It helped me more

For twenty years I was afraid of needles, and it never got me anywhere. In July 2026 I turned the methodology I have been using for months to test language models on that fear. To prepare for one medical test, I had help from a model as well as from my doctor. The model got things wrong twice along the way, and both mistakes came to light only because I checked its advice.

This is one person's experience, not medical advice or a how-to guide. Decisions about treatment and medication were always made by my doctor. The model is not a doctor and does not replace one.

The same methodology as for models

It wasn't "I don't like it." It was panic, cold sweat and checkups put off year after year. And yet for half a year I have been running an adversarial AI lab where I test models on long stories. In July it occurred to me what should have occurred to me long ago: turn the lab on myself.

When I test a model, I don't trust impressions, I trust data across runs. So that is what I did with the fear, too. Calming down had never worked, so in came a protocol. Daily exposure with numbers: the wave climbed to 7, dropped to 2, logged. Breathing and muscle preparation treated as training. A checklist for D-Day. The fear stopped being the weather and became a metric.

What the model actually did

I worked in the Claude app with Claude Fable 5. The model could search the web. No miracles, four concrete things.

First, it explained the mechanism to me. A vasovagal reaction is physiology, not "weakness." According to the model, fear has two layers. One is declarative: "I know nothing is happening." The other is an alarm layer, and it doesn't read arguments, only experiences. For years I had been trying to talk a layer into something when it doesn't speak my language.

Then it helped build the protocol. One of my fears was the very sensation of feeling the needle go in. That's why several of the layers were aimed at exactly that. Each one was meant to block a different link in the chain:

numbing cream on the skin, covered with plastic wrap, not just spread on (for years, that was the difference between "I felt nothing" and "it hurt"),
a vibrating ladybug held against my arm to dull the prick,
lying down, not sitting, so I wouldn't faint,
tensing my leg muscles to counter the drop in blood pressure,
a thinner needle,
commentator mode: narrating in my head what is happening, like a play-by-play of an event ("the subject is lying down now, the nurse is disinfecting"); before the procedure it worked for a while, but in the moment itself I forgot all about it,
a reward at the end.

Above all, I rehearsed the whole thing in advance. Every day I went through it from the waiting room to the very end and wrote down how high the wave of fear climbed and where it dropped back to. That way my body was learning that the arc ends in calm, not horror.

No single layer was a miracle on its own. I was counting on them adding up.

It found evidence in my own life. For years I'd had experiences showing that my alarm system could be retrained. Only nobody had ever analyzed them. The model pulled this out of what I told it and showed it to me as data. That was perhaps the most valuable part.

And it helped me track down a stronger prescription numbing cream nobody had ever told me about. But I was the one who found the lead, while refuting one of its mistakes. The model then identified the product and looked up its entry in the Czech drug database. My doctor decided and prescribed it. That is how it should be.

Where the model got it wrong

So this doesn't come across as an ad: over the four days of preparation, the model slipped up twice.

Once it recommended a "stronger" product. Before ordering, I sent it to the model. The ingredients showed it was the same cream I already had at home, just in a different box.

The second time it presented a guess as a verified fact: that no stronger cream was available here. It admitted this only when I came back with counterevidence. That very counterevidence turned out to be the product my doctor later prescribed.

In the lab we call this the grep rule: a quote that hasn't been verified doesn't exist. With a model giving health advice, it applies doubly. So I check every product and every claim about medication against the registry, the package insert and my doctor. Here too the model made mistakes with confidence, and checking is my job.

D-Day

On the morning of July 24 I lay on the exam table with numbed skin. The nurse was kind, and I didn't even notice the moment it happened. I felt nothing. I burst into tears of relief. Twenty years of history against four days of protocol, and the protocol won.

How much of that was the prescribed cream and how much the rest of the protocol, I can't tell from a single blood draw. In the moment I forgot some of what I had prepared anyway.

It wasn't calm. But I believe that in a year it will be. For the first time, my brain has a sequence stored that ends in relief, not horror.

What I take from it

Where the model helped me most was in what doctors don't have time for: explaining the mechanism, building a protocol tailored to me and showing me data from my own life. Ten minutes of patient "why the body does this" changed more than years of "it'll be fine."

If anything here helped me more than anyone, it was the method, not the model. Data instead of impressions, a protocol instead of willpower. And people exactly where they belong: the doctor and the nurse at the decisions and at the needle.

I went for a follow-up with my doctor, and she said everything is fine.