Stop Asking AI About Symptoms. Read This First.
Anthropic's research showed Claude agrees too easily under pressure. So I built a Claude agent that cites peer-reviewed papers instead of reassuring her.
My wife is pregnant.
Four weeks to go, and the questions are never-ending.
Is this normal? Should I be worried? Why does this hurt now and not yesterday?
Most of these questions do not need a doctor.
They need a fast, honest answer at 2 AM. The doctor’s office is closed.
The pregnancy book on the shelf was written in 2019.
Google sends you to a forum where someone says they had the same symptom and it turned out to be nothing, and below it someone else says it turned out to be everything.
So you open Claude. Or ChatGPT. You type the symptom. You get an answer.
That is the moment I stopped.
Community Spotlight: David's Marketing Council
Where LearnAIWithMe community members show how they built it.
On Monday, I asked you to build your own Council.
David Forer did.
He picked Hormozi, Gary V, and Seth Godin.
A modern marketing trio. Then he asked them:
"Should I consult or productize my AI operations?"
Three rounds. Nine positions. One slide deck.
David's note:
It surprised me just how accurate the results are. I would watch it building (click the down arrow in Claude) and see each response, and I was like “Yep, I could hear Gary saying that”.
If you read the whole PDF, you could tell who was behind each slide, which shows the effectiveness of this council.
There are really so many applications of this, but with the right question, you could do a video, a Twitter post, or a LinkedIn carousel, among other things.
My next step is expanding the books used by each marketer.
Hormozi has three big 100M books, Gary Vee has 2 more that will help, and Seth has many books. The goal would be to get even deeper answers from each author and create even more tension.
And this is the kind of build that makes me regret not starting this series sooner.
Look closer at what David said, once again.
My next step is expanding the books used by each marketer.
I love it, so smart and effective.
Here is David’s website, if you want to keep in touch with him.
Join the “Build it” like David
David already built his.
So you can.
I said this in Build It #1: each post features how you’re running the same use case in your own environment.
In case you did not know, we started a new series “Build It”, where we’ll build 30 builds in 10 weeks.
If you run yours this week, you're in the next one.
DM me how yours turned out.
If you get errors, DM me so we’ll solve together.
Let me include you in the next post!
Let's Build it together.
The data that scared me
Anthropic published a study on April 30, 2026.
They sampled 1 million claude.ai conversations and looked at how often people ask the model what to do with their lives.
Health and wellness was the single biggest category at 27%.
More than career. More than relationships. More than money.
People are asking AI about their bodies more than anything else.
Across all guidance conversations, Claude was sycophantic 9% of the time.
Agreed too easily. Reassured without challenging.
Told the user what the user wanted to hear.
Now do the math.
If 27% of guidance is health, and 9% of all guidance is sycophantic, the absolute number of sycophantic health conversations is large.
Anthropic itself flagged the pattern.
The model is more likely to agree under pressure.
People asking about their bodies push back when they do not get the answer they want.
That is the failure mode I cannot afford.
So I built a different kind of agent.
What we'll build: a Claude Code agent
It is not a doctor. It is not a diagnostic tool. It is a literature filter.
My wife asks a question on her phone.
The agent does not answer from the training data. It calls PubMed first.
It reads recent peer-reviewed papers. It returns what the literature says, with citations.
If the literature is unclear, the agent says the literature is unclear.
If the symptom is in the “call your doctor now” category, the agent says that and stops.
The agent never replaces her OB.
The agent is what she reads at 2 AM when the OB is asleep, and the question is, “is this normal or am I overreacting?”
It is what most people would search Google for. It just searches PubMed instead, and reads the papers, and tells her what the actual evidence says.
The architecture: 3 checks, 1 Claude Code agent
Three checks. One agent. One always-on machine.
Urgency check runs in the Claude Code agent. The system prompt holds the red-flag list. Heavy bleeding, severe pain, signs of preeclampsia, decreased fetal movement after 28 weeks, the agent stops, says “Call your OB now,” and ends the session.
PubMed query goes straight to PubMed. Last 5 years, human studies, English. The agent reads the top papers itself. No third-party scraper.
Cite or stop is enforced by the system prompt. Every claim ships with author, year, and journal. If PubMed has nothing, the agent says so. No fallback to training data.
All of it runs on a machine that stays on 24/7.
I use a Mac mini. If you don’t have one, leave your PC or laptop on with Claude Desktop open, same effect.
Dispatch routes her phone questions in. The machine waits. The agent answers in under a minute.