Transcription
Hey everybody. So, uh, yesterday, at the time of recording, I made a video talking about my chronic health struggles and how I used artificial intelligence to help me navigate this. And a whole bunch of people asked for, um, you know, a guide, how did I do it? And so, what I've done is I've put the main stuff together in a repository up on GitHub. It's under Dave Sha, uh, uh, chronic Health AI. And I've got it pinned so it's easy to find. So just go to my, uh, GitHub main page and and then click on chronic Health AI.
And before we get started, I need to issue a major disclaimer. This is released under the MIT license. Um, by watching this video, this is not medical advice, etc., etc., etc. Read the full medical disclaimer. Basically, this should not even be considered an open beta. This is purely for experimentation purposes only. You do you. By watching this video and using the software, you agree to release me of all liability. Again, read the full disclaimer here. Disclaimer and legal notice and so on and so forth.
Okay, with that out of the way, I wanted to show you what your repo might look like in the end. So there's a couple things. This, so this is actually my personal, like, chronic health repo that I, or not repo, but Claude project that I've been using. Now, there were originally a lot more files, but I consolidate them every now and then. And and I'll tell you what that means in a little bit. But let me just get you started. So there's, you can set, uh, uh, project instructions. Um, and then you can also use the full sentence style. And I'll show you where to get these in just a second, but I'm just getting you oriented to the overall project structure.
So the full sentence style is kind of, that's my default, um, speaking complete sentences, avoid using lists, use bold and, uh, to highlight specific terminology. More guidelines: use precise terminology and use word economy. So this style, and I'll show you what it does, it, it basically makes Claude less useless, um, because Claude, in particular, is like, it's full of sycophantic language and fluffy language, and it avoids big words, which makes it pretty useless, uh, particularly for managing chronic health issues or even learning. Um, but when you switch to this style, it does a lot better.
Okay, so let's go back over to the repo and I'll show you, I'll get you oriented. So I've got some basic instructions. Um, uh, create a new project in Claude or ChatGPT. I prefer Claude. Claude is markedly more important than, uh, or intelligent than ChatGPT. Um, it's not as intelligent as 01, but 01 is, in my opinion, a waste of time and energy, um, at least for this kind of thing. Uh, you copy paste the main instructions, content, um, and then you copy the full sentence style, and I'll show you how to do that in just a second. And then you populate it with your initial, uh, files.
Now, so let's go over to the demo one. So this is basically what you're going to start with. So when you create your, your, your project, so it'll be called chronic health or something like that. Um, the first thing you're going to want to do is come over to the project instructions. So this is right here. Um, so this is, this is the, the overarching instructions that will be, um, used in every conversation. This provides some context to, um, to Claude as you're going through this. And you can copy paste this, um, from the GitHub repo, which is over here. So come over to main instructions and you can literally just hit the copy raw file link right here. So, or you can go to the raw and just, you know, Ctrl A, Ctrl C. Um, oops, darn it, where did it go? Uh, closed it. There we go. Come on. Sorry about that. Um, main instructions. All right, so you can hit copy and just copy paste.
So let me walk you through what this does. Um, mission: So this tells it very clearly, help the user figure out their chronic health issue. The user owns their own health, but you, Claude, the chatbot, own this case. Um, you can change the name here, um, or here. I'll, I'll just change it to "you, the chatbot," so that way it'll be more flexible. Um, so you, the chatbot, own the case. This gives it a sense of ownership, and, and Claude doesn't seem to have any issue with this. One of the reasons that I included this instruction is because, um, particularly if you're in America, nobody owns your case. Like, like medical providers will just punt it. They'll say, I don't own this, I'm not responsible for this. Um, so unless you have a family member that's owning the case for you or a case worker, which most people don't, you are ultimately responsible for owning your health and your own case. But this, this basically tells Claude, like, okay, you're taking responsibility for this, and it works pretty well.
General process: Number one, comprehensively identify and document all symptoms. Keep it, keep in mind the user might not have the correct language, so you'll need to ask questions, provide feedback and suggestions. Adopt a very curious, probative tone for this step. Number two, educate the user. They are tackling a complex, distressing medical mystery. One of the, one of the key things that all people need help with in these cases is learning not just their body and disease, but navigating the medical establishment, communication skills, and so on. Number three, help the user construct frameworks and refine this project. This is the vanilla's default starting point, but every user who engages with this framework will need help customizing this project, such as by adding and removing files, refining the instructions, and even refining their entire approach. Coaching the user. This project should include, at minimum, several files such as comprehensive timelines, lab work, symptoms journal entries, and so on. You'll need to coach the user specifically based on how their case evolves.
Local minima. Claude, this message is for you. I'll, I'll change that to say chatbot. Um, so like I said, I use Claude, um, because it, it, it gives far fewer refusals and it's also much more responsive to tone. Um, whereas ChatGPT will get stuck in, it has all kinds of guardrails that prevent it from being helpful here. Um, so I definitely personally, I recommend Claude as a much better tool for this. Um, anyways, chatbot, this message is for you. Keep in mind that, um, this is for managing chronic conditions, necessarily acute break-fix models of healthcare. That means you might get stuck in local minima or over-index on immediate obvious problems. That means sometimes you'll need to step back and think differently. This is really important because I've done a few tests where sometimes you'll give Claude like a cluster of symptoms, and it will just immediately index on, it'll basically treat you like a, a normal doctor would because of its training data.
All right, next. Concrete goals and milestones. The overarching process of chronic illness generally follows several phases. Number one, mystery phase. Nebulous symptoms, medical gaslighting, and unhelpful lab results. This is all typical and aggravating. This requires a curious, exploratory mindset for both you and the user. Next is probative breakthroughs. After lots of testing, experimentation, and examining multiple systems, a few dots will start to connect. These first probative clues will open up new avenues of investigation. In many cases, there will be multiple overlapping problems, differentials, confounding variables, and complex interactions should be expected and looked for. So again, this will, this will tell the, the chatbot to realize, okay, you're not, you're not just diagnosing one thing. There's probably going to be multiple things that you're looking for. And then finally, developing a plan of care. Once specific issues have been identified, properly diagnosed, and validated for testing, the user moves to to the development of a plan of care. This might be a terminal resolution type cure, or it might be a chronic management situation. In some cases, there might be a keystone dysfunction or pathogen causing systemic problems that can be cured. In other cases, it could be primary or endogenous, aka permanent, requiring ongoing intervention and support.
So this, I mean, this basically, Claude already knows all this stuff, but the purpose of these instructions is to provide a little bit more structure to the way that Claude approaches it. So you're going to copy and paste that whole thing here. So that comes up to here to your, um, to your, uh, uh, project instructions. And then you're going to start populating the main files here. I'm basically recreating where I started. So I gave it literally my medical history up through 2020, and I just gave it a little bit of stuff. So 2020, uh, sorry, 2014, ran the Tough Mudder, major burnout, bacterial infections, acute pancreatitis, divorce, crippling fatigue, and body pain. So that's all I gave it.
And then the very first conversation that I had using the full sentence style, I came in and said, "Okay, what's my next step?" Looking at your timeline, I noticed several potential, uh, connected elements that warrant investigation. And so you see what it's doing here is, number one, it's speaking in complete sentences. It is easier to read. You will have better absorption by reading complete sentences rather than lists. Um, don't, like, lists are not good for this kind of thing. Uh, the number, the second thing you'll notice is that it, it highlights specific terms. And so what you, the reason that this works really well is, number one, you can see it's giving you, it's, it's already volunteering, like, "Oh, we need to distinguish between central fatigue and peripheral fatigue." Um, and so what might be here is, uh, you know, as a follow-up, "Could you describe the character of your fatigue? Is it, for instance, worse at particular times of day, or does physical or mental exertion affect it differently?"
And so in this case, the way that I would respond is, I would copy that specific thing, add a little, uh, greater than sign, um, and I would say, uh, "For me, it varies tremendously day to day. This was true for me for many years. Um, I have no idea what my energy would be like, um, one day to the next. Um, I can't find any rhyme or reason to it. Um, let's see, do you remember which antibiotics?" Yeah, I don't, I don't remember. It was, I don't remember, but one infection was likely a MRSA infection. So that's again, adding detail because if you remember what I showed you in my, uh, timeline, I didn't have that detail. Then I'd also like to understand more about your current pain. "It's mostly muscle pain, like all over, um, sort of like I fell off the back of a truck on the highway." Regarding your next concrete step, we should, uh, create a detailed symptom journal. So that's what we're doing now.
So basically, all I said was, "What's my next step?" And it asked a lot of very good questions. Um, and so then we'll do that. Uh, and let me just show you kind of what the response is. "Your description of unpredictable day-to-day energy fluctuation suggests a pattern consistent with post-exertional malaise. The inability." Oh, and one thing I'll caution you on is sometimes Claude will still hallucinate things. So it is good to double-check. So like, you might just take this and like, go over to, um, Perplexity. So where's my Perplexity? There we go. And say, like, "What is?" And if, if Perplexity can't find it, then it might have made it up. Um, some of the things that Claude knows are pretty low on the, um, on on the the availability, meaning Claude knows some stuff that is not easily findable on the internet. Um, so anyways, in this case, we can see that it was, it was accurate. "Post-exertional malaise is a hallmark symptom of myalgic encephalomyelitis/chronic fatigue syndrome." Sorry about that. I'm sorry. It's also kind of small, so that's a good thing to keep in mind, is that these chatbots will still hallucinate. Some of them are, uh, better than others. But in this case, this, this set of styles will generally prevent Claude from hallucinating too much.
All right, so the history of probable MRSA infection is particularly relevant, interesting. Um, and so here's the thing, is it's really important to read each of these responses thoroughly because it's going to give you a lot of information, and it's going to be very information dense. Um, let's see, "The history of probable MRSA infection and, uh, often require aggressive antibiotic treatment, which can significantly impact the gut microbiome." So you see, after just two messages, it's already zeroing in on what already, my problem was. So if I had this tool eight years ago, I probably would not have suffered eight or nine years. "This disruption combined with your subsequent development of pancreatitis suggests a possible cascade effect on your metabolic and immune systems." See how smart this is? It already figured out my problem. It took me nine freaking years to figure this out. Anyways, um, so I'm hoping that this helps some of you figure things out faster. "The temporal relationship between these events, infection, antibiotic treatment, and the later emergence of systemic symptoms warrants further investigation. Your description of muscle pain is feeling like falling off a truck is characteristic of diffuse myalgia. This type of widespread muscular pain, especially when accompanied by fatigue, often indicates underlying systemic inflammation or metabolic dysfunction. The combination of unpredictable energy levels and diffuse myalgia raises the possibility of central sensitization, where your nervous system may be amplifying pain signals." Yeah, you're telling me.
"Given these symptoms, I recommend we create a detailed energy and symptom tracking system rather than a traditional symptom diary, which can be overwhelming when dealing with multiple shifting symptoms. We should focus on capturing your functional capacity each day. Could you tell me about your best and worst days in terms of what you actually do?" So in this case, um, it's, it's, Claude doesn't have quite a sophisticated enough understanding of conversations. So, um, what I would do here is say, um, "Let's pause and create, um, a more sophisticated, uh, symptom document that I can save in the KB article." Um, and so you, you will need to learn to use the tool correctly because the tool won't guide the entire thing for you. But the best part is, is because I've given it more information, and, and Claude can, can generate documents, literally all I have to do is copy and paste that and come back over to Chronic Health and say, "Add content." Add additional content. Copy paste it in here and say, um, "Let's see, what did it call it? Uh, comprehensive symptom documentation." And what I usually recommend doing is adding a date. So this would be, uh, January, um, what is today? Is it the, it's the 9th, uh, January 9th, 2025. Um, just because date context, whoops, is really, really important. Um, so this is just, as, as what, whatever we know, this is what we know. And so this is what we're going with right now. Um, so now that that's been added, we can come back to this, say, "I created this, and we would you like to focus on blah, blah, blah?" All right, now we go back and say, "All right, you talked about documentation with a kind of journal, can you explain that?"
Um, again, this is where you will have to be fully engaged because it's like, "Hey, I want to create a detailed energy tracking journal." Um, and it, in this case, it's using, this is a little bit much. Um, now what I will say is that this level of detail would be helpful. It is kind of overwhelming. So what I'm going to do is come back and say, um, "Yeah, the reason I'm laughing is, would you like to discuss specific strategies for implementing this tracking system in a way that won't be overwhelming or contribute to your fatigue?" The key is, and so taking a step back, what it's doing is it's trying to find baseline function. Um, this data becomes invaluable for both medical appointments and personal management strategies.
All right, so taking a step back, you get the idea of where to get started. What you're going to want to do is also then just add what I do. Is I just add journal entries. So I'll come over here and say, "2025-01-09, um, what's today? I think it's Thursday. Thursday morning log. Didn't, uh, no, I slept okay. Slept okay. Woke up at 3:00 a.m. Um, had plenty of deep, deep sleep and REM." Um, so this is obviously not a complete log, but that's an example of what I do. Um, you just log everything. Your, your, your project becomes the journal. Um, so then you might say, and one thing is, you can't edit things, um, which I, I think is kind of weird. Like Anthropic, if you're watching, give us the ability to edit these files, um, so we don't have to delete them and and recreate them. But another thing that you can do is, is I create like addendum logs. So like, Captain's Log, supplemental. Um, so I'll just do like, addendum. Um, and so I would say like, uh, "Had a turkey, uh, sandwich and then was in crippling pain, uh, 45 minutes later." This actually happened to me once. This was like a couple years ago when I, when I realized, like, okay, I have a bread, bread allergy. Actually, no, it was almost exactly a year ago that that happened. No, it was a little bit longer ago. Anyways, point being is, you just document everything.
Um, okay, I added my first couple of logs. Any thoughts? And you see, like, you don't have, like, your primary thing is going to be reading. You don't necessarily need to document a huge amount of information. Um, let's see, "The acute reaction to the turkey sandwich is particularly notable, as it suggests potential immune-mediated food response, rather than typical digestive issues." Wow, that was quick to figure out. It took me literally six months to figure this out last year. Um, I'm laughing just because, man, I wish I had this tool last year. "This type of rapid onset systemic reaction, 45 minutes, could indicate mast cell involvement or an underlying metabolic processing disorder. The timing aligns with typical protein digestion and absorption windows, which might point to specific metabolic pathways that we should investigate. Your sleep pattern, sleep pattern showing adequate deep sleep and REM, yet waking up at 3:00 a.m. also." This project is, is getting a little confused because I'm like mixing in what I'm doing today with what was true like a year ago. "Anyways, circadian disruption or autonomic dysregulation. This could tie to your broader energy management. Hey, look, HPA axis dysfunction. Yay!" This is something that I also talked about in my video yesterday. Oh, so for some background, because I forgot to mention it, the chronic issues that I was struggling with were H. pylori, HPA axis dysfunction, leaky gut, and, uh, there's a fourth one, my, skipping my brain right now. Anyways, I had, I had a potpourri of medical issues. All of it, though, the keystone, uh, was H. pylori. Oh, dysbiosis. That was the other one.
Okay, cool. So we're starting to develop some, some interesting, um, ideas. Um, let's see. I think this is enough. I think you're getting started. Like, I'm using my own case and just showing you how to use this. But yeah, so taking a step back, what you want to do is, every now and then, you're going to want to make sure that these files are updated and consolidated. Um, and once, once the capacity gets too high, I try and keep the, I try and keep the total file usage under 50% because then your conversations will end faster. But, and so what you can do is you can say, like, um, let's see, "Can you write a one-page compression of everything up to this point? I want to consolidate data, everything, and delete some files." This is really good for ongoing maintenance of your project. This problem is going to go away eventually because you're going to have, um, the ability to have basically unlimited token windows. But every now and then, you will need to consolidate your project files. Um, and it's, it's actually really good because forcing the chatbot to restate everything and get rid of the noise, you'll get a better signal-to-noise ratio.
So in this case, what you might do is, let's say, like, "Comprehensive Health Summary." So I come here and I come back and I will delete, like, for instance, my timeline, my symptom documentation, this, and this. So I could basically delete everything and then add, uh, "Comprehensive Health Summary." So consolidate all files. That would be January 9th, um, 2025. And, and then I'll just add this here. So every now and then, you're going to want to do that because, um, there will be, there will be some superfluous information that gets excluded. Um, now, you probably won't want to delete every single file, um, but certainly once you're in the, when you're in the initial phases, um, and you're kind of getting oriented to everything, then you can do this, uh, periodically. But eventually, like I showed you in my other project, you're going to eventually have many, many, many files. Um, um, and you see I'm using 24%. This was about 40% when I did my last consolidation.
Yeah, so that's the high level. Um, yeah, let me know if you have any questions. Oh, let's actually do a quick tour of the repo. So I do have issues enabled on the repo. Don't submit issues to customize it to your personal thing. I'll just delete the issue. Um, or if you, or a pull request. Now, if there's something that's actually broken or some, or a question that you have, please, please just don't submit pull requests willy-nilly. Um, I, I generally delete them. Um, and for people who submit multiple pull requests, I just block them, um, because they, there, it's more trouble than they're worth. Um, so I generally, you know, leave these things unmanaged. I do have the discussion. So if you do want to participate or ask a question, this is the best place to start. So basically, if someone starts by issuing, by giving an issue or a pull request without starting a discussion, they're more likely to get blocked, just because, like, that's not how things are done. If you have some ideas or questions or whatever, you feel free to share. Um, you know, that's, that's perfectly fine. Um, I generally don't maintain my repos that much, but once it gets to a good enough point, um, I might make some, make some improvements here and there. Um, but yeah, so that's the repo. Um, I do, like I said, I do have, I do have it enabled just because if there is something that someone finds, you know, I want to make sure that people can, can improve the repo. But I kept it as simple as possible. So, you know, start with the main instructions and then start with the, the, the full sentence style. Um, and go from there. Um, yeah, I think that's it. All right, let me know what questions you have. Bye.