Medical Device Safety Claim Review Prompt

This prompt is designed for regulatory affairs, quality assurance, and legal teams responsible for reviewing marketing copy, product descriptions, or AI-generated content against a medical device's FDA clearance or approval scope. The primary job-to-be-done is to produce a structured, risk-flagged assessment that identifies potential off-label promotion, flags missing disclaimers, and classifies claims by regulatory risk before content is published or submitted for final review. The ideal user is a professional who already understands the device's cleared indications for use and needs a consistent, auditable first-pass analysis to accelerate their review workflow, not a replacement for their regulatory judgment.

To use this prompt effectively, you must provide two concrete inputs: the device's verbatim cleared indications for use statement and the exact content to be reviewed. The prompt is not designed for hypothetical device profiles, general disease-state awareness content, or scientific exchange materials governed by different regulatory standards. It also should not be used to review content for devices that are 510(k)-exempt, unclassified, or regulated as consumer health products without a specific clearance document. The output is a structured JSON assessment that flags claims as 'on-label,' 'potentially off-label,' or 'requires disclaimer,' but it explicitly refuses to make a final compliance determination. The assessment is a triage tool for human review, not a compliance certification.

Do not use this prompt as a substitute for a regulatory submission review by qualified personnel or as a defense in an enforcement action. The model may miss nuanced regulatory interpretations, fail to recognize implied claims that require context beyond the literal text, or misclassify claims when the cleared indications use different terminology than the marketing content. Always route the output to a human reviewer with final sign-off authority, and log both the prompt inputs and the structured assessment for audit trail purposes. If the content includes patient testimonials, comparative claims against named competitors, or statements about safety profiles not explicitly addressed in the clearance, escalate to a regulatory professional before relying on the prompt's output.

This prompt template is designed to be copied directly into your AI harness and adapted by replacing each square-bracket placeholder with the relevant content before sending it to the model. It instructs the model to act as a regulatory affairs reviewer, comparing submitted claims against a device's cleared indications for use. The output is a structured JSON assessment that flags potential off-label promotion, identifies missing disclaimers, and classifies the risk level of each claim. Because this workflow involves regulatory compliance, the prompt enforces evidence grounding and requires the model to cite the specific cleared indication that supports or conflicts with each claim.

textCopy
You are a regulatory affairs reviewer evaluating medical device safety and performance claims against FDA clearance or approval documentation.

Your task is to review the provided claim text against the device's cleared indications for use and produce a structured assessment. You must not invent clearance information, assume indications, or provide legal advice.

## INPUT

**Claim Text:**
[CLAIM_TEXT]

**Device Name:**
[DEVICE_NAME]

**Cleared Indications for Use (verbatim from FDA documentation):**
[CLEARED_INDICATIONS]

**Device Classification and Product Code:**
[DEVICE_CLASSIFICATION]

**Clearance or Approval Number (510(k), PMA, De Novo):**
[CLEARANCE_NUMBER]

## OUTPUT SCHEMA

Return a valid JSON object with the following structure:

{
  "overall_assessment": "COMPLIANT" | "REQUIRES_REVIEW" | "NONCOMPLIANT",
  "claims": [
    {
      "claim_text": "The exact claim extracted from the input",
      "claim_type": "SAFETY" | "EFFICACY" | "PERFORMANCE" | "COMPARATIVE" | "INTENDED_USE" | "OTHER",
      "within_clearance": true | false,
      "supporting_indication": "The specific cleared indication that supports this claim, or null if none",
      "risk_level": "LOW" | "MEDIUM" | "HIGH" | "CRITICAL",
      "off_label_concern": true | false,
      "missing_disclaimers": ["List of required disclaimers that are absent"],
      "recommended_action": "APPROVE" | "REVISE" | "REMOVE" | "LEGAL_REVIEW",
      "rationale": "Brief explanation of the assessment, citing the cleared indications"
    }
  ],
  "aggregate_risks": ["List of cross-cutting risk themes across all claims"],
  "requires_human_review": true | false,
  "human_review_reason": "Explanation if human review is required, otherwise null"
}

## CONSTRAINTS

1. Extract each distinct claim from the input text. A single sentence may contain multiple claims.
2. For each claim, compare it strictly against the cleared indications for use. If the claim goes beyond the cleared indications, mark it as outside clearance.
3. Flag any claim that implies safety or efficacy for an uncleared use, population, or anatomical site as an off-label concern.
4. Identify missing disclaimers such as prescription use statements, contraindication summaries, or adverse event reporting instructions.
5. If the claim text compares the device to another product, verify that the comparison is supported by the cleared indications and does not imply superiority without evidence.
6. If any claim is rated HIGH or CRITICAL risk, requires_human_review must be true.
7. Do not make legal determinations about whether the claim violates FDA regulations. Flag concerns and recommend review.
8. If the cleared indications are incomplete or ambiguous, note this in the aggregate_risks and set requires_human_review to true.

## EXAMPLES

**Example Input:**
Claim Text: "The DeviceX system reduces procedure time by 40% and eliminates the need for follow-up surgeries."
Cleared Indications: "DeviceX is indicated for use in anterior lumbar interbody fusion at L3-S1 in skeletally mature patients."

**Example Output:**
{
  "overall_assessment": "REQUIRES_REVIEW",
  "claims": [
    {
      "claim_text": "The DeviceX system reduces procedure time by 40%",
      "claim_type": "PERFORMANCE",
      "within_clearance": false,
      "supporting_indication": null,
      "risk_level": "MEDIUM",
      "off_label_concern": false,
      "missing_disclaimers": ["No clinical data citation for the 40% reduction claim"],
      "recommended_action": "REVISE",
      "rationale": "The cleared indications do not include claims about procedure time reduction. This claim requires supporting clinical evidence and may require a new 510(k) if it represents a change to intended use."
    },
    {
      "claim_text": "eliminates the need for follow-up surgeries",
      "claim_type": "SAFETY",
      "within_clearance": false,
      "supporting_indication": null,
      "risk_level": "HIGH",
      "off_label_concern": false,
      "missing_disclaimers": ["No data supporting elimination of follow-up surgeries", "No adverse event context provided"],
      "recommended_action": "REMOVE",
      "rationale": "Claim implies a safety outcome not evaluated in the cleared indications. The term 'eliminates' is absolute and requires substantial evidence. This claim presents regulatory risk."
    }
  ],
  "aggregate_risks": ["Unsubstantiated performance claims", "Absolute safety language without evidence"],
  "requires_human_review": true,
  "human_review_reason": "HIGH risk claim identified requiring regulatory assessment before publication."
}

## RISK LEVEL

[RISK_LEVEL]

If RISK_LEVEL is HIGH, apply stricter scrutiny: flag any claim that is not explicitly stated in the cleared indications, require human review for all MEDIUM and above findings, and include a note that the output must be reviewed by regulatory affairs before any downstream use.

To adapt this template, replace the placeholders with your actual data. [CLAIM_TEXT] should contain the exact marketing copy, label text, or AI-generated content under review. [CLEARED_INDICATIONS] must be the verbatim indications for use from the device's FDA clearance letter or approved labeling—do not paraphrase. [DEVICE_CLASSIFICATION] and [CLEARANCE_NUMBER] provide context that helps the model understand the regulatory pathway. The [RISK_LEVEL] parameter controls scrutiny intensity: set it to "HIGH" for promotional materials, direct-to-patient content, or new claims; use "MEDIUM" for internal drafts; use "LOW" only for educational content that does not reference a specific device. After receiving the model output, validate the JSON structure before ingesting it into your review workflow. Any output where requires_human_review is true must be routed to a qualified regulatory affairs professional before the claim is published or used in product materials.

This prompt is designed to operate as a structured review microservice within a larger regulatory affairs or quality management system. It should not be exposed directly to end users or content creators. Instead, wire it behind an internal API that accepts a claim string, device clearance context, and intended use parameters, then returns a structured assessment payload. The prompt is stateless by design—each claim review is independent—so horizontal scaling is straightforward. However, because the output directly informs regulatory compliance decisions, the implementation harness must enforce strict validation, audit logging, and human-in-the-loop approval before any finding reaches a compliance record or corrective action workflow.

Integration pattern: Build a thin orchestration layer that (1) retrieves the device's cleared indications, 510(k) or PMA summary, and labeling from your regulatory information management system, (2) assembles the prompt with the claim under review plus that retrieved context, (3) calls the model with response_format set to the defined JSON schema, (4) validates the output against the schema and business rules, and (5) routes results to a review queue. Use OpenAI structured outputs or Claude's tool-use mode with strict schema to enforce the output shape at the API level rather than relying solely on prompt instructions. For model choice, prefer a model with strong instruction-following and low hallucination rates on classification tasks—GPT-4o or Claude 3.5 Sonnet are appropriate starting points. Avoid smaller or older models that may conflate regulatory categories or miss off-label signals.

Validation and retry logic: After receiving the model response, validate that classification is one of the allowed enum values (within_scope, potential_off_label, requires_disclaimer, unsubstantiated, insufficient_context), that risk_level is low, medium, or high, and that every evidence_reference field contains a non-empty string pointing to a real document section. If validation fails, retry once with the validation error injected into the prompt as additional [CONSTRAINTS] context. If the second attempt also fails, flag the review for manual completion and log the failure for prompt improvement analysis. Do not silently accept malformed outputs—a missing classification is a compliance gap, not a minor formatting issue.

Human review and audit trail: Every claim review, regardless of automated classification, should be written to an immutable audit log that captures the original claim text, the full prompt sent, the raw model response, the validated output, the reviewer's identity, and any override decisions. For claims classified as potential_off_label or unsubstantiated with risk_level: high, route to a regulatory affairs reviewer before the finding is surfaced to the content owner. The human approval step should present the model's assessment alongside the source clearance documents so the reviewer can verify the classification against primary evidence. Never allow the model's output to auto-populate a 483 response, warning letter reply, or promotional material revision without human sign-off.

What to avoid: Do not use this prompt in a real-time content generation pipeline where marketing copy is checked inline during drafting—latency and the need for human review make batch or submit-triggered review more appropriate. Do not treat the model's classification as a definitive regulatory determination; it is a triage and flagging tool. Do not skip the retrieval step and ask the model to recall device clearance details from training data—always provide the specific cleared indications and intended use as [CONTEXT]. Finally, do not log claim text or review results to non-compliant systems; ensure your logging pipeline meets your organization's document control and data integrity requirements for regulatory records.