How to create your own AI performance coach: Optimizing your nutrition, recovery & injury management

1. 0:00 – 4:55

   Introduction to Lucas’s athletic background and injury history

   1. LWLucas Werthein0:00

      I have always been super active into sports, really constantly pushing myself to the limits of what my body can do, and naturally, that means injuries, right? And for me, it became a little bit too much, and so as soon as ChatGPT launched, I started experimenting with aggregating this data so that I can get a more clear synthesis of what I can do to actually optimize the body.

   2. CVClaire Vo0:24

      One of the things that I have never seen anybody do yet, I've seen a lot of folks drop in their daily workouts or their food diaries, but I have not seen MRIs and imaging here. And what important context for somebody who's an athlete to say, "Not only is this how I'm performing on an output basis, [chuckles] but this is actually, like, the structural setup under the hood," so it's really interesting, that combination of data into these files.

   3. LWLucas Werthein0:48

      I'm wanting to demand of my body to feel, like, twenty-five in a forty-year-old's body, and it's interesting to think, what if every person could have a coach that organizes all this action into clarity, right? And part of what we've been talking about is that not everyone is looking for this type of performance. Most people don't need six-packs or match prep, but they could use help with the basics, right? Eating less processed food, sleeping better, moving more, and I think an AI coach could meet people where they are and actually give them the necessary nudges and contextualization of information that they need to be a better version of themselves.

   4. CVClaire Vo1:25

      [upbeat music] Welcome back to How I AI. I'm Claire Vo, product leader and AI obsessive, here on a mission to help you build better with these new tools. Today, we have Lucas Werthein, Head of Technology at Cactus, who has done work for basically everybody: Apple, Coca-Cola, MTV, and even Beyoncé herself. But today, we're not going to talk about product development. We're going to talk about how Lucas has built a wellness coach inside ChatGPT to optimize his nutrition, his workouts, and keep him feeling twenty-five, even though he's a little bit older than that. This is a really fun episode with some practical insights for people just trying to make their lives better with AI. Let's get to it. This episode is brought to you by WorkOS. AI has already changed how we work. Tools are helping teams write better code, analyze customer data, and even handle support tickets automatically. But there's a catch: these tools only work well when they have deep access to company systems. Your copilot needs to see your entire code base. Your chat bot needs to search across internal docs, and for enterprise buyers, that raises serious security concerns. That's why these apps face intense IT scrutiny from day one. To pass, they need secure authentication, access controls, audit logs, the whole suite of enterprise features. Building all that from scratch, it's a massive lift. That's where WorkOS comes in. WorkOS gives you drop-in APIs for enterprise features, so your app can become enterprise-ready and scale upmarket faster. Think of it like Stripe for enterprise features. OpenAI, Perplexity, and Cursor are already using WorkOS to move faster and meet enterprise demands. Join them and hundreds of other industry leaders at workos.com. Start building today. Lucas, welcome to How I AI. Thanks for being here.

   5. LWLucas Werthein3:21

      Thank you. Uh, glad to be here, and thank you so much for inviting me to be on this wonderful podcast and show.

   6. CVClaire Vo3:28

      What I'm excited about is so much of How I AI so far has really been how I AI for business, and I really want to show your use case because it really is a personal how I AI, and how you can actually use AI in your daily life to really make improvements and build something for yourself. And so tell us about the story that got us to what you're going to show us today.

   7. LWLucas Werthein3:56

      I have always been super active into sports and, and would consider myself a pretty competitive person, and so that means really constantly pushing myself, uh, to the limits of what my, my body can do and, and how much I can deal with in terms of the, the stress of, of these sports. Um, and naturally, that means injuries, right? And so I've had surgeries on my foot, two surgeries on my knees, um, surgery on my shoulder, and this is through, through various sports, from surfing, uh, to Muay Thai, uh, to playing tennis, weightlifting, and kind of changing it up over the years as the injuries come and needing to move into new sports. And we were, you know, joking about this before the show, but obviously, as we enter forty,

2. 4:55 – 6:11

   The challenge of synthesizing siloed health data

   1. LWLucas Werthein4:55

      um, things start becoming a little bit more dire, and you start paying attention, more attention to how your body feels and reacts daily to the things that you didn't feel before. And I started becoming really obsessed with how I could optimize my body. You know, I'm forty years old. I'm running a company. I play competitive tennis, I lift weights, and I'm recovering from all these old injuries, and I'm trying to keep up with these teenagers on the tennis court, playing these amateur tournaments and running around, and, and I'm, I'm wanting to demand of my body to feel, to feel like forty, um, to feel like twenty-five, sorry, uh, in a forty-year-old's body. And, you know, data is so siloed, um, and to make sense of everything that people tell you, that professionals tell you, and put it together is actually really hard, right? You get blood tests, you go to the nutritionist, you go to the physical therapist, um, you get data from your Whoop.... y- you know, nutrition plans, InBody scans, and for me, it became a little bit too much.

3. 6:11 – 9:57

   Building a GPT to optimize performance and recovery

   1. LWLucas Werthein6:11

      And so as soon as the GP, uh, ChatGPT launched, um, I started experimenting with aggregating this data so that I can get more, a more clear synthesis of what I can do to actually optimize the body, right? Because the problem isn't the lack of data, the problem is the lack of synthesis and putting it all together. And I started having a few breakthroughs, actually, and it started helping me feel better and perform better, and I just started using it on a, on a daily basis. So, you know, it came from a need of getting injured and trying to perform and getting back on the horse, to now actually having interesting technologies that are a- allowing me to, to, to have really, really specific actions that I can take to actually perform better.

   2. CVClaire Vo7:07

      And what I wanna reflect on before we get into actually what you built, which is gonna be really interesting to see, is, you know, you strike me as a person, and you've described yourself as a person, that is pretty proactive about seeking out data, seeking out advice, going to medical professionals, getting different advice, reading. And so it's not that you're not informed, it's not that you don't have access to experts, but for all this data and all this effort and all this access, you were still struggling a little bit with, with some things here and there. And it's pretty amazing to me that what you're telling me is: "You know, even given that whole portfolio of things that I've put against my, my body and my wellness goals, this AI tool that I built was actually one of the things that helped me unlock a couple of things that had been bothering me for, for a really long time." So I'm really interested that, you know, that last mile of optimization is really being driven by, by this tool that you built, and it's, it's pretty cool to see somebody who, um, is a deep expert still get a lot of value out of, out of going even further.

   3. LWLucas Werthein8:15

      It's, it's, it's a really interesting conversation because I think we all see it in our numerous interactions in the field of health and wellness. You know, you were describing that you, you do a lot of PT. I'm dealing with an elbow injury right now, and I was having a conversation with my PT today, and they were telling me about a patient that had to have surgery for their elbow after a while, and I said, "You know, it's really interesting because the, the doctor makes the diagnosis, you guys are treating the patient, but this person needed to go into surgery because it-- there, there's a missing link." There wasn't someone looking at this guy's stroke and saying, "Well, you need to change your tennis grip like this," or, "You need to change..." Essentially, a biomechanics specialist, right? And, and, and to me, that's really interesting because there, there is always a lack of communication, and the information is a little bit siloed. And I think that what I'm about to show, when you start thinking about the possibilities of, of how this can scale, even to not through performance, you know, we're talking about the edge of the edge of the edge of trying to gain a little bit to, to, to be better, um, but we're gonna... I can talk more about that, but I, I think it's really interesting how, you know, essentially, this is a performance strategist, right? It's, it's trained to personally think about my joints, optimize my energy, extend my peak, and you'll see that it, it answers me filtered through rules I've created, um, and that helps me to sort of compile all this information that is usually really disparate and really separate, but yeah.

4. 9:57 – 13:54

   Demonstrating the data types integrated into the AI coach

   1. CVClaire Vo9:57

      Okay, so let's actually, let's go ahead and, and show it because I'm really interested to see how you got to something that meets your standards, which from what I can hear, are, are quite high. So you're gonna pull up your screen and show us this WellCoach GPT that you built.

   2. LWLucas Werthein10:13

      Yeah. Well, let's, let's get into it. Let me, um, let me dive into this. So when I configured this, this GPT, I fed it a few files that were important, um, for it to have the context of what I wanted, um, information to be reflected on, right? And so here you'll see that it has an X-ray of my left knee, my left knee, my right knee. It has my physiological cycles. This is a CSV file coming in from Whoop data, um, my journal entry, so what I described in my day-to-day, am I stressed? Do I have anxiety? Did I sauna? Did I do compression therapy? My workouts that are logged on a daily basis, and essentially my strain and how hard I worked, and my sleep data on how much good sleep did I get, REM sleep, deep sleep, so a lot of, lot of data being fed into, into here. In addition to that, I mentioned I had a couple of knee surgeries, so it has the MRI of my knee, pre-surgery, post-surgery, and it has a few blood exams from, uh, this year and last year, so three different blood exams. So it can compare the evolution of the tests and how I'm doing. In addition to a nutritional plan from a nutritionist/dietician that helps me think about food as fuel and how I can perform better based on, on fuel, and an InBody scan that essentially measures, um, percentages of fat and muscle and distribution, um, across the body. And so it's using all these files, um, to think about, to have context around myself, and so that, that was an important element to be able to gather this data manually inputted into this GPT.

   3. CVClaire Vo12:09

      ... What I think is interesting about this for folks that are listening or watching are a couple things. One is all this data is in all these different formats, right? So you have imaging data from MRIs and X-rays. You have, like, s- semi-structured data from sleep, from a wearable. Um, you have blood tests in PDF form, where it's gotta parse a bunch of stuff, a textual nutrition plan. And what I love about AI now, for people that maybe haven't built some of these tools for themselves, is you can just dump all that data in, and you don't have to worry about, is it clean? Is it organized? Is it structured? Just put it in. And then one of the things that I have never seen anybody do yet, I've seen a lot of folks drop in, like, their daily workouts or their food diaries, but I have not seen MRIs and imaging here. And what important context for somebody who's an athlete to say, "Not only is this how I'm performing on an output basis, [chuckles] but this is actually, like, the structural setup under, under the hood." So it's really interesting, that combination of, of data into these files. And then how is this GPT set up to actually work? What are the instructions? Would you walk us through that?

   4. LWLucas Werthein13:20

      Yeah, just to add something, I think you mentioned something really interesting around how the data is structured, and it's also coming in from different languages, right? Um, because I spend a lot of time in Brazil, some of my exams are in Portuguese. Um, a lot of them are in English, but some of them are in Portuguese, and so I don't need to worry about that either. Um, I just dump it in, and it processed that information. Um, and so I, I think that's also a, a valid point about how easy it is now to, to port that data into something that can unify it. Yeah, you're absolutely right. Let me tell you

5. 13:54 – 16:31

   Configuring the GPT with clear performance goals and boundaries

   1. LWLucas Werthein13:54

      about how I configure this, right? And so I'm telling it to act as my performance strategist and health optimization coach. It has access to my physiology, labs, imaging, wearables, and I want it to coach me like I'm a high-performance operator. That's really important. I'm not trying to be a professional athlete. I want it to un- understand that I want to perform, but I'm balancing, I'm balancing tennis, lifting, recovery, and mainly running a company, which takes the majority of my time. And I, I don't want to be the most competitive person in the world. I don't want to be the best. My main objective is for it to safeguard my, my joints and to amplify my output and to extend my peak. I wanna feel healthy and pain-free. That's super important. But I do want to perform like a 25-year-old in a, in the body of a 40-year-old, so I give it that, that, that instruction. Um, when, when I share my prompt, um, I wanted to interrogate it through my contacts, right? Looking at my, my blood exams, my scans, my Whoop, um, other, um, specific information that it has. I want it to flag what we call red and yellow zones, right? Um, we see this in a lot of wearables or early signs of overtraining, under-fueling, inflammation, and it's important. I want high ROI actions. Um, no fluff, no hacks, nothing that hasn't been proven, and I wanna be kept inside this zone where I'm moving pain-free. I can play at high-output tennis. I don't break down. Um-

   2. CVClaire Vo15:38

      I'm, I wanna reflect to you something that, one, I think is personally interesting, and two, I think is er- interesting from a prompts perspective. So, you know, the top of your prompt is very common. "You act like a blank. You are a blank," that sort of instructive point to the LLM to give it a role. What I think is really interesting about this last bullet point here is it's the opposite side of that coin, which is, "At the end of this, I want to be X, Y, and Z." And so, you know, say, you're a performance coach. That's the kind of your role. My role is I'm running a company. I wanna feel 25 in a 40-year-old body, and I wanna be rested, move pain-free, and play tennis. Like, it's a very clear input-output structure, and then the human in me wants to reflect, "These are very reasonable, nice goals." So again,

6. 16:31 – 17:50

   Setting realistic expectations for the AI coach

   1. CVClaire Vo16:31

      you know, we're talking about this, like, hyper-optimization, and at the end of the day, you wanna wake up, you wanna feel good, you wanna engage in your company, and you wanna be able to play the sport that you wanna play. And so I think the kind of idea of, like, a, a, a role and then a really realistic outcome for yourself is a nice framing for something like this kind of personal coach prompt.

   2. LWLucas Werthein16:51

      Absolutely, and, and super important for it not to tell me to go get ozone therapy-

   3. CVClaire Vo16:58

      [chuckles]

   4. LWLucas Werthein16:58

      ... um, or to go sit inside a hyperbaric chamber.

   5. CVClaire Vo17:02

      Yeah.

   6. LWLucas Werthein17:02

      Um, which, which may work. I'm not necessarily giving it a ding, um, but I, I do want things that, that are accessible in my day-to-day-

   7. CVClaire Vo17:12

      Yeah

   8. LWLucas Werthein17:12

      ... um, and that are proven and that are scientifically backed. That's really important to how I want it to think about, um, the recommendations.

   9. CVClaire Vo17:22

      Yeah, and the other thing I see a lot when people prompt is they go to these extremes, where they're like, "You are the best in the whole world, and you're going to make me the most elite," blah, blah, blah.

   10. LWLucas Werthein17:33

       Right.

   11. CVClaire Vo17:33

       And you know what I like about this is you're saying, "I'm getting good outcomes by, like, pulling in the bounds of reason on both sides [chuckles] of those-

   12. LWLucas Werthein17:40

       Yeah

   13. CVClaire Vo17:40

       ... and having reasonable roles and reasonable expe- e- expectations." So it's a really good insight from a prompting perspective. Uh, let's go down, and show me-

   14. LWLucas Werthein17:48

       Yeah

   15. CVClaire Vo17:48

       ... show me what you optimize for.

7. 17:50 – 21:47

   Creating nutrition, training, and recovery frameworks

   1. LWLucas Werthein17:50

      Yeah, so, so I think part of this is the context that, that I mentioned of how, for example, it has my nutrition plan, right? And so when you think about performance, so much performance is about how you eat and how you rest, and so having the basis of how I want it to eat has been absolutely fundamental for it to think about the recommendations, right? And so I want it to stick-... to my nutrition plan, unless there's data, a driven reason to adapt. Um, so this is all based on fueling and avoiding inflammation, right? I wanted to prioritize energy, um, stable glucose, uh, low inflammation, and muscle retention. Number two, thinking about training and load management, you, you can't overtrain. If you overtrain, you burn out, and so I needed to think about balancing strength, endurance, and mobility, um, because I need to protect my knees and my shoulders, and my joints, which have been messed with in, in surgery. And so when thinking about recommendations, um, we can't overload, um, the HRV, we can't be outside of sort of the readiness score, um, and I need it to help me pull back because I will overtrain. I, I, I do wanna get better and better and better, right? And so one of the things we hear about the most when studying and thinking about, um, performance is people don't pay enough attention to recovery and to rest, and so this is super important for me. The third one, again, going back to recovery and, and regeneration, is, uh, sleep is the main factor here, right? And yes, PT, mobility, sauna, cold, massages, mindfulness need to be important and not optional. They're part of the training cycle because they're part of recovery, so I need recommendations of how to have it give me nudges so I can maintain those up on my day-to-day. And lastly, these, this idea of, of tracking and, and feedback loops. It's integrating data across Whoop, InBody LABS, diet, journal entries, and I need it to cross-validate the decisions and not recommend something that is not aligned, um, with what I have fed it, just, like, from pulling some random thing from the internet.

   2. CVClaire Vo20:15

      One of the things I, I wanna reflect on here that I've said in other podcasts, more in sort of a business context, is when people are designing these GPTs, I really read these prompts, and I'm like, "Oh, they reflect, like, the perfect employee, or they reflect the perfect team." And when I'm looking at this, this sort of reflects how, in an ideal world, all these experts that are supporting you, your doctor, your PT, your coach, would all be fully integrated, aligned on a strategy, like, consistent in their recommendations, data-backed. But the reality is, when you bring a team of individual experts together, one, they're all gonna come with their unique point of view. Two, it's very hard just, just tactically to stay aligned on recommendations and kinda resolve things across the board. And so what I like about this is, you know, y- ideally, you'd be able to sit all those experts in a room with you and say, "Hey, hey, guys, [chuckles] this is how I want you to take care of me." But because that's not actually practical, what you're doing is bringing some of the data and the insights those people have, your own ambitions and goals, and then sort of like putting it in this system that will operate optimally for you, uh, over time.

   3. LWLucas Werthein21:28

      Absolutely, and, and we'll talk more about the vision and, and, and a little bit of provocation of w- where I think this will, will go-

   4. CVClaire Vo21:36

      Yeah

   5. LWLucas Werthein21:37

      ... and how this is a, a, a prototype of something that, um, will be much bigger and that many, many, um, practitioners, health systems, physicians will

8. 21:47 – 24:25

   Establishing hard boundaries and anti-prompts

   1. LWLucas Werthein21:47

      adopt in the future.

   2. CVClaire Vo21:48

      Yeah. Okay, and then you do what everybody does, which is you give it a bunch of stuff to do and then a bunch of stuff to not do.

   3. LWLucas Werthein21:56

      Exactly. So hard, hard boundaries, right? No pushing past the volume and intensity when metrics show under Recovery. My Whoop is showing a red or a yellow, I can't go train hard. Don't give me supplements that are unknown. Don't, you know, don't tell me to go take, uh, creatine or anything that, although it's super popular at the moment-

   4. CVClaire Vo22:18

      Yeah

   5. LWLucas Werthein22:18

      ... right, that, that we, we don't know is absolutely measurable, scientific fact-

   6. CVClaire Vo22:22

      Yeah

   7. LWLucas Werthein22:22

      ... and has ROI. Don't give me novelties. Um, stick to, to what actually works, to perhaps even ancient data, and act on red flags. Uh, if, if I tell you there's a lot of soreness or low HRV or decreased sleep quality, that means perhaps I'm getting sick. Don't, don't push when I can't push.

   8. CVClaire Vo22:45

      Mm-hmm. Yep. Great, and I think this, again, for people, I'm just kinda giving the meta com- commentary, which is it's a very common prompt structure for anybody trying to build something for themselves is, like, give you a role, give the GPT a role, give it a goal, um, give it some input and data, give it, uh, an anti-prompt, I say, which is, like, tell it what not to do. And then I like that you're closing on, like, the, the check that it's all following the rules and this is how I want you to respond piece. So we can go through that really quickly, and then maybe show a couple examples.

   9. LWLucas Werthein23:16

      Totally. So values, right, precision, energy, adaptation, kinetics. It's all about movement. It's all about energy. It's all about precision, and then the tone, right? Like a coach, be clear, tactical, no fluff, no lectures. Connect the dots that's super informed about everything that we're talking about, and prioritize what matters, um, this week, not vague long-term theory around what's possible.

   10. CVClaire Vo23:43

       Great, and so you know, it's very clear you put a lot of thought into this. Did you also use ChatGPT to help you, like, craft the structure of this prompt?

   11. LWLucas Werthein23:51

       Many times.

   12. CVClaire Vo23:52

       Yeah, I, I, I, I can tell from the emojis. [chuckles] Uh-

   13. LWLucas Werthein23:54

       From the emojis, yeah.

   14. CVClaire Vo23:55

       You can always tell.

   15. SPSpeaker23:56

       This podcast is supported by Google. Hey, everyone, Shresta here from Google DeepMind. The Gemini 2.5 family of models is now generally available. 2.5 Pro, our most advanced model, is great for reasoning over complex tasks. 2.5 Flash finds the sweet spot between performance and price, and 2.5 Flash Lite is ideal for low-latency, high-volume tasks. Start building in Google AI Studio at ai.dev.

9. 24:25 – 27:30

   Example: Managing nutrition around special events

   1. CVClaire Vo24:26

      ... Okay, awesome.

   2. LWLucas Werthein24:27

      Yeah.

   3. CVClaire Vo24:27

      So this is a great deep dive into instructions, and I hope people are paying attention to it, because, one, like, what a great prompt. Thank you for sharing. It's super useful, and two, just the structure of it is very classically, uh, uh, well set up for setting up a GPT. So whether it's this topic or another one, I think people can learn a lot from how you've set it up. So, how does it work? Show, show me what, what are common things that you, that you would do with this GPT?

   4. LWLucas Werthein24:55

      Yeah, let me give you, let me give you an example. I'll give you a few fun ones here. Um, so this morning, I, I woke up, and my wife told me that we have a, a birthday dinner that we need to attend. Um, good friends of ours are gonna celebrate an omakase, uh, birthday dinner, which means plenty of rice and sake. And so how should I manage my day to balance the fact that I'm gonna indulge in the evening? And this may be simple, um, but actually, it's really interesting that it, it thinks about how to change my actual diet, right? In the morning, I would usually eat something post-training, but here it's saying, basically, like, "Eat only protein, minimal carbs." Um, at lunch, it's saying the same thing, and so it's guiding me how to go along my day, um, based on the fact that I'm going to indulge in the evening, um, and have something that's gonna be a little bit different and not necessarily feel destroyed. And so it's prepping me for something that's going to happen, and it's really useful because I don't have to think about it. I prompt it, I asked it, it gives me a response, and I try to adapt, right? And so, okay, great, it told me to have, um, no carbs. I took a picture of my breakfast, a little bit of eggs, avocado, coffee, um, and then it, it feeds me information about, "Okay, that's fantastic. Add a little bit of pepper for inflammation." You know, I'm, I'm very cognizant also that, uh, this is, this is... Most people, um, this is not what most people need, right? If we think about most people, they perhaps just need to move a little bit more, sleep a little bit better, not eat processed food. Um, and I, I, I'm very cognizant also that this is for something very specific that I'm personally looking for, but it's very useful to how I can then program my day and how I can think about the next day as well.

   5. CVClaire Vo27:02

      Well, I think the other thing is, yes, your goals are maybe, um, a higher level of what the kinda like baseline person might have for their own performance health goals. And at the end of the day, the like, "I'm going to a birthday party later, and I don't wanna feel crummy tomorrow. Are there any things that I can do before this birthday party to keep me from feeling crummy?" is, like, a very applicable problem, I think, one. And the second thing

10. 27:30 – 28:24

    Accessibility and affordability of on-demand coaching

    1. CVClaire Vo27:30

       is, you know, it's really, on-demand coaches and nutritionists and experts are expensive. They are inaccessible to a lot of people, and just this sort of short loop, like, "Am I doing the right thing? You know, give me an answer." And I like this piece that you showed us, which is like, "Look, I did it," and you get, like, a very short blip of a good positive feedback loop, can actually help people reinforce habits that I think compound over time. And so, you know, I do think something like this makes some of these, like, tactics, you know, that, that sound very basic, a little bit more accessible, a little bit easier to implement, and gives you sort of an on-demand feedback loop that, as social human beings, we res- [chuckles] we respond to. And so I don't think it's kind of too, too far aground from what most people would find useful.

11. 28:24 – 29:31

    Practical examples and real-life scenarios

    1. LWLucas Werthein28:24

       I think that's a, that's a great point, and the fact that you now have a coach in your pocket-

    2. CVClaire Vo28:31

       Yep

    3. LWLucas Werthein28:31

       ... is super interesting because things change. Um, another scenario, I said it the other day, is I said I was going out to dinner. I prepped the day, but there was a change of plans, and we went to a party, and we were gonna drink and have a bunch of gin and tonics and get home at 3:00 a.m. And I didn't need to think so much about what I needed to do because I prompted it, gave it that information, and then it reacted based on that change of plan. And so having that be accessible now to so many people, whether you are able to make that change at home because you have access to food, but even if, you know, you need to go eat at Chipotle, and it can tell you the things that you can eat or that you should order at Chipotle because that is your only option, I think is a super interesting point of just how accessible, um, good information for you to be optimal, um, is becoming.

    4. CVClaire Vo29:28

       So I would love to see... I think this is a really great sort of,

12. 29:31 – 34:19

    Using AI for injury management and recovery planning

    1. CVClaire Vo29:31

       like, day-to-day practical example. I'm curious if you have anything that shows a little bit more of the kinda like physiology side of things. You know, you mentioned a lot at the beginning injuries and protecting joints and making progress. Has, have, has any of that come into play in, in your coaching? You know, we see the, the nutrition side, but I'm curious if there's anything else there.

    2. LWLucas Werthein29:53

       Totally. Um, let me show you.

    3. CVClaire Vo29:56

       What I also wanna call out for folks that I love as you're scrolling is context changing, content changing. It's like, "Now I wanna talk about nutrition, now I wanna take a screenshot of my workouts, now I wanna do, do this and do that," and, and the chat can kind of adapt to all that information and not need you to follow any rules or any schedule or any structure, so I think that's really interesting. And then I love that this is in Portuguese, some of it, and then switches to English. Uh, I, I caught that [chuckles] on the screen share.

    4. LWLucas Werthein30:30

       ... So I think this one's really interesting. Um, I have been dealing with not a tennis elbow, but an elbow injury, and, um, I went to a doctor. I, I gave it, um, the diagnosis of the doctor, um, I gave it the prescription of physical therapy, and basically, I talk about my pain. I would talk about my pain on a daily basis, and I would take pictures, and I would record movies of how I'm feeling pain, and basically, it would confirm what the doctor, um, has said. You know, I tell it the doctor discarded tennis elbow, um, and it's like, "You know, I've been, I've been off the courts for a week. How much longer?" And the doctor has told me, and the PT has told me, but I'm trying to test it if, if it's going to say something different, and it's saying the same thing. And really frustrated, you know, I'm following all the prescriptions, I'm, I'm doing all the exercises, but it's like it's not, not getting better. And then one day, I actually go into PT, and it does feel better because of something specific that they did with electrical currents and strength training at the same time. And so now I'm prompting it to think about if, you know, based on the evolution of the recovery that I've been telling it, will I be able to play this amateur tournament that I, I wanna play on September 18th? And so it's thinking about how many weeks are left, what's realistic, you know, decision checkpoints, um, what it thinks, and then I ask it to put, put everything on the table to think about the recovery, um, the recovery, uh, plan, um, for, for how we're gonna do this. And honestly, it is exactly the same thing the PT told me, which is really interesting, but it is contextualized in a way that I can digest. And now, sort of the anxiety of me every day thinking, "Man, is the pain gone? Is the pain gone? Is the pain gone?" is eased a little bit because I can manage the expectations of what will happen in just a very visual manner that you don't usually get, um, from your PT or from your doctor. And so I'm contributing it this information so it can think like them, but perhaps process and sy- synthesize the data a little bit better because it has so kn- so much knowledge about myself and what I'm doing and how I act.

    5. CVClaire Vo32:55

       Yeah, one of the... You know, I wanna just go back and reflect on what you just showed us, because I think there's a couple really interesting things here for people to listen to. One is, I think people really underuse what you just showed, which is a video or a picture circling a thing into, [chuckles] into ChatGPT. I found that that's such a useful, um, a useful kind of workflow for folks that are new to AI and not sure what it can do for it. Uh, I don't know if this is an appropriate metaphor or not, but I live in a hundred-and-fourteen-year-old house. It's, like, very similar to living in my forty-year-old body, and, uh, we, [chuckles] you know, we have leaks here or cracks there or bubbles here or whatever, and I'm constantly taking a picture of something, circling it, and saying, "What could this-- You know, tell me what this could be." And so you can do that. You know, I have a-- It's not tennis elbow, it's I sit at my laptop elbow, um, and put my, my arm on my desk at a bad angle. I know I do it. Um, but taking that and just saying, like, "I've got pain here, not, not here, not here, but, like, here. Um, what could that possibly be?" It's a really good use case. I think people also don't know, uh, a lot of these models can process video really well, and so that is another input you can put in, and, you know, kind of do the thing that they make you do at PT, which is, like, "I can go to here, but not further," or, "I can do

13. 34:19 – 37:25

    Validating expert opinions and translating medical advice

    1. CVClaire Vo34:19

       this." So I think that's a really interesting workflow. I think the second thing people are using ChatGPT for a lot is just validating expert opinions. Not to dismiss expert opinions, but you know what? My, you know, personal doctor is not on demand twenty-four hours a day. When I leave the office, that's about as much as I'm gonna get for them. So being able to go back and saying, "Can you re-explain this to me? Is there anything else it could be?" uh, just gives you a more accessible outlet for sort of validating some of that stuff. And then the last thing I would say is, often when you leave a, certainly in the health profession, but an expert, and they give you some takeaways, right? They give it to you in the format they give it to you. They explain it to you verbally, they text it to you, they give you a little takeaway sheet, and you're like: "No, I want this, but I need it in a day-by-day plan until September," or, "Can you re-explain this to me in this format?" And I also think this ability to grok the same information through a different format by having an LLM translate it, it's really useful, especially when it's information from an expert where you may not understand the terms or the language or the mechanics. And so I think those three things are really interesting use cases of, of AI, and you can see them all in just this one flow.

    2. LWLucas Werthein35:40

       I think that's a really fantastic point, and I think we can extrapolate and think about what, what I'm doing here for myself, you know, manually uploading MRIs, Whoop data labs, just exposes a much bigger opportunity. Um, and, and AI could be a, a missing synthesizer of personal health, and I think that healthcare has obviously an interoper- interoperability problem, and the data is siloed. And it's interesting to think, what, what if every person could have a coach that, that organizes all this action into, into clarity, right? And part of what, what we've been talking about is that not everyone is looking for this type of performance. Most people don't need six-packs or match prep, but they could use help with the basics, right? Eating less processed food, sleeping better, moving more.... and I think an AI coach could meet people where they are and actually give them the necessary nudges and con- contextualization of information that they need to be a, a better version of them- of, of themselves.

    3. CVClaire Vo36:50

       Well, what's really funny about this is I'm thinking about you as, you know, a more high-performance athlete, operator. I was just reflecting, I wanna make this for... My eight-year-old wants to get much better at basketball. Like, that's his performance goal, and I'm like, "Oh, you could take the same framework," right? He's eight, he's got this much time. You know, we have to walk to the basketball court. What, what do we do? And you can do everything from videos to, to, um, you know, pictures, all that kind of stuff. And so I think it's just really interesting to think about this, no matter what your goals are, setting up a framework like this that can help

14. 37:25 – 43:27

    Vision for the future of AI in personal health coaching

    1. CVClaire Vo37:25

       you day by day increment your way to them. So before we, before we wrap this up, you know, you've, you've kind of, um, talked about it a little bit, but, you know, what do you think the future of this is? Are everybody gonna make this themselves? Do you think there's a product here? Like, what, what do you think is the gap between I have a GPT, and everybody can kinda do this on their own?

    2. LWLucas Werthein37:52

       Yeah, I, I think there's a really interesting notion when you think about this, and you think about how this can potentially scale in the near future. I, I see a vision in which in five years or less, everyone will have access to a personal AI health coach. And not to replace doctors, but to help us show to doctors, um, show up more informed, and to live healthier between visits, and to make these micro decisions every day. Additionally, I do think that the doctors will also have this, and so our AI will talk to the doctor's AI, and it's interesting to think about what are the spaces that need to be designed, and what type of interactions will occur once that happens. It's gonna be a different world because they will have all the context. Um, they will meet having all the context, and when the doctor and the patient show up, there's just much more clarity, um, to have the conversation. And so I think that the, the, the future of health isn't just about medical breakthroughs, it's a lot about synthesis and the ability to turn this overwhelming amount of data into something that's simple and very, very personalized. I also believe in an era of seamless capture. You know, we're talking about manually uploading all these things to the GPT, but it will be seamless. We will have microsensors around, potentially in your bloodstream, tracking information, glucose, hormones, smart fabrics, eventually toilet sensors measuring microbiome, hydration, and it will all be ambient, passive, and invisible. And I think that there's a world where the healthcare leaders will eventually sell their knowledge as trained AI models. You know, imagine having a coach in your pocket that's been trained not just on you, but on decades of patient data from institutions like Mayo Clinic or Admin Health.

    3. CVClaire Vo39:59

       Yeah.

    4. LWLucas Werthein40:00

       Um, it, it, it, it's, it's just really interesting to think about what happens when you combine that passive data and you put it next to the guidance that's grounded in the best medical science and personalized to you. And I think this also gives the ability for the doctor to get out from in front of the computer and be a storyteller-

    5. CVClaire Vo40:23

       Yeah

    6. LWLucas Werthein40:23

       ... and a long-term strategist, and to have this hyper-personalized aspect around food, supplements, habits. Um, and so I, I think that I laugh because I think that in 10 years, we'll look back and chuckle, um, about how much manual effort we've put into health tracking, and we'll think about just like no one today types in GPS coordinates into their phone, no one will manually log workouts or meals, and health data will capture itself, and we will have coaches in our pockets to go, um, back and forth and, and evolve. And so just last words, I do think that it's important to reflect that this is absolutely not about gimmicky. Um, it, it, it's, it's really a precision tool for consistency. I am looking for high ROI choices, and it's helping me do that through injuries or food. And I do think that when we think about a potential population-level impact, that's where this becomes powerful to imagine, um, what this can be. Um, and it's actually something that my company does quite a bit and, you know, maybe, maybe this is not the place to talk about it, but it, it's something that I'm seeing a, a fast adoption to how health systems, and wellness companies, and doctors are thinking about this. So I think there's, there's a brave new world coming.

    7. CVClaire Vo41:52

       Well, I'm ex- I'm excited for, for that world. I was just think... I s- I spent a hot minute in health tech, and I was like, "Where is my FHIR Epic, uh, uh, MCP [chuckles] I can plug into?" Uh, and then, and then I'm gonna go wild. But again, I, I think that is- there's a future... I wanna bring it back to people who are maybe watching the podcast saying, "How does this apply to me? I'm not an athlete." And, and just the use cases that I think of are caregiving. When you're a caregiver of an elderly relative, you have so much information, so many specialists, so much points of data that you have, that p- you know, you go visit, you have pictures, you have recordings, you have all this stuff, and just having this, this coach to maybe help with the caregiving journey is one. I think kids are super interesting. I think athletics is very interesting. I think there's probably product market fit for people that, um, apparently there are many on this podcast recording right now that are-... 39 turning 40, but wanna be 25. Um, and so I just think, like, you know, maybe people listening don't have the exact same goals, but the framework really applies to a lot of healthcare and wellness challenges, um, or, or goals for people. So I'm really excited that you showed this to me. It gave me so many- I have so many ideas. I won't bore everybody with my, um, achy hip and elbow, but I have... And, and sleep issues, uh, which I can attribute to my kids. But I think it's, it's a really interesting, and it's inspired me to wanna go build a couple of these for different, different coaching topics I have. Okay, well, we are-- this was fa- fabulous. I wanna do kind

15. 43:27 – 48:48

    Other AI workflows: synthetic clients and AI co-founders

    1. CVClaire Vo43:27

       of three lightning round questions. One is just, you mentioned a couple other workflows, so this is your, your favorite, but can you just tell me a couple other maybe that you use in work? Uh, maybe flash us one or two that you think are, are interesting that people can think about.

    2. LWLucas Werthein43:42

       Yeah, absolutely. Um, I'm also happy to share, um, the prompt-

    3. CVClaire Vo43:47

       Yeah, great

    4. LWLucas Werthein43:47

       ... so that you and others, um, can configure their own GPTs.

    5. CVClaire Vo43:50

       Yeah, we'll put that in the show notes.

    6. LWLucas Werthein43:52

       Cactus is a firm that works at the intersection, um, of physical space and digital technology, and we work for healthcare and wellness, uh, developing, uh, essentially digital products and thinking about physical space. It's an intersection of a consulting firm and a design firm, and so many times we're thinking about new products, new services for our clients, and this is an example where we have taken the client, who is a brilliant, fantastic doctor, but she is extremely busy, and so we have synthesized a lot of this information from articles and other podcasts and things that are available on the internet, uh, and we have vetted this information so that we can ask it questions when she's unavailable and try to get the work to 80 or 90% of where we think she would agree with, so that when we present it to her, it's not taking time that could be skipped over, because we have a lot of how she thinks and how she makes decisions.

    7. CVClaire Vo45:02

       I, I love this. We have seen one or two synthetic bosses on, on the podcast, and I love this from a kinda like consulting firm, design firm perspective, which is, like, synthetic clients are very interesting ways, because, as you sa- you know, just going back to everything we've been talking about, your expert's not always available, your client's not always available, people's time is scarce. But if you know enough about how they might react to information, you can not only give yourself some insight, but you can also give, um, your team insight into how they might react to things. And then, you know, I, I, I love these, uh, as people make them for their bosses or for their clients. A little pro tip to the bosses and clients out there, you want, you wanna make this even easier, make one of yourself that [chuckles] you can share with people, um, because you probably have the best understanding of what's important to you and how things matter. And so it's one of those tips I tell everybody to do is go replicate yourself in a GPT to give your team a first-line pass of feedback, and then sometimes you end up like me, whereas you build that GPT, and then it accidentally becomes an enterprise software, [chuckles] software business, which is how my company started. So I think this is a, this is a great idea. And then one other thing we, you, we talked about before the show, and maybe you can just voice over what you do here, is you have an AI co-founder as well, so lots of synthetic people. Tell us a little bit about that.

    8. LWLucas Werthein46:26

       Yeah, my... I love my, my co-founders. They're, they're brilliant. Um, and I do not need, uh, an, an AI co-founder, but this is a, a, a new world that we're living in. Uh, the, the, the company that we run is a distributed firm. We're no longer in an office, and we no longer have access to each other, um, you know, by going and, and tapping on them and saying, "Hey, do you have a minute?" And you have to schedule a call or call them, and many times, you just need a little bit of a partner to think about a potential thorny problem, um, that you would only think with your co-founder. And it's really interesting to load it up with data around how your co-founder thinks, how you think, some of the problems that you're going through, and being a voice to brainstorm with you so that you're not starting from a blank slate, which is something obviously you hear a lot, but it's really helpful. It's almost like business therapy.

    9. CVClaire Vo47:26

       Yeah. Well, [chuckles] on that point, I'm laughing to myself 'cause as you were describing this, I was thinking, "Oh, maybe I can save my husband a little bit of trouble if I make a synthetic Claire," and he just double-checks, like, "Should I? [chuckles] Should I say or do this to Claire before-

    10. LWLucas Werthein47:41

        Yeah

    11. CVClaire Vo47:42

        ... before I do it?" But I, I do think that there's- we're in this interesting world where, you know, wanting the expertise of someone on demand is, is not always possible, and AI has made an approximate version of that possible, and it's not, you know, it's not the real thing. Um, but it does help you in, in the moment, um, especially in a distributed world where, you know, you don't wanna @mention your colleague in Slack at, you know, 11 o'clock when you're thinking about a problem. And so I found similarly that AI can be a really great, uh, again, co-pilot or partner, um, in some of those moments where you just need a quick check.

    12. LWLucas Werthein48:21

        Can I just mention, going back just, just to the synthetic client, uh, just so I don't get in trouble with-

    13. CVClaire Vo48:27

        Yeah

    14. LWLucas Werthein48:27

        ... with my clients, it's important to highlight that none of the information that we put in these synthetic clients are proprietary. They're all available on the internet, and so we're not training any of the models with the information that we get from our clients. It's just an exercise-

    15. CVClaire Vo48:44

        Publicly available

    16. LWLucas Werthein48:44

        ... that we run through presentations publicly available. Yeah, so-

    17. CVClaire Vo48:47

        Yeah.

16. 48:48 – 51:36

    Final thoughts on AI reliability and evolution

    1. CVClaire Vo48:48

       Great

    2. LWLucas Werthein48:48

       Um-

    3. CVClaire Vo48:49

       ... Well, this is awesome. You, you know, I'll wrap with our, our final favorite question, which is, you are clearly an expert prompter, which I love to see, but when your coach is giving you bad advice or, you know, the AI is not responding how you like, you seem like a very reasonable person, so you probably act quite politely, but what is your go-to tactic? How do you, how do you get AI back on track? Um, do you ever find yourself frustrated? What do you do?

    4. LWLucas Werthein49:17

       Not frustrated, um, and I think this ties into how the models have evolved. As we see the iteration of models, I see definitely an evolution of how it hallucinates less or it makes up less things. I do think that putting guardrails around how we're allowing it to think and not necessarily access outside information makes it a little bit easier. Um, and so when it gets something wrong, um, I see it as the evolution of technology. This is brand-new technology. It's gonna get it wrong, and I try to perhaps help it, like I help my children when they, when they get things wrong. [chuckles]

    5. CVClaire Vo50:02

       I, I have said this consistently on this podcast, the answer to that question is always a reflection of your parenting tactics and, and strategies. Well, Lucas, this has been amazing. Thank you so much for sharing your coach. I'm actually gonna go spin off into ChatGPT. I have a really good idea for one that maybe I'll share in the show notes as well. I appreciate you joining How I AI. Where can we find you, and how can we be helpful?

    6. LWLucas Werthein50:24

       Well, you can find me at... I'm actually off social media.

    7. CVClaire Vo50:28

       Oh, um, so we can't find you.

    8. LWLucas Werthein50:30

       I'm off social media.

    9. CVClaire Vo50:31

       Where can we find-

    10. LWLucas Werthein50:32

        You can't find me

    11. CVClaire Vo50:32

        ... your company?

    12. LWLucas Werthein50:34

        You can find my company. It's Cactus.is, as in Sam. Um, and that's where I spend most of the t- my time, to be honest. So if you wanna find me, just, just go to Cactus.

    13. CVClaire Vo50:46

        Great. Go, go to Cactus, and find him at that amateur tournament, tennis tournament on September 18th.

    14. LWLucas Werthein50:53

        Or, or challenge me to, to a tennis match. That's, that's the other way-

    15. CVClaire Vo50:56

        Challenge you to tennis

    16. LWLucas Werthein50:56

        ... to get my attention. Yeah.

    17. CVClaire Vo50:57

        Perfect. Well, thank you so much for joining. I really appreciate this. It's been a great conversation.

    18. LWLucas Werthein51:02

        All right. Thanks, Claire. It's a pleasure to meet you, and thanks for having me on the show. It's been wonderful. [upbeat music]

    19. CVClaire Vo51:08

        Thanks so much for watching. If you enjoyed the show, please like and subscribe here on YouTube, or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at howiai pod.com. See you next time! [upbeat music]

Episode duration: 51:36