Client Work // 04 · Healthcare AI · Prompt Engineering · LLM Architecture
When Seconds Matter, Doctors Shouldn't Have to Hunt for Information
EHR systems were built for documentation. Not for thinking under time pressure. This is a prompt architecture and interface concept built around one constraint: what should AI never do in a clinical setting?
Role
Lead UX Designer · Prompt Engineer
Type
Prototype · Ongoing
Stack
LLM · Prompt Architecture · React
Status
Ongoing. No hallucinations observed in tested scenarios.
The Problem
"A physician has 60–90 seconds to scan a patient's history before entering the consultation room."
A physician walks into a consultation with about ninety seconds before the patient looks up. In that window they need to know what matters. Not everything. What matters. The electronic health record gives them everything. Critical alerts sit next to routine observations. Missing information looks identical to present information. The system does not tell you what it does not know. The cognitive cost of sorting signal from noise falls entirely on the doctor, every time, before the conversation has even started.
"14 of 21 major studies reported an association between clinician burnout and clinically significant medical errors. The AMA frames this as a system failure. Not a lack of physician resilience."
Source: PubMed burnout–error meta-analysis; Armstrong Institute (Johns Hopkins)
My Role
Lead UX Designer · Prompt Engineer
I started from the wrong end of the problem. Not what can the AI do, but what should it never do. That question shaped every decision. The AI should not interpret. It should not diagnose. It should not present uncertain data as though it were certain. Its job is to transform a fragmented pile of records into a card a physician can read in under ten seconds. Surface what is critical. Flag what is missing. Stop there. That constraint is not a limitation. It is the design.
- Problem framing and research synthesis
- Prompt architecture and constraint design
- Output schema and card UI design
- Hallucination testing and edge-case validation
The Solution
The Patient Summary Card
I designed a prompt engineering architecture that uses an LLM as a clinical structuring layer. The system is designed to extract clinically relevant details, surface critical flags first, and output a clean structured card. Built to minimise hallucination and avoid clinical interpretation. Clinical judgment stays with the physician.
Margaret H., 54
Female · GP visit · Dr. A. Osei
Alerts
HbA1c overdue, 14 months
Guideline: retest every 6 months for Type 2 DM
Cardiology referral unresolved
Referred 8 months ago, no outcome recorded
Visit Reason
Fatigue and shortness of breath, onset 6 weeks ago, gradual
Active Conditions
Current Medications
Patient Voice
"The tiredness is affecting my work. I'm worried the breathing could be something serious."
Data Quality
DEFAULT STATE · ALL DATA PRESENT
David K., 67
Male · GP visit · Dr. A. Osei
Alerts
Allergy record incomplete
Sulfa allergy logged, no reaction severity documented
Visit Reason
Routine review, knee pain management
Active Conditions
Current Medications
Patient Voice
No patient-stated concern found in record
Data Quality
INCOMPLETE DATA STATE · MISSING FIELDS SURFACED
Yusra M., 71
Female · GP visit · Dr. A. Osei
Alerts. Review Before Entering.
Warfarin + new Ibuprofen script, interaction risk
Prescribed by out-of-hours GP 3 days ago, not reviewed by this practice
INR result outstanding, 19 days
Last INR: 3.8 (above therapeutic range). Retest overdue.
Falls risk, 2 incidents in past 4 months
Formal falls assessment not completed
Renal function, last checked 11 months ago
Relevant given Warfarin dosing and age
Visit Reason
Dizziness and unsteadiness on feet, onset 2 weeks ago
Active Conditions
Patient Voice
"I feel unsteady when I stand up. I am frightened of falling again."
ALERT-HEAVY STATE · CRITICAL FLAGS PRIORITISED
Before vs After
Before
- Raw EHR data — unformatted, inconsistent, dense
- Physician scans multiple screens to build context
- Critical flags buried inside narrative text
- Missing fields silently absent, not flagged
- Working memory overloaded before patient enters the room
After
- Structured clinical card — consistent, scannable output
- Single view designed to surface clinically relevant context
- Alerts ordered by severity, visible immediately
- Missing fields explicitly flagged in tested scenarios
- Physician enters the room with context already formed
The Architecture
Prompt Engineering as UX
The logic layer does what a good UX system always does: it reduces unnecessary decisions for the user. Five rules guided the prototype output.
Zero filler
No introductory text, no hedging language. Output was constrained to begin with the first JSON key.
Alert-first ordering
Overdose risks, unresolved allergies, and missing critical data surface before any summary content. A doctor must see risks before context.
Confidence calibration
The prompt was designed to avoid diagnosis, recommendation, or interpretation. It structures and surfaces. Clinical judgment remains with the physician.
Human layer
One patient-stated concern extracted verbatim from notes. If none exists, marked as missing. Not inferred in tested scenarios.
Confidence ceiling
The prompt was designed not to diagnose or interpret. It surfaces, and flags what it doesn't know.
FIG A. CLINICAL LOGIC ENGINE. Prompt architecture across input, logic, data, and component layers.
Research Grounding
The Evidence Base
Working memory ceiling
~7 items
A typical pre-consultation EHR data load far exceeds working memory capacity, forcing physicians into triage mode before the patient enters the room.
CLT in medicine (AMEE Guide No. 86)
AI vs full record review
Reported review-time reduction
One study found AI-generated summaries could reduce review time while maintaining clinical accuracy.
Comparing AI- vs. Clinician-Authored Summaries (medRxiv, 2025)
Diagnostic safety
Burnout and diagnostic-safety research
Burnout is associated with production pressure and path-of-least-resistance decisions. The AMA frames this as a system failure. Fix the interface, not the person.
PubMed meta-analysis; Armstrong Institute; AMA research
Quality of interaction, not duration
Substantial satisfaction increase
Patients who feel "seen" have higher enablement scores. The human layer in the card gives physicians the raw material to create that moment, even in 15 minutes.
Enablement After Consultation in Primary Healthcare (Dove Press, 2025)
Next
"The most important decision in a safety-critical AI system is often what you refuse to let it do."