rchitecture** that enforces source verification, contextual precision, and diagnostic boundary limits.

1. Context Injection Framework

Provide structured physiological and symptomatic parameters to reduce ambiguity. Frame inputs as research requests rather than diagnostic queries.

[ROLE] Medical Research Assistant
[CONTEXT] User: 34M, sedentary developer, 6 months persistent lower back pain, no trauma history.
[QUERY] List differential diagnoses commonly associated with these symptoms.
[CONSTRAINT] Do not provide treatment plans or dosages. Cite only peer-reviewed sources or major clinical guidelines (e.g., WHO, CDC, Mayo Clinic).

2. Constraint & Fact-Check Directives

Inject explicit boundary conditions to prevent hallucination and scope creep. Use system-level directives to enforce source limitation and uncertainty acknowledgment.

[DIRECTIVES]
- If data is insufficient or outside training cutoff, explicitly state uncertainty.
- Reject medical myth propagation. Cross-reference claims against established clinical consensus.
- Format output as: Symptom Mapping -> Potential Pathologies -> Recommended Clinical Verification Steps.
- Never replace professional medical evaluation.

3. Decision Matrix Implementation

Integrate a risk-based routing logic before query execution:

• Critical/Emergency (Chest pain, dyspnea, stroke signs, severe trauma) → Bypass AI. Immediate clinical escalation.
• Personalized Treatment (Dosage, surgical decisions, prescription adjustments) → Bypass AI. Direct specialist consultation required.
• Non-Critical Discomfort (Chronic back pain, eye strain, mild GI issues) → Use AI for preliminary differential mapping, followed by scheduled clinical visit.
• Informational/Literacy (Terminology, lab value interpretation, biological mechanisms) → Optimal AI use case. Apply structured prompts and cross-verify.

4. Medical Translation Application

Repurpose LLMs as a Clinical Terminology Decoder. Anonymize PHI (Protected Health Information) and request plain-language explanations of lab panels (LDL, HDL, Triglycerides, AST/ALT) to improve patient-clinician communication efficiency.

Pitfall Guide

1. The Direct Diagnosis Trap: Asking "What disease do I have?" forces the model into unfounded probabilistic guessing. LLMs lack clinical examination capabilities and will generate plausible but unverified conditions.
2. Hallucination Blindness: Trusting AI-generated citations, drug names, or treatment protocols without cross-referencing peer-reviewed literature or clinical databases. Always verify source links and pharmacological claims.
3. Ignoring Data Cutoff & Bias: Assuming the model reflects current clinical guidelines. Training data may contain outdated protocols or internet-sourced medical myths, leading to stale or biased recommendations.
4. Privacy Leakage (PHI Exposure): Inputting identifiable health records, exact lab values with names, or sensitive personal history into public LLM endpoints. Always anonymize data before processing.
5. Skipping the Verification Loop: Treating AI output as final medical advice. AI should only serve as a preliminary research aid; all clinical conclusions require human professional validation.
6. Confident Tone Fallacy: Misinterpreting the model's fluent, authoritative phrasing as clinical certainty. LLMs optimize for linguistic confidence, not diagnostic accuracy. Treat outputs as hypotheses, not facts.

Deliverables

• 📘 Health-Query Prompt Blueprint: A structured architecture template for safe medical LLM interactions, including context framing, constraint injection, and output parsing rules.
• ✅ Pre-Query Safety Checklist: A 6-step validation workflow covering risk categorization, PHI anonymization, source constraint enforcement, and verification routing before executing any health-related prompt.
• ⚙️ Prompt Configuration Templates: Ready-to-use JSON/Markdown prompt schemas for three operational modes: Symptom_Mapping, Lab_Value_Decoder, and Medical_Terminology_Research, pre-loaded with safety directives and citation constraints.