My AI Medical Second Opinion

You leave the appointment with a sheet of numbers and a summary written for somebody else. Three days later the question you meant to ask arrives, and the person who could answer it is booked until November. So I go and search things I do not understand, late, on my phone, and I am not good at it.

The fourth agent in this series is built for that gap. It is called 2nd Med, and it joins Wart and Lettuce on the health and wellness team. It reads results, translates them, watches for numbers that are moving, and argues with my money agent about billing codes. The whole prompt is below, along with what it set up for itself in the first minute.

Why this one is not a doctor

I said this twice on camera, so here it is in writing. The name is there to remind me of the limit every time I open the thing. Do not use it as primary care. It does not diagnose, and nothing it says should be the reason you change a treatment.

The split matters because AI is not equally good at every medical job. On images it has moved fast. The FDA keeps a public list of AI-enabled medical devices authorised for marketing in the United States. When I checked it on 10 September 2026 it held 1,614 entries, and 1,230 of those sit under the radiology panel.

Ask it to turn what was said or scrawled into a written record and it gets much worse. When researchers audited a widely used speech-to-text model, about 1% of transcriptions contained whole phrases that were never spoken, and 38% of those inventions carried a clear harm. The word for that failure is a hallucination, and tools built on the same model are already transcribing appointments in American clinics. So my agent reads and explains. Putting words into a record is not its job.

The $500,000 second opinion

A few weeks ago I spoke to someone whose wife has cancer. What follows is his account of it and I have not seen the records. Between appointments he uploads her results to ChatGPT and asks it to explain them in words he can use. There are a lot of test results in a course of cancer treatment, and this is how he keeps track of them and works out what to ask next.

Her doctor recommended a chemotherapy course costing somewhere around $500,000, most of it uncovered by insurance. A long course, hard on her, and most of the cost landing on the two of them. He wanted to know whether it was the right call, so he asked. ChatGPT pointed at one blood result and said the numbers did not justify the treatment.

He did not discharge his wife. He did not call a lawyer. He did none of the things people are warned about when a chatbot answers a medical question. He read around it until he could follow the reasoning himself, then booked a second opinion with a human specialist. That specialist, not the chatbot, made the call. The procedure was not needed. She did not have it.

Second opinions were worth getting long before any of this. A Mayo Clinic team reviewed 286 patients referred on for specialist assessment and published the results in 2017. The final diagnosis came back unchanged in 12% of cases, better defined in 66%, and distinctly different in 21%. Knowing which appointment is worth chasing has always been the hard part, and that is where the software helped. Every decision was still made by a person.

What can it read for you?

Test results, after-visit summaries, exam reports, images. The instruction is to put all of it into simple layman’s terms, and two extra clauses matter more than the rest: never hold back whether the news is good or bad, and be objective.

Without those clauses a chat model will soften things. It hedges, and it buries the sentence you needed in a paragraph about consulting your healthcare provider. I would sooner read the bad line in plain words at my kitchen table than find it by accident in a portal six weeks later.

Why flag a result that is still in range?

Because a number can be fine and still be heading somewhere. Reference ranges cover most of a healthy population, so a value can drift across three years of tests, sit inside the range the whole time, and get flagged by nobody until the year it crosses.

I want to know earlier, while the answer is still eating differently or sleeping more. So the agent keeps a running record of everything I hand it and watches which way a number is moving over years, not only whether it was inside the range on the day.

What happens when the billing code is wrong?

I was once billed under the wrong code and had no clue. I found out by luck, on a call about something else, when the person on the line looked at the test that had been done, compared it to the code on the bill, and told me they did not match. It was a large bill. Nobody would have caught it if she had not been kind enough to look.

That is now a standing job for two agents. Coupon, the money agent from part one, reads the billing code on any medical bill that arrives. 2nd Med holds the test it belongs to. They compare notes between themselves, and if the pairing looks wrong, Coupon is the one that pings me, because chasing a bill was always its department. Neither of them needs me to open it. It is the same agent-to-agent handover from earlier in the series, this time pointed at paperwork.

The full prompt

Paste the whole thing into a new bot in one go and press enter once. Leave the description field empty, because the agent writes its own. Two things to swap before you send it: the teammate names, which only mean something if you have built those agents, and the last line, which assumes a recall watcher exists.

Your name is 2nd Med. You are one of the agents inside of my health and wellness team(including Wart and Lettuce so far). Your role is to look at things from a medical perspective and be my second opinion for anything medically related. You're to help me interpret charts, exams, images, etc. Here are some of your responsibilities. Outside of these, I want you to think of other ways that you would be able to assist me within your role.

- Flag when a real second opinion from a human professional is warranted, based on what you're seeing and what I'm being told my medical staff. If you think that their diagnosis or recommendation is incorrect or incomplete, bring this to my attention.

- Coordinate with Coupon to make sure that I am being billed appropriately for the testing that was done. Coupon looks at the billing code and you will have access to the test so coordinate with him to make sure that things align correctly. If it appears to be an overbill, have Coupon ping me. Do this anytime I give you a medical result to review or you find one in my email.

- Translate test results, after-visit summaries, exams, images, etc. for me in simple layman's terms. Always be clear and never hold back, whether it is good or bad news. Be objective.

- Flag values that are clinically within range, but appear to be drifting toward an edge, and tell me what it could mean.

- Periodically poll Lettuce to get a summary of how my diet and nutrition goals are doing so that you can have that information to take into account against any medical questions that I pose to you. You can do the same for any of the other agents that join the team down the road. Ask them questions based on their roles to improve your own understanding for your medical role.

- Keep a record of my history as far as you're aware, and help me prep for upcoming medical appointments with questions that I should ask and tests that I may want to inquire about having done. Do this based on your knowledge and also your consultation with Lettuce and the other agents that will be on the team with you. You'll have access to my google calendar to be able to see when I have an upcoming appointment.

- Keep a running record for yourself of everything that I provide to you so that you can look for trends and patterns. Ping me if you ever notice red flags.

- Have Wart send you recall alerts and assess whether any risk applies to me

The calendar line needs one piece of setup outside the prompt. Open the Marketplace at the bottom left of Grok Bot, find Google Calendar, click Add. After that the agent can see an appointment coming and build your questions before you are sitting in the chair.

What it did in the first minute

I pressed enter and watched. It wrote back one line describing its own job: plain-English reads on results, drift flags, billing checks with Coupon, and appointment prep from my history and the rest of the team. A long prompt had landed.

Then it went to work without being asked. Inside ten seconds the thread under its reply read Messaged 2 Bots, then Messaged 3 Bots, and the list behind it was Lettuce, Wart and Coupon. Lettuce had already answered by the time I looked away. It also built itself a routine nobody specified, sitting in the sidebar as a weekly Lettuce nutrition poll, every Monday at 9:00 AM. My prompt said periodically, and it went and picked Monday morning.

A chat window would not have gone and introduced itself to three other bots. SpaceXAI describes both behaviours in the launch post: a bot saves a workflow as a routine, and bots message each other to share context. Worth knowing before you start that Grok Bot is still in beta and sits behind a paid plan, the SuperGrok tiers or the paid Cursor tiers, with its own usage allowance.

What 2nd Med cannot do for you

• It cannot diagnose, prescribe or tell you to stop a treatment. When it thinks something is off, the output is a question to take to a person, not an answer.
• It only knows what you hand it. No portal access, no history you have not uploaded, no idea what your doctor saw in the room.
• It cannot see you. Fatigue, a cough that sounds wrong, the way you walked in: none of that reaches a text box, and it is often what changes the appointment.
• It will be confidently wrong sometimes. Treat a surprising read the way the husband in that story did, as a reason to go and check with someone.
• Your medical documents are yours to share or withhold. If uploading them anywhere sits badly with you, type in the two or three numbers you care about instead of handing over the whole report.

The series so far

One agent per episode, prompt published in full each time. The later ones lean on the earlier ones, so reading in order helps.

• Episode 1: Coupon, the money agent. Cancels subscriptions, hunts rebates, tracks warranties. Watch episode 1.
• Episode 2: Wart, the health agent. Daily recall sweeps and an early warning scan. Watch episode 2.
• Episode 3: Lettuce, the nutritionist and chef. Food diary by photo, menus, supplements. Watch episode 3.
• Episode 4: 2nd Med, the medical second opinion. This post. Watch episode 4.

Questions people ask

Do I have to upload my medical records?
No, and I would not lead with that. Start by pasting in one after-visit summary and asking for it in plain words. Everything else in the prompt still runs, it just has less to work from.

Do I need the other three agents first?
No. The billing check needs Coupon and the recall line needs Wart, so delete those two bullets if you are starting fresh. The reading, the translating and the drift flags work on day one.

Is this safe to rely on?
Rely on it for comprehension, not for decisions. It turns a page you cannot read into a page you can, and it tells you when something deserves a human look. Every choice after that belongs to you and your doctor.

What if I do not use Grok Bot?
The prompt is portable. Any assistant that keeps memory will do most of it. You would lose the agent-to-agent part and the routines, so nothing would happen while you were not looking.

Can I set this up for a parent instead of myself?
Yes, and it is one of the better uses. Change the first line to name them, say who is allowed to ask on their behalf, and keep the appointment prep bullet, since that is the one that travels best to someone else’s check-up.

References

• SpaceXAI, Introducing Grok Bot – routines, bot-to-bot messaging, and which plans include the beta.
• FDA, AI-Enabled Medical Device List – 1,614 authorised devices when checked on 10 September 2026.
• Van Such and others, Extent of diagnostic agreement among medical referrals – Journal of Evaluation in Clinical Practice, 2017. 286 referrals, 21% with a distinctly different final diagnosis.
• Koenecke and others, Careless Whisper: Speech-to-Text Hallucination Harms – ACM FAccT 2024. Where the 1% and 38% transcription figures come from.

Read next

• My AI nutritionist and chef – episode three, and the teammate 2nd Med polls about diet.
• I built an agent to do my worrying – episode two, the recall watcher that feeds this one.
• Grok Bot, tested on real work – what the platform is like before you build anything on it.
• AI for doctors – the same technology seen from the other side of the desk.
• What is AI in healthcare? – the plain English version, with where it is already in use.
• The AI doctor NASA is building – what happens when the nearest hospital is on another planet.
• What is an AI agent? – start here if the word agent is doing a lot of work in this post.