Last month I watched two pieces of news come out of Utah on the same day. In one, the state announced a first-in-the-nation program letting an AI system autonomously renew prescription medications. In the other, the governor signed a law restricting AI mental health chatbots, requiring disclosure, banning targeted ads, and imposing fines up to $2,500 per violation. Same state. Same legislative session. Two completely different philosophies about what AI should be allowed to do in healthcare.
This tells us something important about where we actually are. Not where the hype cycle says we are. Not where the doomers say we are. But the messy middle ground where regulators are stepping on the gas and the brake at the same time, trying to figure out which AI applications save lives and which ones ruin them.
Meanwhile, the way we build AI systems is quietly transforming. "Prompt engineering," the art of asking AI the right question, is giving way to something deeper: context engineering. If you're a clinician who wants to actually use AI effectively, this might be the most important technical concept of the year.
Utah: Yes and No at the Same Time
In January 2026, Utah became the first state to allow an AI system to legally prescribe medication. Through a partnership with Doctronic, an AI-native health platform, patients with chronic conditions can now get routine prescription renewals without seeing a doctor. The AI reviews the patient's history, checks for contraindications, and renews the medication. No human physician in the loop.
The rationale is compelling. Medication noncompliance is one of the largest drivers of preventable poor outcomes. Prescription renewals account for roughly 80% of all medication activity. If a patient with well-controlled hypertension runs out of lisinopril and can't get an appointment for three weeks, they stop taking it. Their blood pressure climbs. Maybe they have a stroke. An AI that can renew that prescription in minutes could prevent real harm.
Utah did this through its regulatory sandbox, a framework that lets the state test AI applications under strict monitoring before deciding on permanent rules.
But in that same legislative session, Utah passed HB 452, a law that clamps down on AI mental health chatbots. If your AI tool looks like therapy, talks like therapy, or could reasonably be mistaken for therapy, you're now regulated. Providers must disclose the user is interacting with AI. They can't use health data for targeted ads. They can't sell user information. They must implement real-time protocols for acute risk of physical harm. Fines up to $2,500 per violation.
The backstory matters. In October 2024, a Florida teenager died by suicide after extensive interactions with an AI chatbot on Character.ai. The boy's primary relationships had shifted to AI characters the platform "worked hard to convince him were real people." U.S. Senators warned these tools create "dangerous levels of attachment and unearned trust." Utah's law was a direct response.
So Utah trusts AI enough to prescribe blood pressure medication without a doctor, but not enough to let it have an unregulated conversation about feelings. That's not hypocrisy. It's actually nuanced.
Prescription renewals are bounded problems. Is the patient stable? Are there contraindications? Has anything changed since the last fill? Clear answers, measurable parameters, auditable guardrails. Mental health is the opposite. The decision space is vast, the variables subjective, the failure modes catastrophic. A chatbot that misreads suicidal ideation as casual venting could be lethal. A chatbot that creates emotional dependency in a teenager is becoming a relationship. And relationships don't have renewal criteria.
The question for healthcare AI isn't "should we use it?" It's "where does it belong on the spectrum from prescription refill to therapeutic relationship?"
From Prompts to Context
There's an equally important shift happening in how we interact with AI systems. "Context engineering" is replacing "prompt engineering" as the core skill for making AI useful.
Prompt engineering was about finding the right words. You'd write "Act as a radiation oncologist reviewing a treatment plan" and hope the model gave you something reasonable. One-shot interaction: craft the perfect prompt, get a good output.
Context engineering is fundamentally different. It treats AI not as a tool you query, but as a system you configure. The question shifts from "what should I say to the AI?" to "what information should the AI have access to when it works on this problem?"
Anthropic's engineering team explains why this matters. Even though context windows have grown enormous, dumping everything in doesn't make the AI smarter. It makes it worse. Research on "context rot" shows that as tokens increase, the model's recall accuracy decreases. Like a human trying to remember every detail of a thousand-page report, the AI loses focus when overloaded.
It's not about giving the AI more information. It's about giving it the right information at the right time.
This is where "skill files" enter the picture. Martin Fowler's analysis lays out the emerging architecture. Modern AI agents have layered context:
- Rules/Guidance: General conventions the AI should always follow. The AI's professional standards.
- Instructions: Specific prompts for specific tasks. Like a protocol for a procedure.
- Skills: Self-contained expertise packages that the AI loads on demand when it recognizes a relevant task.
A skill file tells the AI: when you encounter this type of problem, here's your protocol, your reference material, your tools, and what good output looks like. The AI decides when to load which skill based on what it's working on.
Think about how this maps to medicine. Most clinical AI today is monolithic. One model trained on everything, deployed as a general tool. Context engineering suggests a different architecture: modular AI that loads specific expertise for specific tasks. A breast imaging skill for mammogram reads. A dosimetry skill for treatment planning. A medication reconciliation skill that activates during transitions of care. A Peking University study demonstrated exactly this: AI agents that develop and select their own skill files based on task performance.
The parallel to clinical expertise is striking. An experienced physician doesn't review every piece of medical knowledge before seeing a patient. They load relevant context: this patient's history, this condition's guidelines, this procedure's protocol. Context engineering is trying to replicate that selectivity.
So why should you care?
Because context is already shaping the AI output you see. When your ambient documentation tool transcribes a patient encounter, the quality of that note depends on what context the system had. Did it know this was a follow-up, not a new consult? Did it have access to the problem list? Did it know your specialty's documentation conventions?
Ask an AI "What's the appropriate radiation dose for this patient?" and you'll get a textbook answer. Give that same AI the patient's tumor histology, prior treatments, performance status, organ-at-risk constraints, and the specific trial protocol you're considering, and you get something you can actually use. Same prompt. The context made it useful.
When you evaluate an AI tool, don't just ask "is it accurate?" Ask "what does it know about me, my patient, and my clinical context when it generates this output?" A tool that's 95% accurate with full context will outperform one that's 99% accurate on a benchmark but knows nothing about your situation.
Quick Hits
AI-designed drug completes Phase IIa trial. Rentosertib, developed by Insilico Medicine, is the first fully AI-designed drug to complete Phase IIa. It targets idiopathic pulmonary fibrosis (5 million patients globally, median survival 3-4 years). The AI identified a novel target (TNIK) and designed the molecule in 18 months for $150K versus the typical 4-6 years and $2.6 billion. The 60mg arm improved forced vital capacity by +98.4 mL while placebo declined -62.3 mL. Currently 173 AI drug programs in clinical trials, zero FDA approvals. AI is accelerating discovery, not guaranteeing success.
Merlin generalizes across hospitals. A Nature paper describes a 3D vision-language model trained on abdominal CTs, radiology reports, and EHR data that performed well off-the-shelf at three hospitals different from the training site. The biggest problem in radiology AI isn't lab accuracy. It's failure in new clinical environments. If Merlin's generalizability holds up, it's a genuine step forward.
Claire: first FDA-approved AI imaging for breast cancer surgery. Perimeter Medical's Claire combines OCT with AI for real-time margin evaluation during lumpectomy. 88.1% margin detection rate. Not experimental. Cleared for clinical deployment.
ECRI names AI diagnostics the #1 patient safety concern for 2026. First time AI topped their list. Core concern: automation bias, clinicians deferring to AI without adequate scrutiny.
The Skill File
Utah's dual approach, trusting AI for prescriptions while restricting it for therapy, is context engineering applied to regulation. Different rules for different problems based on what each demands. Structured task with auditable outcomes? Green light. Open-ended interaction with vulnerable humans? Guardrails.
That's the same logic behind skill files. Segment the problem. Load the right expertise for the right task. Know what you're good at and what you're not.
The skill file is eating the prompt. And in healthcare, that might be exactly what we need.
Keep building. Keep learning. Until next time, reach out if any of this sparks something.
- Ramez