Top-Tier AI, Information Parity for Health—Managing My Own Health with Claude Code

This entire episode is personal sharing. It does not constitute medical advice of any kind.

Friends still working at companies, at the big tech firms, keep asking me the same thing: an agent as expensive as Claude Code—beyond me using it to earn money in my own business, what can it actually do for me personally? Today I’ll share one answer: I use Claude Code to manage my own health.

First, some background. For two of the past few years I was at Xingshulin, running a “personal physician for entrepreneurs” business: helping entrepreneurs manage their health, connecting with a lot of top-tier hospital specialists, delivering online health management consulting. So I have a decent sense of what serious health management looks like. My core view: AI isn’t here to replace doctors entirely. It’s here to help doctors keep our health records in far more detail, so that when we go see a doctor or go in for tests, he sees more of the picture and arrives at a better diagnosis and treatment plan.

A six-month follow-up plan

I had my last full checkup around late December and early January. That plan came from the AI, built off my earlier checkup reports and my genome sequencing report. It started by pulling every annual men’s checkup package the big health providers offer, with what each costs and what tests each includes. It analyzed them, recommended an annual package for me, then picked out a set of add-ons on top of it. And at the end it very thoughtfully flagged two or three tests the annual package simply can’t cover, then told me which department to book at a hospital like Beijing Hospital to get those done.

Six months have passed, and things happened in those six months. I developed allergy symptoms for the first time in my life, for instance. I grew up in Beijing and had never had allergies, so chalk it up to six months of short sleep plus all the late nights with AI. Every symptom I developed in this stretch, every treatment and test I had, I synced all of it to the AI. It built me a six-month follow-up plan that spelled out exactly which few things deserved my attention most.

A few examples, one by one:

• Cubital tunnel syndrome: go back to neurology for follow-up. Anyone who lives at a computer knows this one—it’s an occupational hazard. Keep your elbow bent while typing long enough and your little finger goes numb. The plan it drew up this time is an EMG plus high-frequency ultrasound of the ulnar nerve at the elbow, for an accurate, quantified read on how it’s progressing. On feel alone, all I can say is “more numb” or “less numb.” Over these six months I added vitamins for it and did exercise-based therapy, and these tests let me put a number on whether anything actually improved, and whether surgery is next.
• Fatty liver / liver fibrosis: its recommendation is grounded in my genome sequencing results. My family genetics make me more prone to liver fibrosis. I’m a Liaoning–Shandong mix and I can hold my drink, but the genes say that while I metabolize alcohol reasonably well, alcohol as a liver-damaging substance hurts me more than it hurts most people. My last checkup showed mild fatty liver, so it drew up a set of quantified liver tests.
• Clot-related markers: my hematocrit runs high, and I also ran a guesthouse in Yunnan for four years. Live at altitude long enough and your blood cell count rises, which makes clots more likely. A routine checkup doesn’t usually go looking for these. But the genetic risk is established, and the AI knows my full history of conditions and treatments, so it tells me to watch these particular markers.
• Plus the cardiovascular panel you should be doing in your forties; this year’s allergy problem (I’m on ebastine now, and it hands me a set of medication warnings, because it knows that among the dozen-plus supplements I take, curcumin interferes with the metabolism of ebastine’s main compound and pushes blood concentration into dangerous territory); the subtyping for my elevated uric acid; and targeted early cancer screening based on family genetics—it didn’t tell me to test for everything, it flagged tumor markers for two specific cancers, and it also lists the ones I tested six months ago that don’t need repeating this cycle.

It gives me a plan I can act on, too. What annoys us most in practice is not knowing which department, which hospital, how to book. It can look all of that up and make recommendations.

The key is being traceable to a source

And these recommendations don’t come with the routine you get from some agents: “sorry, I was wrong, I hallucinated again.” A genuinely good agent, given a good workflow and some real data sources, can produce content it stands behind.

In my setup, the PubMed research API is wired into Claude Code. Every judgment, decision, or recommendation carries the reference behind it: based on this specific paper, combined with your current situation, here’s the recommendation. Every reference link, every source, traceable.

Supplements, finer-grained than a personal physician

I’ve been taking supplements for a long time, and back when I ran the personal physician business I was a personal physician customer myself. But when my doctor fine-tuned my supplement regimen, there was no way he could get anywhere near as granular as the AI does.

Buying them, too. I used to order from Tmall, but honestly I lost my nerve. Those imported supplements all claim to be sourced overseas, and since I studied art I notice capsule color—every batch I bought looked different. Counterfeits are rampant. So starting last month I switched to iHerb (a US supplement site that ships from Hong Kong) and wired Claude Code into its Hong Kong site. It can see the exact composition of each supplement, the price of a given package size, and whether it’s out of stock. Then it gives me a purchase list: what to adjust, what dose, why that dose, when to take it (with breakfast, every other day, with dinner, or before bed). Detailed recommendations on all of it, and it tells me which of my own circumstances each one rests on. Where a marker is still uncertain, it flags the recommendation as preliminary and says it’ll give a sharper target once the other tests are done.

It even explains the science. The broccoli sprout powder I take isn’t a conventional supplement, and it lays out what the active compound is and what conditions degrade the effect. Or take the noise a while back about fish oil being useless. It goes and finds the papers and tells me plainly which population each null-result study looked at, what the intervention was, how big the sample was, and under what conditions the effect vanished—and under what conditions it does show up. CoQ10 and fish oil may not deliver much in the areas they get hyped for, but for lowering my blood lipids and improving specific organ function they still earn their place. Watch influencers online and all you get is the conclusion. A report you get after wiring in these sources and letting it read your own data is genuinely something money can’t buy.

Information parity

Never mind how little a Claude Code subscription costs per month. Even if I went to the single best health management physician working today, he wouldn’t have the time, the energy, or frankly the capacity to read all my reports. Think about it: the genetic data alone runs over 1.2 million rows, and that’s only the portion current science can tie to specific health issues and drug responses. Already that large. No doctor could ever produce a report this detailed.

So beyond using AI to “pay it to work for us,” top-tier AI and top-tier agents are enormously useful for ordinary people too—it really is a breakthrough product for information parity, for parity of intelligence.

That’s it for today. Leave a comment if there’s something you want to hear about, and I’ll see you in the next one.

Source: EP0032_audio.mp3 · ASR model gemini-2.5-pro (chunked parallel) · full text of the roughly 10-minute original recording

[00:00] A lot of people have been asking me lately: look, you run your own company, so you use AI to do business and make money. But for people like them who are still at a company, working at a big tech firm, what does an agent as expensive as Claude Code actually do for them personally, beyond paying out of pocket for work? So today I’m going to show you how I use Claude Code to manage my own health. This is purely me sharing what I do—there’s no medical advice here. Because as you know, over the past few years, for two of them I was at Xingshulin doing

[00:26] the personal physician business for entrepreneurs. Meaning we helped entrepreneurs manage their health. We’d reach out to a lot of top specialists at Class-A tertiary hospitals to do what amounted to online health-management consultations. And honestly, now that we have AI, this is something we can do at a pretty decent level at home ourselves. And not that it fully replaces the doctor—just that it can help the doctor manage our health records in far more detail,

[00:52] so that when we do go see a doctor and get checked at a hospital, the doctor can consider things more comprehensively and arrive at a better diagnosis and treatment plan. So first, let’s look at—I had my last physical late last December, early January, I think. And for that physical, AI drew up my testing plan based on my previous physical reports and my genome sequencing report. At the time I used AI to get

[01:19] every men’s physical package Meinian Health offers—how much each package costs, and which tests it includes. And after AI analyzed all those items, it recommended an annual checkup package for me, and on top of that it picked out a series of add-ons for me. And in the end there were maybe two or three items that couldn’t be done at Meinian. So it very thoughtfully told me—it said I could go to a hospital like Beijing Hospital, which department, what kind of appointment to book, to go get those supplementary tests done.

[01:44] So it’s half a year later now, and a few things did happen in that half year. For instance, for the first time in my life I developed allergy symptoms. Because before, having grown up in Beijing since I was a kid, I’d never had allergy symptoms. It might be because I’ve slept less this half year, plus staying up late using AI and so on—for the first time I developed allergy symptoms. So any symptom I had during this period, any treatment I did, and the tests I had done recently, all of it got synced to AI. And it went and built me this six-month follow-up plan.

[02:09] And in here, very clearly, it tells me what the few things most worth watching are. So let’s go through them one by one. For example: the first one is, go to neurology and re-check my cubital tunnel syndrome. Anyone who’s on a computer a lot knows cubital tunnel syndrome—it’s a desk-drone’s disease. Your little finger goes really numb, and it’s numb even when you’re not moving it, brought on by holding your arm bent for long stretches while typing on a keyboard. So the plan it drew up for me this time,

[02:34] the follow-up plan, says to get an EMG plus high-frequency ultrasound of the ulnar nerve at the elbow, so we can understand how the condition is progressing more precisely and quantitatively. Because if I only go by how it feels, all I can say is it’s more numb or it’s less numb. Because over this half year I’ve been deliberately supplementing certain vitamins, and doing some exercise-based therapy to relieve the problem—a series of tests like this can quantify for me whether the problem has actually improved,

[03:00] and whether I should go get surgery next to solve this thing. Then there’s the classic one, fatty liver. You can see that its recommendation is actually based on some of the results from my genome sequencing. And in my genome sequencing, my family genetics make me more prone to developing liver fibrosis. If you’ve seen my earlier videos you’ll also know that even though I’m a textbook Han Chinese, the fabled Liaoning-Shandong mix, pretty good at drinking, what my genes tell me

[03:25] is that even though I metabolize alcohol reasonably well, alcohol, as a substance that causes liver damage, actually does more harm to me. So since my last physical showed mild fatty liver, it went and drew up some quantitative liver tests for me, including tests for certain thrombosis-related markers.

[03:50] Because of my hematocrit—and I did run a guesthouse in Yunnan for four years, and if you live at altitude a long time your blood cell counts go up too, so you’re more prone to clots. And in a routine physical, nobody specifically tests those markers, or nobody even looks at them. But once the genetic risk is a known quantity, and once AI knows my whole history of illness and treatment, it’ll specifically have me watch certain markers, or add certain tests.

[04:16] And after that, a whole run of cardiovascular tests—I am over forty, after all, right? And then there’s the allergy issue that came up this year. I took different types of allergy medication, and some worked and some, some didn’t. And the drug I’m on now, ebastine—it’ll also give me a whole set of medication reminders. Because it knows I’m currently taking maybe a dozen-odd nutrients, and among those nutrients, curcumin actually affects the metabolism of

[04:42] ebastine’s main active ingredient, which raises the drug’s blood concentration and creates a degree of danger. So it actually gives me a whole set of alerts like this. Then the metabolic stuff, right—diabetes, tests for insulin resistance. And then including high uric acid from previous tests: say I don’t have gout, but my uric acid has always been high. So it’ll use certain tests to determine what my uric acid subtype is,

[05:07] and which drugs should be used down the line if it gets more serious. And with the nutrient supplements, there’s the whole business of not overdosing on them. Because when you take nutrients, the bottle might just say take this so many times a day, and when to take it, and roughly how much, right? But actually it’s different for every person, right? Everyone’s body weight is different, and everyone’s absorption of different nutrients—whether the precursor substance or the nutrient itself—is different too. All of that actually matters.

[05:33] So for some people this dose just isn’t enough, and for others this dose is too much. And there are actually a lot of targeted tests in here too, including early screening for certain cancers, which also ties back to my family genetics and some of the markers in my current test results. Would it have me test for everything? No, but it roughly—I remember what’s flagged in here is that it recommends I check the tumor markers for two, two cancers. And then there are some things I already had tested half a year ago but don’t really need re-tested this half year—it lists those out too.

[05:58] And it gives me a plan I can actually act on. The most annoying thing for us day to day is that I don’t know which department to book, right? So which department, or even which hospital, and how to book the appointment—it can look all of that up for me and then give me the matching recommendations. And the recommendations in here aren’t like a certain other bot, right, where it’s “sorry, I was wrong, right, I hallucinated again, this time I’ll give you a responsible answer.” It’s not like that. A genuinely good agent, given a good workflow and some data sources, can produce responsible

[06:24] content like this. And it can also go do a second round of verification. So this Claude Code of mine actually has PubMed’s paper API wired into it, meaning for every judgment, decision, or recommendation it makes in here, it has the corresponding reference document—like, based on such-and-such a paper, and given what your situation is now, here’s the recommendation I’ve made. All the reference-document links and sources

[06:49] are fully traceable. And beyond that, I’ve also been taking nutrients all along, and back when I was in the personal-physician business I was a personal-physician customer myself. And when my doctor was doing precision nutrient management for me, there was no way to get anywhere near as granular as what AI does. So let me show you—this is the nutrient shopping list AI built for me. I used to have the doctor prepare it, and now, without a doctor,

[07:14] I used to buy from a Tmall storefront, but honestly, I really don’t dare buy there anymore. Why? Because for those imported nutrients I’ve been taking all along, they all claim to be overseas-sourced, but when it arrives—even granting that there can be batch differences, sure, the packaging does vary—but really, the color of the capsules. I studied art, right, so I look at the capsule color. I deliberately kept a capsule each time, and every purchase was different. I really don’t dare buy it. I think honestly,

[07:39] counterfeits are just too rampant now. So after going back and forth on it a bunch, starting last month I’ve been buying my nutrients straight from iHerb, which is an American nutrients site. And they ship from Hong Kong, so I plugged my Claude Code into their Hong Kong official site, and it can actually see the specific ingredients of which nutrient, and what a given, given package size

[08:04] currently costs, and whether it’s out of stock, and so on. And then it’ll build me a shopping list saying, here’s the adjustment I recommend for your nutrients, and the specific dosage, and why this dosage, and when I should take this nutrient—like with breakfast, or every other day, or with dinner, or before bed. It has a detailed recommendation for all of it, and it tells me what the current recommendation is based on.

[08:29] And for some recommendations it’ll say, because certain tests of yours, certain markers, are still uncertain, this is a preliminary recommendation, and once you’ve finished the other tests it’ll give me a more, more accurate recommended value. And with a specific nutrient, if you haven’t given it the exact parameters—like the broccoli sprout powder I take, which isn’t really a traditional nutrient supplement—it’ll go explain the difference to me, like what exactly the active compound in it is,

[08:55] and what kinds of things would make it work less well. Oh—like a while back there was that thing going around online that fish oil is useless, right? It blew up everywhere. But it can also go, like I said, find the reference literature and tell me clearly why the people saying it’s useless have a valid case, right—what situation they were targeting, what experiment they ran, how many samples, proving that under what conditions it has no effect. But it’ll also tell me under what,

[09:20] what conditions it actually does have an effect. Take coenzyme Q10, or fish oil, these—sure, in some of the miracle-cure areas the marketing claims, the effect may not be that strong. But for me, for lowering my overall blood lipids, for improving the function, the function of certain specific organs, it is still useful. So the thing is, when we watch those influencers online, all we get is the conclusion. But if we have Claude Code wired into these information sources, and it understands our own

[09:45] information, then the report we get truly is something you couldn’t buy with money—never mind how little a Claude Code subscription costs per month, right? If I go find a personal physician, even the most top-tier health-management doctor out there right now, he doesn’t have the time or the energy, or frankly the ability, to read all my reports. Think about it: the genetic data alone is over 1.2 million rows, and that’s only the part current scientific research can confirm is related to certain health issues or to medication—and that’s already that much.

[10:10] There isn’t a single doctor who could produce a report this granular. It really is—beyond us paying AI to work for us, for ordinary people, top-tier AI, a top-tier agent, is also enormously helpful. It really is a breakthrough product for information parity, for intelligence parity. All right, that’s a wrap for today. If there’s something you want to hear about, drop a comment. See you next episode, bye.