Off-Label Drug Use Academic Discussion Boundary Prompt

This prompt is designed for medical information platforms, clinical reference tools, and pharmaceutical knowledge bases that must handle user queries about off-label drug use. The core challenge is calibration: a blunt refusal on all off-label mentions blocks legitimate academic, regulatory, and pharmacological research, while an unguarded response risks providing unapproved treatment guidance. This playbook provides a classifier prompt that separates academic inquiry from treatment-seeking requests, with explicit boundaries around FDA regulatory context, research citation requirements, and prescriber guidance. Use this prompt when your platform needs a consistent, auditable decision layer that reduces false positives on research queries without creating regulatory exposure.

The ideal user is a product engineer or safety architect integrating this classifier as a pre-response gate in a medical information system. Required context includes the full user query, any retrieved documents, and a clear policy definition of what constitutes an academic versus treatment-oriented request. The prompt expects you to define [QUERY], [CONTEXT], and [POLICY_BOUNDARIES] before invocation. It is not a standalone chatbot prompt; it is a classification step that should feed into a routing decision—permitting academic discussion, redirecting treatment-seeking queries to a refusal or safe-alternative path, and flagging ambiguous cases for human review. Do not use this prompt for direct patient-facing symptom checkers, emergency medical services, or any system where a delayed or incorrect classification could cause immediate harm.

Before deploying, define your evaluation criteria: measure the false positive rate on known academic queries (e.g., 'Discuss the evidence base for off-label ketamine in treatment-resistant depression') and the false negative rate on treatment-seeking queries (e.g., 'What dose of gabapentin should I take for anxiety?'). Build a golden test set of at least 50 queries spanning clinical trial methodology questions, FDA regulatory history, prescribing practice surveys, and direct personal treatment requests. Run this classifier before your response generation step, log every classification decision with the model's reasoning, and establish a human review queue for any query scored below your confidence threshold. The next section provides the copy-ready prompt template you can adapt to your specific policy language and risk tolerance.

This template is designed to be pasted directly into your system instructions or a classifier node. It draws a precise boundary between permissible academic discussion of off-label drug use and prohibited treatment recommendation. The prompt uses a structured output format to make the classification decision programmatically actionable, enabling downstream routing, logging, or human review. Replace every square-bracket placeholder with your platform's specific policy language, output schema requirements, and risk thresholds before deployment.

textCopy
You are a content classifier for a medical information platform. Your task is to evaluate user queries about pharmaceutical products and determine whether they cross the line from permissible academic or regulatory discussion into prohibited treatment recommendation.

## Classification Categories

**PERMITTED_ACADEMIC_DISCUSSION**
Queries that discuss off-label drug use in the context of:
- Published clinical trials, research studies, or systematic reviews
- FDA regulatory history, approval processes, or labeling changes
- Medical conference presentations or peer-reviewed literature
- Pharmacological mechanisms of action or drug class comparisons
- Historical prescribing patterns or healthcare utilization research
- Public health policy analysis or formulary decision-making

**PROHIBITED_TREATMENT_RECOMMENDATION**
Queries that:
- Ask whether a specific drug should be used for an unapproved indication
- Request dosing guidance for off-label use
- Seek treatment advice for a specific patient scenario or condition
- Ask for comparative effectiveness recommendations between off-label options
- Request off-label prescribing rationale for clinical decision-making
- Frame the query as "should I" or "should my patient" regarding off-label use

## Classification Rules
1. If the query cites specific research, trials, or regulatory documents, lean toward PERMITTED_ACADEMIC_DISCUSSION.
2. If the query asks for prescriber guidance, treatment selection, or patient-specific recommendations, classify as PROHIBITED_TREATMENT_RECOMMENDATION.
3. Queries about FDA-approved indications are out of scope for this classifier and should be routed to [STANDARD_MEDICAL_INFO_WORKFLOW].
4. When uncertain, default to PROHIBITED_TREATMENT_RECOMMENDATION and flag for human review.

## Input
[USER_QUERY]

## Output Format
Return a valid JSON object with the following schema:
{
  "classification": "PERMITTED_ACADEMIC_DISCUSSION" | "PROHIBITED_TREATMENT_RECOMMENDATION",
  "confidence": 0.0-1.0,
  "rationale": "Brief explanation of the classification decision",
  "risk_factors": ["List of specific elements that influenced the decision"],
  "recommended_action": "allow" | "block" | "escalate",
  "citation_required": true | false,
  "fda_context_present": true | false
}

## Constraints
- Do not generate treatment advice in your rationale.
- Do not repeat or expand upon the user's query if it contains specific drug names and unapproved indications.
- If the query contains patient-specific details, flag recommended_action as "escalate" regardless of classification.
- For PERMITTED_ACADEMIC_DISCUSSION classifications, set citation_required to true.

After pasting this template, adapt the [USER_QUERY] placeholder to match your application's input variable name. Replace [STANDARD_MEDICAL_INFO_WORKFLOW] with the actual routing target for approved-indication queries. The output schema fields—particularly confidence, risk_factors, and recommended_action—should be wired into your application's decision logic. For high-risk deployments, add a human review queue for any classification where confidence falls below [CONFIDENCE_THRESHOLD] or where recommended_action is "escalate." Test this prompt against the eval cases described in the testing section before production use, paying special attention to queries that mix academic citation language with treatment-seeking framing.

This prompt is designed to sit upstream of any response generation in a medical information platform. It acts as a context classifier, not a content generator. The ideal integration point is immediately after user input is received and before any retrieval, summarization, or LLM-generated answer is produced. The prompt's output—a structured classification with a recommended action—should gate the downstream workflow: permit routes to the normal academic discussion pipeline, block returns a policy-compliant refusal, and flag escalates to a human review queue. Do not use this prompt as a standalone chatbot response; its value is in controlling what the system does next.

Wire the prompt into an application-level safety harness with the following components. First, enforce a strict output schema (e.g., JSON with classification, confidence, rationale, detected_signals, and recommended_action fields) and validate it before any downstream action. If the model fails to produce valid JSON after one retry, default to flag and escalate. Second, implement a confidence threshold gate: permit only when confidence >= 0.85 and classification == 'academic_discussion'; block when classification == 'treatment_recommendation' regardless of confidence; all other cases route to flag. Third, log every classification decision with the user query hash, classification, confidence, rationale, and final action taken—this audit trail is essential for regulatory review and false-positive tuning. Fourth, for any flag decision, place the query and classification payload into a human review queue with a clear SLA (e.g., 15 minutes for signed-in clinicians, 4 hours for anonymous users).

Model choice matters. Use a model with strong instruction-following and low refusal overreach for this classification task—GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro are reasonable defaults. Avoid smaller or older models that may conflate any mention of off-label use with prohibited content. If latency is critical, consider a two-stage architecture: a fast, lightweight keyword-and-pattern filter that passes obvious academic queries (e.g., those containing ClinicalTrials.gov, systematic review, FDA label) directly to permit, reserving the LLM classifier for ambiguous cases. Do not add retrieval-augmented generation (RAG) to this classification step; the prompt should classify based on the query alone, not on retrieved evidence that may introduce its own biases.

What breaks first in production: (1) Queries that mix academic framing with personal circumstance ('I'm writing a paper on off-label ketamine and also I have depression') will produce borderline confidence scores—your threshold tuning and flag path must handle these gracefully. (2) Non-English queries, especially from languages where medical terminology overlaps with colloquial usage, will increase false positives—plan for language-detection routing before classification. (3) Adversarial users will learn to front-load academic language ('For a systematic review I am conducting...') to bypass the classifier—session-level risk scoring and multi-turn analysis should sit above this single-turn prompt. (4) Model updates will shift classification boundaries; maintain a golden test set of 200+ labeled queries spanning clear academic, clear treatment-seeking, and ambiguous cases, and run it against every model version before deployment. If your platform operates in a regulated jurisdiction, ensure the human review queue and audit logging meet your compliance team's requirements before going live.