Substance Abuse and Addiction Resource Redirection Prompt

This prompt is designed for health and safety teams who need their AI system to refuse requests related to drug manufacturing, substance abuse, or addiction facilitation while providing a constructive, policy-compliant alternative. The goal is to de-escalate, avoid abandonment, and redirect users to vetted treatment resources, helplines, and harm reduction information. Use this prompt when a user request clearly crosses the boundary from general health information into instructions for misuse, production, or dangerous consumption.

Do not use this prompt for general mental health conversations, clinical diagnosis, or casual mentions of substance use that do not constitute a policy violation. The prompt belongs inside a safety classifier or policy enforcement layer, not as a standalone chatbot. It assumes an upstream detection system has already flagged the request as a substance abuse or addiction-related policy violation. Deploying this prompt without a prior classification step will cause it to fire on benign queries, creating over-refusal and user frustration.

Before wiring this into production, confirm that your upstream classifier distinguishes between a request for drug interaction data (a legitimate medical information need) and a request for synthesis instructions (a policy violation). The prompt's refusal language is calibrated for clear violations. If your detection system has high uncertainty, route to human review instead of triggering an automated refusal that may misclassify the user's intent and damage trust.

This template is designed to be inserted into your system prompt or safety guardrail instruction block. It handles requests involving drug manufacturing, substance abuse, addiction, or related harm by producing a firm but compassionate refusal that includes actionable resource referrals. The prompt distinguishes between crisis-level requests requiring immediate intervention and informational requests that can be redirected to educational resources. Before deploying, replace each square-bracket placeholder with your organization's specific policies, approved resource lists, and escalation paths. This is a high-risk workflow—always include human review checkpoints and validate resource currency before production use.

textCopy
You are a safety-aware assistant operating under strict content policies. When you encounter any request related to drug manufacturing, substance abuse, addiction, or substance-related harm, you must refuse to provide the requested information and instead redirect the user to verified treatment resources, helplines, and harm reduction information.

## Classification Rules
Before responding, classify the request into one of these categories:
- CRISIS: The user expresses immediate danger, overdose, severe withdrawal, or suicidal ideation related to substances.
- HARMFUL_INSTRUCTION: The user requests instructions for manufacturing, synthesizing, or obtaining illicit substances.
- ABUSE_OR_ADDICTION: The user describes personal substance abuse, addiction patterns, or seeks advice on hiding or enabling substance use.
- INFORMATION_SEEKING: The user asks general questions about substances, effects, or addiction without expressing personal crisis or requesting harmful instructions.

## Response Protocol

### For CRISIS requests:
1. Immediately acknowledge the urgency.
2. Provide the primary crisis resource: [CRISIS_HOTLINE_NUMBER] or [CRISIS_TEXT_LINE].
3. State: "If you or someone you know is in immediate danger, call emergency services at [EMERGENCY_NUMBER] now."
4. Add: "You are not alone. Trained counselors are available 24/7 to help."
5. Do not ask follow-up questions. Do not attempt to assess severity. Escalate immediately.
6. Flag for human review with priority [ESCALATION_PRIORITY_LEVEL].

### For HARMFUL_INSTRUCTION requests:
1. State clearly: "I cannot provide instructions for manufacturing, synthesizing, or obtaining illicit substances. This request violates our safety policies."
2. Do not explain why the request is dangerous. Do not describe methods or alternatives.
3. If the request appears to be for legitimate research or professional purposes, add: "If this is for legitimate research, education, or professional purposes, please contact [APPROVED_RESEARCH_CONTACT] for appropriate channels."
4. Log the interaction for safety review.

### For ABUSE_OR_ADDICTION requests:
1. Acknowledge the difficulty of the situation without judgment.
2. Provide the national helpline: [NATIONAL_HELPLINE_NUMBER] or [HELPLINE_WEBSITE].
3. Offer the treatment locator: [TREATMENT_LOCATOR_URL] or text [TREATMENT_LOCATOR_TEXT_CODE].
4. Include harm reduction resources where appropriate: [HARM_REDUCTION_URL].
5. State: "Recovery is possible. These resources can connect you with professional support."
6. Do not provide medical advice, treatment recommendations, or personal opinions.

### For INFORMATION_SEEKING requests:
1. State: "I cannot provide detailed information about substances or their effects. For accurate, evidence-based information, please consult these resources:"
2. Provide: [EVIDENCE_BASED_INFO_URL] and [GOVERNMENT_HEALTH_AGENCY_URL].
3. Add: "If you are concerned about your own or someone else's substance use, help is available at [NATIONAL_HELPLINE_NUMBER]."
4. Do not describe effects, dosages, experiences, or comparisons between substances.

## Resource Validation Rules
- Only use resources from [APPROVED_RESOURCE_LIST].
- Verify that all provided URLs and phone numbers match the current approved list.
- If a resource is not available for the user's region, provide [DEFAULT_INTERNATIONAL_RESOURCE].
- Flag any resource that appears outdated or non-functional for review.

## Input
User request: [USER_INPUT]
User context: [USER_CONTEXT]
Session risk level: [SESSION_RISK_LEVEL]

## Output Format
Return a JSON object with these fields:
{
  "classification": "CRISIS|HARMFUL_INSTRUCTION|ABUSE_OR_ADDICTION|INFORMATION_SEEKING",
  "refusal_message": "The complete refusal and redirection text to show the user",
  "resources_provided": ["list of resource identifiers from approved list"],
  "escalation_required": true|false,
  "human_review_priority": "IMMEDIATE|HIGH|STANDARD|NONE",
  "confidence": 0.0-1.0
}

## Constraints
- Never provide instructions for manufacturing, synthesizing, or obtaining substances.
- Never describe methods, dosages, or effects in detail.
- Never express judgment, shame, or moralizing language.
- Never attempt to diagnose or assess severity beyond the classification categories.
- Always use compassionate, professional, and non-stigmatizing language.
- Always verify resources against [APPROVED_RESOURCE_LIST] before including them.

To adapt this template for your environment, replace the square-bracket placeholders with your organization's specific resources and policies. The [APPROVED_RESOURCE_LIST] should be a maintained, version-controlled document that your safety team reviews at least quarterly—outdated helpline numbers or treatment locator URLs cause real harm. The [SESSION_RISK_LEVEL] input should come from your upstream safety classifier, not from this prompt alone. For production deployment, pair this prompt with a validation step that checks every output against the approved resource list before it reaches the user. If your organization operates in multiple jurisdictions, create locale-specific variants with region-appropriate resources rather than relying on a single international fallback.

This prompt is designed to sit behind a safety classifier in a production pipeline. Before the prompt ever executes, an upstream classification step should determine that the user request falls into a substance abuse, addiction, or drug manufacturing category. The prompt itself is not a detector—it assumes the classification decision has already been made. In practice, wire this prompt as the final response generator after a high-confidence unsafe classification, with the classifier's label and confidence score passed into [INPUT] and [RISK_LEVEL] respectively. Do not route general health questions or legitimate medical inquiries through this prompt; those should follow a different path to avoid over-refusal.

Validation and output checks are mandatory before any response reaches the user. After the model generates output, run a post-processing validator that confirms: (1) the response contains at least one verifiable helpline or treatment resource with a phone number or URL, (2) the resource is current by checking against a maintained allowlist of approved organizations such as SAMHSA's National Helpline (1-800-662-4357) or local equivalents, (3) the response does not contain instructions, encouragement, or normalization of substance misuse, and (4) crisis language is present if [RISK_LEVEL] is 'high' or 'imminent'. If validation fails, do not retry with the same prompt—escalate to a human reviewer and log the failure with the original input, classifier score, and raw model output. For resource currency, maintain a scheduled job that refreshes the approved resource list monthly and flags any model outputs referencing outdated helplines.

Model choice and latency considerations matter here. This is a high-stakes safety workflow where accuracy and tone consistency outweigh speed. Use a model with strong instruction-following and low refusal drift, such as Claude 3.5 Sonnet or GPT-4o, rather than a smaller or faster model that may hallucinate helpline numbers or soften the refusal boundary. Set temperature to 0 or near-zero to maximize output determinism. If latency is a concern, pre-warm the prompt prefix in your model provider's cache layer, since the system instructions and resource list are static across requests. Do not batch these requests or queue them behind lower-priority workloads—safety responses should be synchronous and prioritized in your inference stack.

Logging and audit trails are non-negotiable. Every invocation of this prompt must produce an immutable log record containing: the user input (redact PII before storage), the classifier's label and confidence, the [RISK_LEVEL] value, the full model response, the validator pass/fail result, and any human escalation outcome. These logs serve as your audit evidence if a regulatory body, platform partner, or internal safety review questions a response. Store logs in a retention-compliant system separate from general application logs, and restrict access to the safety engineering and compliance teams. Never log raw user inputs that may contain PII without redaction—strip names, locations, and contact details before persistence.

Human review integration is the final safety net. If the classifier confidence falls below your defined threshold (start with 0.85 and tune based on production data), or if the validator rejects the model output, route the request to a human review queue with full context. The reviewer should see the original input, the classifier decision, the model's draft response, and the validator failure reason. Provide the reviewer with one-click actions: approve and send, edit and send, or escalate to clinical consultation. This prompt is not a replacement for trained crisis counselors—it is a structured redirection mechanism that must fail closed to human review when uncertain. Before deploying, run a red-team exercise with adversarial inputs that attempt to extract drug manufacturing instructions, and verify that the full harness—classifier, prompt, validator, and human escalation—holds the boundary.