mark it, and you know what to do. A quick example, if someone says it's intuitive, you can mark it deaf because intuitive might mean the UI is simple. They already learned it. It matches another tool, or they can recover from mistakes. So your notation becomes a prompt to ask, when you say intuitive, what does that look like in practice? Or if something important comes up, but your mid story, you mark Park and keep them moving. That prevents derailment without losing the thread. Bucket three, direct AI later. This is the handoff. A good AI prompt usually starts with here's what I noticed or here's the hypothesis I'm testing, and here's what counts as evidence. Notations make that easy because you've already tagged the moments that matter. On this slide, you'll see a simple way to tag what you want the model to help you test later. The point is that your prompt becomes almost automatic because you've already marked the moments that matter. So you might write H frequent checking indicates anxiety, not trust. E, language about fear, reassurance, just to be sure, checking after notifications or ALT, checking as a habit or a ritual, or they enjoy the feeling of control. And now your prompt to AI is almost automatic. I suspect frequent balance checking reflects anxiety rather than trust. In the transcript, find evidence supporting and contradicting this interpretation and propose alternative explanations. Quote, the exact lines. That is a very different prompt than summarize the interview. On this slide, I'm connecting the whole loop, what you notice, how you test it, and what you want back from the model. Mark what mattered, then use AI to test it with evidence. So how do you choose a notation without overcomplicating it? Pick based on your context. If you do a lot of interviews back to back, choose notations that help you capture thinking quickly. If your biggest challenge is staying on track, choose notations that steer the session. If your biggest challenge is writing good prompts, choose notations that direct AI later and start with one or two, not six, because the goal is not perfect notes. The goal is a repeatable loop you can actually do when you're tired. And now it's your turn. Choose one technique to practice in your next session. In the next lesson, we'll go deeper into the specific notation set. 5. Lesson 4.1 - Meta-cognition markers: If you've ever re read an interview transcript and thought, Why didn't I notice that in the moment? This lesson is for you? Because most of what makes a researcher good is not the questions on the script. It's your instrument, your ability to notice surprise, tension, confusion, and confirmation as they happen while you're still in the room. Metacognition markers are a simple way to capture those reactions in real time without turning the interview into note taking theater. They're small, consistent marks that tell your future self, something happened here. Come back. In this lesson, we'll use three markers. The first one is for surprise. Something I think works well is using two exclamation marks because it feels like the emotional tone of that moment. Your attention spikes and you want to catch it before it disappears. The second one is for confusion. I like two question marks here because it keeps you honest. It's a quick way of saying, I don't understand this yet, and I'm not going to pretend I do. And the third one is for strong confirmation. When something clearly supports a hypothesis you're already holding. I use two arrows because it feels like a strong vector, a clear direction, and it reminds me to capture the cue so I can test whether it holds across the rest of the interview. The goal is not to be clever. The goal is to leave bread crumbs for your future self. Don't worry about memorizing these. The only job right now is to recognize them when you see them and to understand what each one is for. These markers are rigorous precisely because they are humble. When you mark surprise or confusion, you're not claiming truth. You are capturing a change in your attention, and attention shifts for reasons. Something contradicts what you expected. Something is vague, where it should be concrete, or something emotional and high stakes is showing up underneath the words. If you do not capture that shift, it gets smoothed over later and you end up with a transcript that reads cleanly but hides the real signal. Here's the key constraint. A good marker should take less than a second. You're not writing paragraphs. You're writing a short concrete phrase, and then you add the marker. Example, you might write a quick note like says it's easy, avoids mobile, and then add your surprise marker. Or you might write keep saying it depends and add your confusion marker. Or you might write checks prices first every time and add your confirmation marker. If you can do that, while staying present and keeping the participant talking, you're doing it right. When you use the surprise marker, you're saying, This is not what I expected. There might be a tension worth exploring. When you use the confusion marker, you're saying, I do not understand this yet, and I should not pretend I do. I need a clearer example or a sharper definition. And when you use the confirmation marker, you're saying, This supports a hypothesis I'm holding. Capture it now and test whether it holds across the rest of the interview and across other sessions. Notice how none of these are conclusions. They are placeholders for future verification. And here's what that might look like during an interview. Could you walk me through the last time that you used the app for savings? Okay, it was yesterday evening, actually. I was on the sofa, and I remembered I'd moved some money around earlier in the week, so I opened the app to check where I was at. Mm hmm. I usually check my current account first, then I look at the savings pot, and if it looks okay, I'll move a bit over. Mm hmm. Yeah, it sounds boring, but it's kind of a relief. Like, I've done the responsible thing for the day. When you say it's a relief, what's happening for you in that moment? I would say I feel calmer, but it's it's weird because I wouldn't say I'm stressed about money all the time. Like, I'm fine. It's just, I really don't like the feeling of not knowing, you see? Mm hmm. Basically, if I don't check, I can start thinking, did something go out? Did I forget a subscription? Did I mess up? Mm hmm. So yeah, checking is basically me reassuring myself. Good. And can you talk me through what tends to trigger that urge to check? It's the notifications. 100%, maybe I'll get a weekly spending thing or a message like, you spent more than usual on food, and my brain just goes, Oh, no. What have I done? Mm hmm. Mm hm. But then I also kind of like it because it keeps me honest. So I'm annoyed, but I also I also want it. So that annoyed, but I also want it bit is interesting. What about it is annoying? It's the timing and the tone. If it pings me at, like, 9:00 P.M. I'm already tired, and it's like it's judging me. I know it's not, but it kind of feels like a teacher marking your homework, you see? Yeah. And then I'll open the app, check the balance. And sometimes I'll move money into savings straightaway. Yeah. I know it's maybe not logical. Yeah. It's more like I'm trying to trying to undo the bad you see? Yeah. And when you move money into savings like that, what are you hoping will happen? That I'll be back in control. Also, I have this rule that if I've spent more than I meant to, I'll make up for it by moving something into savings. It's funny because I don't budget. I hate budgeting, but I do have a weekly cap for eating out. And, tell me about that weekly cap. How did you decide it, and what happens when you go over? I picked a number that felt reasonable. And yeah, if I go over, I feel guilty. Not huge guilty, kind of annoyed with myself. And that's when I do the savings move, you see? Mm hmm. Mm hmm. Yeah, I know it's not like a spreadsheet budget. It's just a guardrail. Yeah. But yeah, the app basically triggers the whole cycle. Yeah. Got it. To keep these markers clean, there's a small discipline you can lean on. When you write a surprise marker, keep it in the shape of I expected one thing, but I'm hearing another. When you write a confusion marker, keep it in the shape of, I need a clearer example of this. And when you write a confirmation marker, keep it in the shape of this supports hypothesis H for now. Structure keeps you honest. You're naming your own reaction, not diagnosing the participant. Later in the course, we'll use AI to do something very specific. We retrieve evidence for and against your human signal. These markers make that possible because you've already done the human work of noticing what mattered in the room. The difference is that AI is not deciding what mattered. AI is helping you test what you noticed. Now try it yourself. Do a five minute role play interview. Your only job is to capture exactly three marks. Surprise, confusion, and confirmation. Keep them short, keep them concrete. And if you can do that, without losing the flow of the conversation, you've got the skill. 6. Lesson 4.2 - Emotional arc tracking: In a lot of research notes, we accidentally create a flat record, a pile of quotes, a list of observations, and then later, we try to build meaning from it. But the thing we most need for insight is often missing the shape of the experience, the moments where energy rises, the moments where it drops, the points where someone gets annoyed, relieved, uncertain, excited. This is what emotional arc tracking captures. It's a lightweight way to track energy shifts without interrupting the conversation. So what is an emotional arc? It is a tiny timeline of energy shipped across an experience, and it helps you answer some really important questions. Where did the experience start to work? Where did it start to break? What happened right before the drop? And what did the person need in that moment? Those before moments are where the real opportunities hide. To capture this, we use four simple marks. Up arrow means energy is rising. A down arrow means energy is dropping. A right arrow means things are neutral or steady, and an exclamation mark means there's a frustration spike. This is all you need four symbols. The simplest way to use them is to track the arc against a sequence. That sequence could be a user journey like onboarding, then first action, then setting. Be the sections of your interview guide. Or it could be a task flow like search, compare, check out. Here's a practical tip. Print your discussion guide or journey map and leave a margin on the right side of the page. You're not writing a paragraph. You're just marking the current step with one of those four symbols. So here's what it looked like in practice. Onboarding, neutral, permissions, energy drops. Dashboard, energy goes up, Settings, frustration spike, confirmation, energy recovers. Now, for each of those steps, you add one sentence that captures the trigger, not a quote, not a paragraph, just one sentence that names what caused the shift. For example, next to permissions with a down arrow, you mic write felt uneasy about what access is being requested. And next to settings with an extraomation mark, got irritated because the option was buried and the labels were unclear. Exton dash forward with an up arrow, relieved because the information was immediate and easy to scan. If you can name the trigger, you can design for it. Now, there are a few common mistakes to watch out for. The first one is confusing emotion with opinion. If someone says it's bad, that is an opinion. That is not an arc. But if energy drops when they hit a specific step, that is an arc. We're tracking shifts, not judgments. The second mistake is tracking too many micromoments. Start with four or six steps, not 20. You want the big shape, not every tiny fluctuation. And the third mistake is assuming you know the cause. We what happened right before the shift and treat your explanation as a hypothesis, not as a fact. The single question that makes all of this useful is what happened right before the down arrow or the exclamation mark? Once you have three to five arcs across different participants, you can start to pattern age. Do the drops cluster around the same step? Are the frustration spikes always caused by the same kind of trigger, like unclear tone or bad timing, uncertainty, or hidden settings? And do the ups happen when people regain control, clarity, or speed? I give you a map of leverage points, and it also gives you a clean hand off to AI. Later, when you have a transcript, you can ask AI to explain the moment right before each shift and pull out the evidence for you. Here's an example of what that prompt might look like. You tell the AI that you tracked an emotional arc using the four symbols. Then you ask it to do three things for each down arrow and each exclamation mark. Quote what was happening right before it, infer what need or concern is showing up and offer two alternative explanations, and you ask her to return the results as a table. This is a great way to go deeper in the moments that matter most without having to reread the entire transcript yourself. Alright, now it's your turn. Pick a recent experience you can remember clearly. It could be signing up for an app, booking a ticket, or setting up a new tool. Write a four to six step arc. O line per step. Mark each step with one of the four symbols and add one sentence that captures the trigger. If you can do this in under 2 minutes, you are training the skill we want pattern recognition across time. 7. Lesson 4.3 - Question Cascade: One of the hardest parts of interviewing is staying present while your brain does what brains do. You hear something interesting, and you immediately generate a live follow up question. If you chase them all, the interview turns into whiplash. If you ignore them all, you miss the best fred. The cascade notation is a simple way to capture your curiosity without derailing the participant. You will use two tags. Q arrow is a follow up that is worth asking right now. Q later is for a follow up you want to return to when the timing is better. If you already use the toolkit notations from lesson four, this will feel familiar. The follow up arrow and arrow are the same move. Follow this fret now. Park and Q later are the same move. Hold the fret and keep flow. And depth is a common sub type of Q arrow. When a vague term appears, you ask for a definition in the moment. Again, you can use whatever works best for you, but these are the principles behind it. So what is this really about? It is about making sure your best curiosity survives the moment. You want two outcomes. You keep the participant talking, and you still collect the follow ups that will deepen the insight. Now, let us talk about how to keep the question short. A good cascade question is usually one line. If it takes two lines, it is probably two questions. Here are a few useful shape. What was happening right before that? Can you give me a specific example? What made you choose that option? What would you have expected instead? Write the shortest version that still points to the missing detail. So how do you decide whether to ask now or park it for later? Use QR when the participant is already near the moment you care about and use later when asking now would interrupt a story that is unfolding. When it would send you sideways into a different topic or when it requires contexts you do not have yet. The simple rule is this, if the follow up will make the current answer clearer, ask now. If it opens a new branch, park it. And one practical habit that really makes this work, leave three to 5 minutes at the end of the session to circle back to your later list, treat it as a mini closeout section. Let's see what this could look like in an actual interview. Could you just walk me through how you typically use the app? Yeah, so I usually in the app, I usually try to just kind of check where I'm at. You know, I don't check in daily, and I don't even check in weekly to be completely honest with you. I truly just put my money aside and then maybe check in if I remember, like, I've got the account. So that's what I use the app for most of the time, kind of checking in, see how it's doing. And I go off of, like, is it more red than blue? Because if it's more red, then I may need to call my dad and see what we could do going forward. So I suppose the color kind of helps me with understanding. So, yeah. And when you say more red than blue, what does that mean to you? Like, what do you think the colors are telling you? I don't know the exact numbers behind it. It's more like if it looks red, then I'm like, Okay, something is wrong, or, like, in a good spot. And if it's blue, I'm like, Okay, we're fine. It's kind of like a quick vibe check. And I know that sounds silly, but it's just easier than trying to, like, read everything. You know what I mean? Mm hmm. Got it. Yeah. And what do you do when you see red? So if I do call my dad, and he says, like, Let's move some money, let's do something different, I tend to go on the website than I do on the app. Mm hmm. Yeah, I think maybe it's because my dad knows more about the website than the app. Mm hmm. I know to use the app, but he knows how to use the website more. Mm hmm. Yeah, and maybe that's kind of how I've learned from it. So I generally feel better using the website because I know what I'm doing. Mm hmm. Mm hmm. Okay, so the app is mainly for checking in, and the website is for actually moving money. When you're checking in on the app, what's the moment you're hoping for? Like, what is good news? I mean, good news is it's blue. And then I'm like, Okay, I don't have to deal with it because honestly, money stuff is it's stressful. So if it's blue, I can just, like, carry on with my day. But if it's red, then I'm like, Oh, now I have to do the whole thing. And the whole thing is like calling my dad, going on the website, trying not to mess it up. Yeah. Mm hmm. Mm hmm. That makes sense. Okay, before we wrap, I want to circle back to something you said earlier. You mentioned you feel better using the website than the app when you need to move money. What is it about the website that makes you feel more confident? Um, it's honestly just familiarity. Like, I've watched my dad do it, and on the app, I'm always a bit worried I'll tap the wrong thing. On the website, it's, like, slower and I can see more, and it feels more efficient. Yeah. So I trust myself more there. Okay. Yeah. Notice what happened. I used deep to clarify a term that mattered to their decision making. I used Q later to park a fred so they could keep talking. Then I left time at the end to circle back and get a clean answer. Now, let us talk about how this connects to AI after the session. During the interview, QRO and Q later are flow tools. They help you stay present. But after the session, AI becomes useful for a different job, organizing the questions you did not fully answer and turning them into a plan for the next round. This matters because some Q later questions will get asked later in the same interview, but some will not. They get lost when time runs out. When the participant takes you somewhere more valuable or when the question needs context you did not get. This is exactly where AI helps. So after the interview, you can use AI for three things. First, for questions you parked, but never got to ask. Second, for questions you did ask, but did not get a clear answer to. And third, for turning repeated Q later themes across sessions into probes for the next round. You can skip AI for any Q later questions that you already asked and got a clear answer to in the same session. That one is done. Here's an example of what that prompt might look like. You give the AI your notes marked with QR and Qlater. Then you'll ask you to first mark which questions were answered in the session and which were not. Or the unanswered ones, you ask it to group them into three to five themes, rewrite each question to be neutral and specific and suggest an order that keeps flow. And you ask it to output a short follow up guide you can use in the next interview. This turns your raw curiosity into a structured plan without you having to do the organizing yourself. Alright, now it's your turn. Do a five minute role play. After it ends, write two Q RL follow ups you would ask right away, and two Q later follow ups you would return to later. Keep each question to one line. If you can do that consistently, you will feel a difference in your interviews within a week. 8. Lesson 4.4 - Context anchors: I Transcripts make interviews look cleaner than they were. They capture what was said. They often miss what shaped what was said. Context anchors are how you keep that missing layer. They are short bracket notes that capture constraints and changes. Things like where the person is, what else is happening around them, what device they're on, who is nearby, and what is competing for their attention. These details are not nice to have. They explain why a participant gives a short answer, avoids a feature, changes tone, or becomes cautious. So why does this matter? Because if you do not capture context, you can misread the data. Might label someone unengaged when they are actually exhausted. You might label someone confident when a partner is coaching them off camera. You might label something easy when the person is on a desktop in quiet conditions. A transcript cannot reliably show any of that. Transcripts strip away the environment and interruptions. They strip away device and setup constraints. They strip away social dynamics like who is present. They strip away timing and energy state, and they strip away the reason the pace changes. What counts as context and what does not? A good context anchor has two qualities. First, it is objective or close to objective. And second, it plausibly changes behavior or meaning. Here are some good examples. On found walking to work. Screen glare cannot read small text. At work, keeping voice low. Child interrupts twice, loses freight. Partner in room, answers get shorter. Switch from app to website mid task. What does not count as a context anchor? Opinions, like, This is annoying. Capture those elsewhere. Interpretations like they are embarrassed. Treat those as a hypothesis, not context. The rule of thumb is this, write context that would help a teammate understand why this moment looked the way it did, and you want to write them quickly, keep each one to a single line. There are two useful shapes. The first is the constraint, then the likely effect. For example, night shift, walk up, slower pace. The second is the change moment, then what changed in the conversation? For example, partner enters room becomes more formal or on mobile, one handed avoids typing. Now, when you get to analysis, context anchors help you interpret patterns without flattening them. You can ask questions like, do the same issues show up under the same constraints? Are certain frictions mobile only? Do confident answers correlate with someone else being present? Do emotion shifts align with interruptions or time pressure? And this keeps you honest, and it also makes your prompts to AI more precise. The main value of context anchors is simple. They give AI and future information the transcript will not contain. Sometimes that context will matter. Sometimes it will not. The goal is to reduce the chance of a confident sounding misread. When you use context anchors with AI, there are three things to keep in mind. First, include the relevant contexts before you ask for interpretation. Second, ask AI to flag where the context could be shaping the data. And third, ask for alternative explanations when the context is ambiguous. Here's an example of what that prompt might look like. You give the AI your interview notes along with a few contact anchors. Then you ask it to use the contacts to avoid over interpreting short answers, hesitations, and topic changes. For each context item, you ask it to list one or two ways it could be shaping behavior, one or two ways it might be irrelevant, and to label every inference with a confidence level, high, medium or low. You ask her to quote exact lines for any claims. This forces the AI to reason carefully instead of just pattern matching on the surface. Let me give you a quick example of how this works in practice. Say the participant says, Yeah, I mean, it's fine. I don't really use that. Your note might be context in open plan office whispering short answers. Later, when you interpret that it's spine, you read it differently. Maybe they were not dismissive, maybe they were just keeping their voice down. The point is to understand the conditions the data came from. So you and AI do not over interpret what happened. Alright, now it's your turn. Do a short roleplay and capture three context anchors. One for an environment or interruption, one for a device or setup constraint, and one for a social dynamic or timing factor. Keep each one to a single line in the bracket format. If you can spot these in real time, you will be capturing a layer of insight that most researchers mith entirely. 9. Lesson 4.5 - Observations vs interpretations: There is a habit that creeps into almost every researcher's notes, and it is so natural that most people do not even notice they are doing it. You see someone hesitate and you write confused by the layout. You see someone tap quickly and you write found it easy. You see someone sigh and you write frustrated. Every one of those is an interpretation dressed up as a fact. And the problem is not that the interpretation is wrong. It might be right. The problem is that once it is written as a fact, nobody questions it, not you, not your team, and definitely not AI. This lesson is about building a simple habit that protects your rigor. You are going to learn to separate what you saw from what you think it means using two prefixes, O for observation and I for interpretation. An observation is something you could point a camera at and everyone would agree on. The participant hesitated for 6 seconds before tapping B. That is an observation. Anyone watching the recording would see the same thing. An interpretation is your meaning making. Money anxiety, fear of misclick, decision fat. Those are interpretations. They might be accurate, but they're not facts. They are hypotheses. The habit we're building is simple. When you write a note, ask yourself, could someone else see this exact thing on the recording? If yes, it is an O. If you're adding meaning, it is an I. This matters more than you might think, and here's why. When you hand notes to AI and ask it to find pageants, AI treats everything you wrote as equally true. It does not know which lines are the things you saw and which lines are the things you guessed. So if your notes are full of unmarked interpretations, the AI will build its analysis on top of your guesses and present them back to you as confident findings. That is how overconfident misreads happen, and they are hard to catch because the output looks polished. Separating O from I gives you a clean foundation. The observations stay solid, the interpretations stay testable. So let's talk about how to write a clean observation. The key is to keep your language descriptive and specific. Describe behavior, timing, and sequence. Avoid adjective that carry judgment. Here are a few examples. Hesitated 6 seconds before tapping by that is clean. Scroll past the pricing section without stopping. Clean. Read that error message out loud, then close the tab. Also clean. Now compare those two. Was confused by the pricing. That is an interpretation. You do not know they were confused. You know they scrolled past it, the confusion is your guess. A good test. If you catch yourself writing a feeling word, like confused, frustrated, delighted, or overwhelmed, pause and ask what you actually saw. Describe that instead and move the feeling word to an eyeline. Now, how do you write a good interpretation? The most important thing is to treat it as a hypothesis, not a conclusion. Write it as something that could be tested or disproved. For example, Oh, hesitated 6 seconds before tapping B. I, possibly money anxiety or unsure whether the item is correct. Notice what happened there. The observation is locked in. The interpretation offers two possible explanations, and neither one claims to be the answer. You can also give your interpretation a confidence level, something like I likely comparing to pricing they saw elsewhere, medium confidence. This helps future you and your team know how much weight to put on it, and it helps AI to treat it with the right level of caution. Here's what a real set of notes might look like using this system. Observation hesitated for 6 seconds before tapping by. Interpretation money, anxiety or fear of miss click. Observation scrolled past the pricing table without stopping. Interpretation may not have recognized it as pricing or already decided in skipping detail. Observation said, wait, where did that go? After the page transitioned. Interpretation, unexpected navigation possibly felt like loss of control. Observation smiled and said, Oh, nice when the confirmation screen appeared. Notice that last one has no interpretation line. That's fine. Not every observation needs an interpretation. Sometimes the behavior is clear enough on its own, and adding a guess would just be noise. A good ratio to aim for in your notes is roughly three to four observations for every one or two interpretations. That keeps your notes grounded. Now, let's talk about how this connects to AI. When you give AI a set of notes with clear observations and interpretations, clear O and I labels, you can ask it to do something very specific. Work with the observations first and then compare its explanations to your interpretations. This creates a useful check. If the AI's reading matches your interpretation, that is a signal your guess may be on track. If the AI offers a different explanation, that is worth investigating. Either way, you end up with a more honest analysis. Here's an example of what that prompt might look like. You give the AI your notes with the O and I labels. Then you ask it to first analyze only the observation lines and suggest two to three possible explanations for each without looking at your interpretations yet. After that, you ask it to compare its explanations to your I lines. Where do you agree? Where does the AI see in different possibility? And you ask it to flag any interpretation that feels like a stretch given the observation evidence. This is a powerful way to use AI as a thinking partner rather than a confirmation machine. Alright, and now it's your turn. Think of a recent experience you had with a product or service. It could be anything an app, a website, a self checkout machine, a booking flow, and then write four to six lines total. Aim for three to four observations and one to two interpretations. Remember, observations pass the camera test. Interpretations are your best guesses written as hypothesis. If you can do this quickly and consistently, you're building a skill that will make every analysis you do more trustworthy, whether you're working with AI or without it. 10. Lesson 4.6 - Contradiction mapping: Alright, this will shock you, but people do not always do what they say they do. That is not a judgment. It is one of the most reliable patterns in UX research. Someone says a feature is easy, but they avoid using it. Someone says they do not care about a setting, but they spend 2 minutes configuring it. Someone says, I trust this app, but their voice gets quieter and they start hedging. These gaps between what a person says and what they actually do are some of the richest materials you will ever collect, and they are easy to miss because transcripts flatten them. A transcript gives you the words. It does not always show you the behavior that contradicts those words. This lesson teaches you a simple way to catch and record these contradictions in real time using two tags says and does. When you write a says and does pair, you are capturing attention, and tensions are where insight lives. They point to unmet needs, workarounds, learned helplessness, social desirability or simply a gap between how someone thinks about their experience and how they actually live it. The important thing is that we're not trying to catch people in a lie. That is not what this is about. People are not being dishonest. They are being human. We all narrate our own behavior in ways that are slightly tidier than reality. Your job is to notice the gap and hold it with curiosity. There are two kinds of contradictions worth paying attention to. The first is a behavior contradiction. This is when what someone said and what they do are misaligned. For example, says, I always check my budget before buying. Does tabs buy without opening the budget screen. That is a clear behavior contradiction. The second is a tone contradiction. This is subtler. The words sound fine, but something in the delivery suggests otherwise. For example, says, It's fine, I don't mind. But the voice is flat, the pace slows down, or they immediately change the subject. The words say one thing, the energy says something else. Both kinds are worth capturing. Behavior contradictions are easier to spot. Ton contradictions take more practice, but they often point to the moments people are least comfortable talking about directly. How do you write these in your notes? Keep it simple. Write the says line first, then the does line right underneath. O line each. Here are a few examples. Says is really easy to use. Does avoids the feature entirely and ask the family member instead. Says, I check this every week. Does app usage data shows last login was three months ago. Says, I don't care about the design. Does spends 45 seconds adjusting the theme colors. Says? Yeah, that makes sense. Does rereads the same paragraph three times. Notice that you're not explaining the contradiction. You're just recording both sides. The explanation comes later. Now, there's an important skill here that takes a little practice, keeping your notes curious rather than judgmental. It's tempting to write something like claims it is easy but clearly cannot use it. That sounds like you're catching the participant out. And that is not the energy we want in our notes because it closes down thinking. Once you clearly cannot use it, you have already decided what is happening. Instead, write the contradiction as a pair and leave the meaning open. Says, it's easy, does, avoids it, that is enough. Tension is visible. You can explore it later. If you want to add a note about what the contradiction might mean, use a question rather than a statement. Something like mismatch. Is this social desirability, or do they genuinely experience it as easy but choose not to use it for a different reason? That keeps you in inquiry mode. When you get to analysis, contradictions are powerful because they cluster. If three out of five participants say a feature is easy, but none of them use it unprompted, that is a pattern worth investigating, and it is a pattern that a simple sentiment analysis would completely miss because the words are all positive. This is also where the connection to the previous lesson comes in. If you have been separating observations from interpretations, you can now lay out contradiction mapping on top. The says line is close to a quote. The does line is an observation. And the question you ask about the gap is an interpretation held lightly. Now, let's talk about how this works with AI. The main value of CS and DS pairs for AI is that they give the model a specific kind of tension to work with. Instead of asking AI to find insights, which is vague, you can ask her to focus on the mismatches and generate questions from them. This matters because AI is very good at summarizing what people said. It is much less good at noticing when behavior contradicts those words, especially if the contradiction is subtle. By flagging the contradictions yourself, you're giving AI the most interesting material to work with. Here's an example of what that prompt might look like. You give the AI your notes with a says and does pairs. Then you ask you to describe the tension in one sentence for each mismatch. Suggest two to three possible explanations framed as questions and flagged whether it is a behavior contradiction or a tone contradiction. You tell it not to resolve that tension, but to present it as an open question. And then you ask it to look across all participants and find where the same contradictions repeat. This gives you a map of where the real friction lives, not just where people said they had friction. Alright, now it's your turn. Think of a refund conversation or interaction where someone said one thing but did another. It does not have to be a research interview. It could be a friend recommending a restaurant they never go to or a colleague saying a process is straightforward while spending an hour working around it. Write two to three contradiction pairs. One line for says, one line for does, keep it factual, keep it curious, and resist the urge to explain the gap. If you can slot these in real time during an interview, you will be capturing the kind of insight that no transcript and no AI summary can give you on its own. 11. Lesson 5 - Working with AI in partnership: Lesson, we're going to talk about AI in a really practical way, not the hype and not the fear, how do we use it as solid researchers? Here's the headline. AI is not the boss. We are. AI can help us move fast. It can help us get unstuck, and it can help us stress test our thinking. But it can also do this thing where it sounds incredibly confident while quietly making things up. And if we're not careful, it can nudge us into conclusions that feel tidy and convincing, but that aren't actually true. So our goal is not to get answers from AI. Our goal is to use AI to speed up the work, to keep the quality bar high. And for the rest of this course, we're going to use one simple loop, draft, critique, verify, and document. That's it. That loop is what lets us use hechPT, Claude, Gemini, whatever is next without tying our skills to one tool. And I want you to hold on to one metaphor. We keep our hands on the steering wheel. AI is the satnav. Alright, let's start with the most useful question of all. What even is an LLM. An LLM, a large language model, is basically a system trained to predict the next word. It has seen a massive amount of text, and it has learned patterns of how humans tend to write. So it can produce language that feels fluent, coherent, and honestly, sometimes a bit spooky. Here's the key bit. Fluent does not mean true. LLMs are great at the shape of an answer. They can draft, summarize, reformat, brainstorm, and help us spot possible patterns. And for research work, that's genuinely useful because so much of our job is turning messy inputs into something clearer. What they can't do automatically is know what's accurate for your project. They don't know your project. They don't know your users. And they definitely weren't in the room with your participants. So the best mental model is a very fast collaborator. Helpful, creative, sometimes surprisingly sharp. Not a witness and not a source and not the person we quote in a stakeholder meeting. And once we really get that, a lot of the confusion and hype around AI starts to calm down. Now, if we're going to use AI, we need to know how it fails because it fails in pretty predictable ways. There are three classic ones that I see all the time in research work. Number one, hallucination. That's the polite word for it made something up, a feature that doesn't exist, a quote that no one said, a neat little key insight that sounds plausible but isn't actually grounded in anything you actually collected. Number two, overgeneralization. This is the two people mentioned it, so it must be a universal truth problem. Suddenly, we got users hate on boarding or everyone is confused by pricing. Maybe. Or maybe it was just two people with a specific context on a specific day. Research is all about context, and AI will flatten that context if we let it. Number three, confident tone. This one is sneaky. Even when the model is just guessing, it can sound calm, certain, and authoritative. When you're moving fast, it's really easy to mistake confidence for correctness. So the fix isn't become a prompt wizard. The fix is to use a workflow that makes it hard to accidentally believe something that isn't true. And that's where our loop comes in. So let's do it. Alright, so here's the thing that we're going to lean on for the rest of the course. It's simple, it's repeatable, and it keeps us honest. It's this loop, draft, critique, verify, document. The reason why I love it is because it stops us treating AI like an answer machine. Instead, we use it like a power tool. Fast, helpful, and still something we're responsible for. Let's walk through it. Draft. We use AI to get a first version on the page, not because it's perfect, but because it gives us something to react to. This could be a theme list, a report outline. Anything that normally starts with a blank page. Step two, critique. Now we switch gears. We ask the tool to critique what it just made. What's missing? Where is it vague? What's biased or leading? What assumptions has it smuggled in without telling us? This is where we turn a nice output into a useful draft. Step three, verify. And this is the part that makes it research. We check the output against reality, against the brief, against our notes, against transcripts, against actual quotes. If the tool makes a claim, we ask, Where's the evidence? And if we can't back it up, we don't ship it. Step four, document. Finally, we write down what we decided and why, what we're confident about. What's still a hypothesis? What's unknown? This is how we keep our work defensible, especially when we're moving quickly. Just to be really clear, this loop isn't a one off thing. It's basically the spine of our whole process. If something matters, we attach evidence. That's the quality bar. So that's the foundation. We use AI to move faster, but we stay accountable for quality. We don't outsource judgment. We use the loop, draft, critique, verify, document. If you remember nothing else from this lesson, remember this. When something matters, we attach evidence. That's how we stay useful and that's how we stay trustworthy. 12. Lesson 5.1 - Prompting that respects rigor: At this point in the course, you have built a set of notation habits. You know how to track emotional arcs, capture context, separate observations from interpretations and flag contradictions. Now we're going to talk about what happens when you hand those notes to AI, because the way most people use AI with research data is paste in a transcript, type something like, what are the key insights, and see what comes out. And the output usually looks good. It is well structured, it's confident and easy to read. The problem is that easy to read and accurate are not the same thing. You give AI a vague prompt, it fills in the gaps with pattern matching. It guesses what you probably want, and it presents those guesses as findings. This lesson is about giving AI a specific job instead of a blank canvas. You're going to learn a small set of prompt patterns that test your thinking rather than replace it. The core idea is this, a good research prompt tells AI what to look for, what format to return, and what not to assume. It treats AI like a research assistant who is fast and thorough but has no judgment. You are the one with judgment. The prompt is how you direct the assistance effort. We're going to cover four prompt patterns. Each one is designed for a different moment in your analysis, and each one is built to keep you honest. The first pattern is confirm and contradict. This is for when you already have a hypothesis and you want to pressure test it. The shape of the prompt is, here's my hypothesis, find evidence that confirms it and evidence that contradicts it. Quote the exact lines. This is powerful because it forces the AI to look in both directions. If you only ask, does the data support X, the AI will almost always say yes and find supporting quotes. That is confirmation bias built into the prompt. By asking for both sides, you get a more balanced picture. Here's an example. My hypothesis is that participants feel more confident on the website than in the app. Find evidence that supports this and evidence that contradicts it. Quote exact lines from the transcript. The second pattern is retrieve evidence. This is for when you need to back up a claim with specific data. The shape is quote the exact lines that support X. Do not paraphrase. This is useful when you're writing up findings and you need to anchor your claims in real data. It is also a good check on yourself. If the AI cannot find a direct quote, that might mean your claim with an interpretation rather than a supported finding. One important detail. Always tell the AI not to paraphrase. If you leave that out, it will often reword the quotes to sound cleaner, and then you lose the participant's actual language. The third pattern is counterexamples. This is for when you have a theme forming and you want to check whether it holds. The shape is, I see a theme around eggs, find moments that do not fit this theme. What is different about them? This is one of the most valuable prompts you can write because themes are easy to overapply. Once you see a pattern, your brain wants to see it everywhere. Asking for counterexamples forces you and the AI to look at what does not fit. And the exceptions are often where the real nuance lives. The fourth pattern is compared early versus late. This is for when you want to understand how someone's experience or attitude shifted over time. The shape is compare what the participants said and did in the first half of the session to the second half. What changed? What stayed the same? Connects directly to the emotional arc tracking you learned earlier. But instead of you doing the comparison manually, you're asking AI to pull the evidence from both halves and lay it out side by side. This is especially useful for longer sessions where it is hard to hold the whole arc in your head. So those are your four patterns. Now, let's talk about what goes wrong because there are a few common failure modes that are easy to fall into. The first one is the leading prompt. This is where your prompt already contains the answer you want. Something like explain why the participant found the checkout process frustrating. That prompt assumes frustration. Better version would be, what was the participants experience during checkout? Quote relevant lines and note any shifts in energy or tone. The second failure mode is the vague prompt, something like, what are the key insights or summarize the important findings. Now, these sound reasonable, but they give AI no direction. The output will be generic and confident, which is a dangerous combination. And the third failure mode is the prompt that asks AI to feel. Something like, how did the participant feel about this? AI does not know how someone felt. It can only pattern match on language. A better version is, what did the participants say and do during the moment? List possible interpretations. The last thing I want to cover is how to store your prompt so you can reuse them. Once you find a prompt that works well for your research, save it. Keep a simple document or note with your go to prompts. You can organize them by stage, prompts for during analysis, prompts for after synthesis, prompts for writing up findings. Over time, this becomes your personal prompt library, and the beauty of it is that each prompt encodes a research habit. Find counter examples is a habit. Quote exact lines is a habit. Compare early versus late is a habit. The prompts are just the way to make those habits consistent and repeatable. Let me leave you with one all purpose prompt that brings several of these patterns together. You can copy this and use it as a starting point. I noticed g. Find evidence that confirms it, contradicts it, and what I might be missing. Quote exact lines. Do not paraphrase or infer emotions. If the evidence is ambiguous, say so. This prompt does four things at once. It starts with your hypothesis. It asks for evidence in both directions. It demands exact quotes, and it gives the AI permission to say, I'm not sure. Which most prompts do not do. The last part matters because AI defaults to confidence. If you do not explicitly tell it that ambiguity is okay, it will resolve every uncertainty into a clean, sounding answer. Alright, now it's your turn. Pick one of your own notes from a previous lesson. It could be an observation, an interpretation, a contradiction pair, or an emotional arc. Write two prompts for it. One should use the confirm and contradict pattern. The other can use any of the four patterns we covered. Keep each prompt to three or four lines. If it is longer than that, it's probably doing too many things at once. If you can write a clear specific prompt in under a minute, you're ready to use AI as a real thinking partner in your research. 13. Lesson 5.2 - Evidence trail workflow: Now the bit that keeps us honest. When people say AI analysis, what they often mean is paste the transcript, ask for insights, and hope for the best. And yes, you'll get an answer. It just might be a beautiful paragraph that you can't defend. So we do something slightly more disciplined. We build an evidence table. If you've never done this before, here's how it works. We're making a little bridge between the raw transcript and the insights that we're willing to put our name on. And the rule is quote first, meaning second. So a row in an evidence table might look like this. What I love about this is that it's simple. It doesn't require a special tool, and it stops you from accidentally writing a report that's 90% vibes. Also, notice what we're not doing yet. We're not trying to summarize the whole interview into a grand theory. We're just collecting the pieces of evidence that will matter later. And now we can use the AI like what it is, a fast assistant that can read and summarize. It's totally fine to ask questions like, how did this participant do X? What did they struggle with? What confused them? What did they expect to happen? Long as we add guard rails that make the output usable in synthesis. So here are my favorite guardrails. Anter in a table. Every claim must include a supporting quote. Separate what happened versus what it might mean. And if the evidence is weak, say that explicitly. Yes, you can have a bit of fun. You can ask. What would you rename these features based on the participant's mental model? Or write the frustration moment as a one line story. Just don't confuse this creative output with evidence. It's just an aid for thinking, not a finding. Alright, confidence. When we put a confidence level next to a finding, we're not trying to sound scientific. We're just doing something much simpler. We're telling the truth about how solid the evidence is because in research, there's a huge difference between this person had a rough moment, and this is a reliable pattern that should influence product decisions. Confidence is our way of signaling that difference. So here's a human way to think about it. When you read a quote, ask yourself. Did they say it clearly? Did they give a concrete example? Did it connect to the thing that we're actually researching? If the quote is specific and unambiguous, that's already a good start. And then ask is this just one moment or did it show up more than once? That more than once could be the participant repeated it in different words in the same interview. Or you've heard the same thing from other participants, or you can see something similar in behavioral data. So drop off rage clicks or support tickets. So a simple confidence scale can be high confidence means this feels solid. The evidence is clear, and it's not hanging on one fragile quote. Medium confidence means this looks real, but I want to confirm it. It might depend on the participant type, the scenario, or the specific flow we tested. Low confidence means interesting, but I'm not ready to build decisions on it yet. It could be a misunderstanding, a one off or just not well supported. And here's the secret weapon. If you want to sound credible without being boring, add one more sentence, what would raise confidence? For example, we hear this from two more participants, this becomes high. If analytics shows drop off at this step, this becomes high. If we test the revised UI and the confusion disappears, we can close this. Limitations are just your honesty clause. They are the reasons a smart stakeholder shouldn't overgeneralize like small sample, transcript quality, unusual participant, very specific context. Stating limitations doesn't weaken your research. It stops someone else from misusing it. Alright, let's make this real. Pick one interview transcript. Pick two research objectives. Now build yourself an evidence table with the help of AI. Let's aim for about ten quotes total. For each quote, we want one sentence on what it means and a confidence level, so low, medium or high. And then do the part that turns it into actual research, right, two candidate insights, and under each one pays the two to three quotes that support it. If you can do that, you basically learned the core skill. Everything else is just scaling it up and staying honest while you do. Good luck, and I'll see you in the next lesson. 14. Lesson 6 - Calibration: making partnership real: Throughout this course, you have been building two kinds of skill at the same time. One is your ability to take sharp, structured notes during a research session, and the other is the ability to use AI as a thinking partner after the session. This lesson is about putting those two skills side by side and seeing where they overlap, where they diverge, and what that tells you about your own patterns as a researcher. We're going to use a simple tool called the calibration matrix. It is a two by two grid, and the purpose of it is not to score yourself or score the AI. It is to help you see your blind spots so you can get better over time. Here's how it works. So the horizontal, the X axis is about you. And on one side, what you noticed and on the other side, what you did not notice. The Y axis in the vertical is about AI. On one side, what AI picked up, on the other side, what AI missed. That gives you four quadrants. Now, let us walk through each one. The first quadrant is shared ground. This is where you and AI both notice the same thing. For example, you both picked up that the participant dislikes long quizzes. This is reassuring. It means your observation is well supported and unlikely to be a stretch. The second quadrant is AI caught, and this is where AI noticed something that you missed. Maybe AI flagged that the participant used the word trust six times during the session, and you did not pick up on the repetition. This quadrant is where AI earns its keep. It is good at counting, spotting repetition and catching patterns across large amounts of text. The third quadrant is my edge. This is where you noticed something that AI missed entirely. Maybe you caught that the participants tone was sarcastic when they said, Oh, yeah, that was easy. AI read the words at face value. You read the subtext. This is your human advantage, tone, body language, context, and the subtle things that do not show up in a transparent. And the fourth quadrant is the blind spot. This is where neither you nor AI noticed something. By definition, you cannot fill this one on your own, but you can start to populate it over time by comparing your matrix with a colleague's analysis of the same data. Or by revisiting sessions after a gap and noticing things you missed the first time. Now, the quadrant I want you to pay the most attention to is the third one, my edge, because it contains an important question. When you notice something that AI missed, it could be your edge. You spotted a real signal that the machine could not detect. That is valuable. But it could also be your bias. You read something into the data that is not really there. And AI did not confirm it because there was nothing to confirm. The honest question to ask yourself is, is this my edge? Or is this my bias? And the way to answer it is to look for evidence. Can you point to a specific observation, a quote, a behavior? If yes, it is probably your edge. If your evidence is mostly a feeling or a hunch, it might be a bias worth examining. This is not about doubting yourself. It is about staying calibrated. Let me walk you through a worked example using the banking app interview from earlier in the course. So you interviewed Alex about the banking app with the red and blue colors. You took your notes during the session. After the session, you ran the transcript through AI and asked for a summary. Now you compare your notes to the AI summary and build the matrix. Shared ground, you both notice that Alex uses color as a quick decision shortcut. Red means something is wrong, blue means everything is fine. This showed them clearly in the transcript and in your notes. AI caught it. AI flag that Alex mentioned calling the dad freight separate times in the session. Noted it once, but you did not track the repetition. The frequency suggests this relationship is more central to Alex's financial behavior than you initially thought. My edge, you noticed that Alex's tone shifted when talking about using the app to move money. The words were neutral, but the energy dropped. You marked this in your emotional arc. AI summarized the section as prefers website for transactions, but missed the emotional weight behind it. Blind spot. After comparing with a colleague, you realized that neither of you explored why Alex checks the app only when remembering the account exists. There might be a notification or a trigger design opportunity there that nobody probed. Now, for each quadrant, you want one clear action. For shared ground, move the finding forward with confidence. It is well supported. For AI cord it, verify the pattern, go back to the transcript and check whether the repetition is meaningful in context or just a speech habit. For my edge, protect the observation, write it up with evidence so it does not get lost. This is the kind of insight that makes your work distinct. For blind span, add it to your question list for the next round. This is where future research should probe. Now let's talk about how to generate this matrix using AI. The prompt is straightforward. You give AI your notes and ask it to compare them against its own summary of the transcript. So here's how it works. You give AI the transcript and your notes. You ask it to first produce its own summary without looking at your notes. Then you ask it to compare the two and output the matrix with two bullets per quadrant. For the M edge quadrant, you ask it to suggest whether each item is more likely your edge or your bias, and to explain why. Give you a structured self check that takes about 5 minutes and gets more valuable every time you do it. That brings us to the last part of this lesson, making this a habit. If you do research on a regular cadence, whether that is weekly discovery interviews, fortnightly usability tests or monthly stakeholder conversations, the calibration matrix becomes a ritual. After each session or after each batch of sessions, take 5 minutes to build a quick matrix. You don't need to be exhaustive. Two bullets per quadrant is enough. Over a few weeks, you will start to see your own patterns. Maybe you consistently miss repetition, maybe you're great at catching tone shifts. Maybe AI keeps surfacing word frequency patterns that you overlook. That is useful self knowledge, and it keeps the partnership honest. You're not outsourcing your thinking to AI. You're not ignoring what AI offers. You are calibrating. Alright, and now it's your turn. Take a short transcript or use the one that I provide. Run it through AI and get a summary. Then compare the AI summary to your own notes. Build a two by two matrix with one or two bullets per quadrant. For the M edge quadrant, ask yourself honestly, is this my edge or could it be my bias? You can do this in under 10 minutes, you have a calibration habit that will sharpen your research skills for as long as you practice it. 15. Lesson 7 - Ethics + responsibility: Is about trust. Before we get into tactics, let's set the tone. Using AI in research can be genuinely helpful. It can also go wrong in very boring, very human ways. A participant shares something personal, and it ends up somewhere where it shouldn't stakeholder reads an AI summary and assumes it's the truth. A quote gets cleaned up a little too much, and suddenly it's not really a quote anymore. So, this isn't an ethics lecture. This is a set of habits that protect people and protect your work. And we're going to keep it practical. We talk about what to share and what not to share. What to disclose and how we say honest when we're moving fast. So hands on the steering wheel, AI can help, but we decide what makes it into the research. Here's a simple mental model. What can go wrong? To keep this practical, we're going to use a few plain categories. It's not just because we love frameworks. I mean, we do. It's because it's easier to spot risks when you have buckets. So here are the big ones privacy and security. Are we exposing personal or confidential data? Transparency. Are we being honest about AI? Fairness? Are we missing or distorting certain groups experiences? Accountability. Who is responsible if something is wrong? Human oversight. Are we letting the tool decide or are we deciding? If you can remember those, you'll catch 90% of the problems before they happen. Let's start with the one that causes real damage. A. Here's a rule that'll save your career. If you wouldn't post it on red it, then don't paste it into a random AI chat. So here's some examples of what not to paste. Names, emails, phone numbers, and addresses. You'll hear the term PII used a lot in the industry. That means personal identifiable information. So don't paste raw session links that include someone's name, or any medical, financial, or highly sensitive personal info. No clients strategy, unreleased roadmaps, or internal intellectual property. Just don't paste anything that your company or your client has not approved for external processing. Yes, you can still use AI. You just need to work with a redacted transcript or a safer tool. And this is why the redaction habits from Lesson six matter. Now let's talk about disclosure. Disclosure is where people get weird. They either hide AI use because they're nervous or they overshare in a way that makes participants feel like they're being watched by robots. So let's make it simple. Participants need to know. Will AI be used at all? How will it be used during the session or after? Will they interact with AI directly? How is their data being protected? Can they opt out of AI involvement? Will their data be used to train models? Say yes or no, clearly. Also, a quick reality check telling people AI is involved can change behavior. Some people will trust you less, some will perform, and some will hold back. So let's frame it in a calm and reasonable way. AI is used to save time on admin tasks like transcription. A researcher reviews everything. You can opt out. And stakeholders need something different. They need to trust the work. So we include a short block in the report. What tools were used? What the AI did, and what humans checked. No drama, just transparency. Now, let's talk about bias. AI has a personality. It likes things neat. It likes things average. And if you're not careful, it will turn a messy set of human experiences and turn it into one smooth story that feels true while quietly sanding off the edges. And that's what we mean by average washing. So let's make it concrete. Imagine you interviewed five people for an onboarding flow. People said it was fine. Two people said it was confusing. One person said, I didn't do it at all because I thought it would share my data with my employer. A lazy AI summary will often come back with something like users found on boarding confusing and wanted clearer guidance. And that's not wrong, but it's also not where the real risk is. That one person's fear about data sharing might be the difference between a minor UX tweak and a trust disaster. Example two, the default user that it invents. If you don't tell the AI who your participants are, it fills in the blanks. It will sometimes assume the default user is confident with tech from the majority culture and using the product in a fairly standard way. So if you're researching a group that is not that say, accessibility needs, low digital confidence or a niche workflow, you have to name it. Otherwise, the AI will gently flatten it. So here's the habit we built. Every time AI gives you a tidy summary, we do a quick follow up. Cool. Now show me the contradictions. So we ask, who had a different experience? What didn't fit? What surprised us? What would be easy to miss if we only looked at averages? And we keep coming back to quotes because quotes are where the nuance lives. Let's talk about human oversight, how we stay responsible without becoming paranoid. This is where we keep our hands on the steering wheel. In practice, human oversight just means we're clear on what the AI is allowed to do for us. So it can help us draft the first pass, organize messy notes, summarize a chunk of a transcript, and suggest possible interpretation. What it can do is take responsibility. That part stays with us. So here's a simple way to work before anything leads to a laptop. So before it goes into a report or a deck or slack message to stakeholders, we do a quick integrity check. Asked, Are the codes real and copied accurately? Can we point to evidence in each insight? Did we accidentally turn a hypothesis into a fact? Did we write down any limitations so someone doesn't overgeneralize? And when the stakes are high, we make it social. We ask someone else to read the outputs, not because we're panicking, but because it's genuinely hard to spot your own blind spot when you've been deep in the work. So, this is the vibe. We use AI to move faster, and then we use human review to stay honest. Reporting integrity. And this is the part where we stop being cute. If a report includes made up quotes, it doesn't matter how well designed the slide deck is. That's not research. It's just storytelling, and it damages trust fast. So here's the standard we hold the quote rule. If it's in quotation marks, it came from word for word. If you paraphrased it, don't put it in quotation marks. Write it as a summary. The composite participant trap. Sometimes people try to be helpful by combining a few participants into one clean persona like story. That's fine for internal sense making, but it's not fine if you present it like a real person said it. So if you ever do a composite, label it clearly. Otherwise, don't the evidence trail. The easiest way to keep this simple is every insight gets a little anchor, a link to notes, a clip, or a set of quotes in your evidence table, not because stakeholders will click every link, but because you should be able to. The AI usage log. And this is the bit nobody wants to do. And it's the bit that quietly makes your work feel professional. An AI usage log is just a small record of what happened. It answers questions like what tool did we use? What did we ask it to do? Did we feed it raw data or redacted version? Did we check it afterward? What did we change? Think of it as a memory aid. And if a client ever asks, How did you analyze this? You don't have to rely on vibes. Also, it protects you, because if the AI output was wrong, you can see exactly where it entered the workflow. Alright. Here's where we actually build the habit. Take your current project and do two small things. First, write your participant disclosure. Keep it short and human. Imagine saying it out loud at the start of an interview. Second, start your AI usage log. Just one entry. Tool, task, what you checked. If you do those two things, you're not just aware of ethics. You're practicing it. 16. Lesson 8 - Project video: The course gives you all the skills and principles. Now let's put it all together in one complete cycle and give you the template to keep doing it on your own. The Capstone project is not a test. There's no pass or fail. It's a practice run of the full workflow from session notes to emdence table using everything you have learned. Here's what the project looks like. You are going to produce four things. The first is a set of post session notes. These are human first notes taken during or right after a session using the notation system you have been practicing. That means emotional arcs, context anchors, observations, interpretations, contradiction pairs, and question cascades. You do not need to use all of them in every session. Use the ones that fit the moment. The second is two to three prompts. These are validation oriented prompts that you write after the session and run against the transcript. Use the patterns from the course, confirm and contradict, retrieve evidence, counter examples, or compare early versus late. The third is an evidence table with five to ten rows. Each row has a claim, the evidence, the source, any relevant notes, and a confidence level with a reason. The fourth is a calibration matrix, the two by two from the course that compares what you noticed to what AI picked up. That is the full cycle notes, prompts, evidence, calibration. Let me walk you through what each one looks like when it's good enough because I want to set realistic expectations. This is a practice workflow, not a polished deliverable. Good enough means someone else could follow your reasoning and check your evidence. For the notes, good enough means you captured the key moments with the right tags. You do not need to tag every single line. Focus on the moments that mattered, the energy shifts, the contradictions, the contexts that shaped the conversation. For the prompts, good enough means each prompt has a clear job. One prompt might pressure test the hypothesis, another might retrieve supporting quotes. A third might look for counterexamples. They should be specific, not vague. For the evidence table, good enough means every claim has at least one piece of real evidence. The source is traceable and the confidence includes a reason. If a row has weak evidence, that is fine. Just label it honestly. Calibration matrix, good enough means you filled in at least two bullets per quadrant, and you ask yourself the edge or bias question for the bottom left. If you want to use a transcript for this exercise and you do not have one of your own, you can use the one I'm providing. That session has enough material to practice the full cycle. Here's the suggested flow. First, watch or read the transcript once without taking notes. Just get the shape of the conversation. Then go through it a second time and take your notes using the notation toolkit. After that, write your prompts and run them, then build your evidence table. And finally, compare your notes to the AI output and build the calibration matrix. The whole thing should take about 30 to 45 minutes once you're comfortable with the tools, and it will get faster each time you do it. Now, let's talk about the templates. You're going to leave this course with a small template pack that you can reuse in your own work. Are not complicated. They are intentionally simple because the goal is for you to actually use them. The first template is the notation sheet sheet. This is a single page with all the tags you have learned the emotional arc symbols, Q arrow and Q later, the context bracket, Oh, and I and says and does. Keep this next to you during the sessions until the tags becomes second nature. The second template is the prompt library. This has the four prompt buttons plus the all purpose prompt. You can add your own prompts to this over time as you discover what works for your research style. The third is the evidence table template. Five columns, claim evidence, source, notes, confidence, ready to copy and fill in. And the fourth is the calibration matrix template, the two by two grid with the four quadrants labeled. These templates are available in the course resources. Download them, copy them into your workspace, and adapt them as you go. Good luck with the project. Take your time and remember, good enough means someone else could follow your reasoning and check your evidence. That is the standard we're aiming for. 17. Lesson 9 - Next Steps: Made it. And before we close, I want to take a few minutes to look back at what you have built, look forward at what comes next, and leave you with something that I hope sticks with you beyond this course. Here's what you now have a notation system that captures what transcripts strip away. Emotional arcs, context, contradictions, the difference between what you saw and what you think it means. A prompt library that turns AI into a thinking partner rather than a shortcut, an evidence workflow that connects every claim to real data with honest confidence levels and a calibration habit that helps you see your own blind spots and get sharper over time. That is a full tool kit, and the good news is, you do not need to use all of it every time. If there's one thing I want you to take away, it is this. Pick one method this week and use it in your next session. Just one. Maybe it is the emotional arc. Maybe it is the says and does pair. Maybe it is writing one good prompt instead of asking AI, what are the insights? Start there. Build the muscle before you build the system. The system will come naturally once the habits are in place. Now let me give you a concrete suggestion for what to do next week. In your next research session or your next conversation that involves learning from another person, do three things. First, use one notation from the toolkit during the session, whichever one feels most natural. Second, after the session, write one prompt and run it against your notes or the transcript. And third, spend 5 minutes on a quick calibration matrix. Even a rough one with one bullet per quadrant is enough. That is the minimum viable practice. One notation, one prompt, one calibration check. If you do that once a week, you will feel the difference within a month. And if you want to track your improvement over time, the calibration matrix is your tool. Keep a simple log. After a few weeks, you will start to see your own patterns, where you're consistently strong, where you tend to miss things, where AI adds the most value. That is self knowledge, and self knowledge is what separates a good researcher from a great one. I also want to leave you with a self assessment question. It is the most important question in this entire course. Did you keep human judgment primary? Because everything we have done here, every notation, every prompt pattern, every template is designed to keep you at the center of the analysis. AI is a tool, a powerful one, but the value you bring as a researcher is your ability to hear what is not being said, to notice the tension between words and behavior, to understand the context that shaped the conversation. Yourself these four questions after any research cycle. Did I form my observations before running prompts? Did I verify AI quotes against the original data? Did I challenge my own interpretations? Did I use the calibration matrix to check my blind spots? If you can answer yes to most of these, you're doing this well. And if you cannot, that is not a failure. It is a signal to slow down and re engage with your own judgment next time. Let me close with something about who you are becoming. Research is changing. AI is changing it, and there are two ways to respond to that. One is to let AI do the thinking and become a person who manages outputs. The other is to sharpen the skills that AI cannot replicate and become a person whose judgment makes AI more useful. You have chosen the second path, and that path has a name. You are becoming the steward of impact. Steward of impact is someone who takes responsibility for the quality of the insights that reach decision makers. Someone who does not just collect data, but shapes how it is understood. Someone who knows that a confident sounding summary is not the same as a trustworthy finding. That is the work, and it matters more now than it ever has because the easier it becomes to generate plausible sounding analysis, the more valuable it is to have someone in the room who can tell the difference between plausible True. You are that person. Your perception steers, AI expands. That is the partnership. You are the one who hears the subtext, reads the context, and holds the contradictions. AI helps you move faster and check your blind spots. Together, you produce better research than either could alarm. Thank you for being here. I hope that these tools serve you well in your research practice, and I genuinely hope you enjoy using them. If you enjoyed the course, please take a moment to leave some feedback and give it a five star review. It really, really helps me out. And if you know someone who would benefit from this, then please tell them about this. Let's go and build something good.