Clinical Trial Data Q&A Prompt Template

This prompt template is designed for research teams, clinical analysts, and pharmacovigilance professionals who need to query clinical trial protocols, results summaries, adverse event reports, and regulatory filings. The prompt enforces strict separation of primary and secondary endpoints, requires trial phase context in every answer, and mandates statistical significance notes when numerical results are cited. Use this prompt when your retrieval pipeline returns chunks from ClinicalTrials.gov records, FDA review documents, published trial manuscripts, or internal clinical study reports. The prompt assumes retrieved context contains structured trial data and will refuse to answer when evidence is insufficient.

Do not use this prompt for patient-level clinical decision support or diagnostic reasoning. Those workflows require separate prompts with explicit human-review gates and medical disclaimer language. This prompt is also unsuitable for general biomedical literature review without trial-specific structure, as it expects fields like enrollment counts, arm descriptions, and p-values to be present in the retrieved context. Before deploying, verify that your retrieval pipeline consistently surfaces trial registry entries, results databases, or regulatory documents rather than press releases or news summaries, which lack the statistical rigor this prompt demands.

The prompt is designed to fail safely: it will output ANSWERABLE: false when the retrieved context lacks trial phase, endpoint definitions, or statistical results. This abstention behavior is intentional and should be respected in your application logic. Do not attempt to override it by loosening the constraints or removing the evidence-sufficiency check. If your use case requires answering from weaker evidence, use a separate prompt with appropriately lowered confidence thresholds and explicit caveats about source quality. Wire this prompt into a pipeline where a human reviewer can inspect abstentions and decide whether to broaden the retrieval or escalate the question.

This template is designed to be pasted directly into your system prompt or user message template when building a Q&A system over clinical trial protocols, results, and adverse event reports. It enforces strict grounding in provided trial documents, separates primary from secondary endpoints, and requires statistical significance context and population limitation notes. The square-bracket placeholders let you inject the specific trial context, user question, and output constraints at runtime without modifying the core instruction set.

textCopy
You are a clinical trial data assistant. Your only knowledge is the trial context provided below. You must not use any external knowledge about drugs, diseases, or medical facts.

[TRIAL_CONTEXT]

Answer the user question using only the information in the trial context above. Follow these rules exactly:

1. **Source grounding**: Every factual claim must be traceable to a specific passage in the trial context. If the context does not contain the answer, state that clearly rather than guessing.

2. **Trial phase context**: Identify the trial phase (Phase I, II, III, IV) from the context and note how it limits the generalizability of the findings.

3. **Endpoint separation**: Clearly distinguish between primary endpoints, secondary endpoints, and exploratory endpoints. Never present a secondary endpoint finding as if it were a primary endpoint result.

4. **Statistical reporting**: When reporting results, include p-values, confidence intervals, and effect sizes if present in the context. Note whether results reached statistical significance. If significance information is missing, state that it was not provided.

5. **Population limitations**: Explicitly note the inclusion/exclusion criteria, sample size, demographics, and any factors that limit generalizability to broader populations.

6. **Adverse events**: When discussing adverse events, report severity grades, frequencies, and attribution assessments exactly as stated in the context. Do not minimize or amplify safety signals.

7. **Uncertainty language**: Use precise uncertainty language. Distinguish between 'the trial demonstrated', 'the trial suggested', 'the data were inconclusive regarding', and 'the context does not address'.

8. **Abstention**: If the question asks for clinical advice, treatment recommendations, or off-label use, refuse to answer and state that you provide information retrieval only, not medical advice.

User question: [USER_QUESTION]

Output format: [OUTPUT_SCHEMA]

Additional constraints: [CONSTRAINTS]

Adaptation notes: Replace [TRIAL_CONTEXT] with your retrieved trial documents, ensuring each chunk includes metadata about the trial phase, NCT number, and publication date. The [OUTPUT_SCHEMA] placeholder should contain your expected JSON structure or markdown format specification. Use [CONSTRAINTS] to inject domain-specific rules such as 'only use results from the final analysis, not interim analyses' or 'flag any sponsor-funded studies'. For regulated environments, add a [RISK_LEVEL] placeholder that gates whether the system should append a mandatory human-review disclaimer. Test this template with golden Q&A pairs where you know exactly which passages should be cited, and build an eval that checks for hallucinated statistics, missing population caveats, and endpoint conflation.

This prompt is designed to operate inside a Retrieval-Augmented Generation (RAG) pipeline where the model receives pre-retrieved clinical trial documents as [CONTEXT]. The implementation harness must ensure that the prompt receives clean, well-chunked, and metadata-rich context before generation. The application layer should handle document retrieval, chunk assembly, and metadata injection (trial ID, phase, sponsor, date) before populating the [CONTEXT] placeholder. The model should never be responsible for searching or selecting documents; it should only synthesize from what it is given.

Validation and Output Guardrails: After generation, the output must pass through a structured validation layer before reaching any user. Parse the model's response to verify that every claim tagged with a citation has a corresponding passage in the provided [CONTEXT]. Use a secondary LLM call or a deterministic string-match check to flag hallucinated citations. Validate that the output schema is respected: primary endpoints, secondary endpoints, statistical significance notes, and population limitations must be present as distinct sections. If the model fails to separate primary from secondary endpoints, or if it makes a claim without a source, route the output to a repair prompt or escalate for human review. Retries: If validation fails, implement a single retry with the original prompt plus the specific validation error message appended to [CONSTRAINTS]. Do not retry more than once for clinical use cases; escalate instead.

Model Choice and Tool Use: This prompt is optimized for models with strong instruction-following and long-context handling, such as GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro. Do not use smaller open-weight models without extensive eval on hallucination rates for clinical text. The prompt does not require tool use or function calling during generation, but the harness should log every input, output, and validation result to an audit trail. For regulated environments, store the full prompt, retrieved context, model response, and validation report as an immutable record. Human Review Trigger: Any output that includes a statistical significance claim (p-value, confidence interval) or a comparative efficacy statement must be flagged for human review before it reaches an end user. The harness should block automated delivery of such outputs and route them to a review queue with the source documents attached.

What to Avoid: Do not use this prompt in a zero-context mode or with web search. The clinical domain requires grounding in specific, vetted trial documents. Do not allow the model to generate patient-specific advice or treatment recommendations; the prompt's scope is trial data Q&A, not clinical decision support. If a user question asks for medical advice, the harness should intercept it before generation and return a refusal message. Finally, do not deploy this prompt without a regression test suite that includes known hallucination-prone questions, missing-evidence scenarios, and edge cases where trials have conflicting results.