Imagine spinning a wheel every hour and getting hit with crushing fatigue, brain fog, lightheadedness, or nausea 20% of the time—randomly, unpredictably, wrecking your ability to function. That's the diagnostic nightmare one AI researcher found herself living after two brain surgeries for a prolactinoma (a tumor on her pituitary gland) last year. The tumor was controlled with medication, but these mystery episodes made planning anything impossible. Walking to a grocery store felt insurmountable. She could barely string sentences together mid-episode. Then she applied the same rigor she uses evaluating AI agents to her own health—and cracked the case in about a month.

The Problem With Modern Medicine

Cardiologist Eric Topol wrote in Deep Medicine that 'patients exist in a world of insufficient data, insufficient time, insufficient context, and insufficient presence.' This researcher—who asked to remain anonymous but works at an AI safety organization—realized a good process with AI solves all four. She isn't claiming models outperform specialists like her neuroendocrinologist (the country's top expert on her condition), but she does make a bold claim: 'an AI-literate patient running a good process with a frontier model can outperform most PCP visits for ambiguous, multi-system symptoms.' The key word is 'process.'

Track Everything, Then Feed It to the Models

The first step in her system is tracking—logging symptoms and possible causes obsessively. She set up an hourly log from 9am-10pm recording energy levels (1-5 scale), symptoms, and notes on potential triggers. Daily logs tracked menstrual cycle day, days since last tumor-medication dose, sleep scores, calories, carbs, exercise, stressors, and average hourly energy. Her stack: Garmin watch for sleep and steps, Cronometer for food intake, a continuous glucose monitor, and Notion spreadsheets auto-populated by an AI agent she built. She emphasizes that 'no amount of context is too much'—especially compared to the ~11 seconds average doctors let patients talk before interrupting. The goal isn't just data collection; it's creating longitudinal baselines so you can actually measure whether interventions work, not just vibe-check them. Without tracking, she's seen her own prolactinoma diagnosis delayed 2.5 years because she never took irregular periods seriously.

Test Broadly, Then Analyze With AI

The second step is testing: blood work and specialized tests run in parallel with symptom tracking. She used Function Health for a panel covering 100+ biomarkers at $365/year—pricier than individual PCP-ordered tests but more efficient for broad coverage. Given her pituitary history, she got a full hormone panel plus an ACTH stimulation test (which her nurse practitioner later suggested independently). At-home blood pressure monitoring screened for orthostatic intolerance patterns. Then comes the real power move: analyzing everything together with reasoning models like Claude Opus 4.8 or GPT 5.5 using high thinking effort. She uploaded months of Garmin data, exports from Cronometer and her CGM, all test results, and asked specific questions—'What hypotheses might explain my symptoms?' 'Why do my symptoms tend to hit in the mid-afternoon?' The models found patterns she couldn't see: roughly half the time, eating a snack improved symptoms. That single insight pushed her toward a dietitian consultation, which revealed she'd been under-fueling by ~300 calories daily for months—likely putting her body into energy-saving mode.

Experiment Systematically

The final step is experimenting with interventions based on analysis, then measuring results against your baseline. Her experiments fell into three risk tiers: no-brainers (iron supplements when ferritin was below range), harmless-but-maybe-useless (creatine post-workouts), and risky/costly (stopping tumor medication or pausing exercise—which she did one at a time under physician guidance). She stopped her medication for four weeks, tested prolactin levels, then restarted at half dose since levels were too low. Her endocrinologist also suspects low estrogen may contribute and is testing it across her next cycle.

Key Takeaways

  • Use reasoning models (Claude Opus 4.8, GPT 5.5) with high thinking effort for health analysis—worth the $20/month subscription cost
  • Create a dedicated project in ChatGPT or Claude to build memory around your specific health context over time
  • Track symptoms obsessively: hourly energy logs + daily metrics + test results give AI something useful to analyze
  • Be specific when sharing data: 'T4 is not free T4; estrogen is not ultrasensitive estradiol'—vague inputs produce wrong outputs
  • Run experiments one variable at a time with clear success/failure criteria based on your pre-intervention baseline

The Bottom Line

This isn't about AI replacing doctors—it's about patients using these tools to show up to appointments already armed with data and hypotheses, turning ambiguous symptoms into trackable, verifiable problems. 'You don't need to understand medicine,' the author writes. 'You don't need to be technical. You don't need anyone's permission. Because by default, nobody is coming to save you.' That's hacker energy applied to healthcare, and honestly? It's exactly how more people should be approaching their own wellness journeys.