Violence and Harm Prevention Redirection Prompt

This prompt is a specialized, production-grade safety instrument for trust-and-safety teams and safety engineers. Its sole job is to generate a firm, non-negotiable refusal when a user request has already been classified as involving threats, violence planning, or dangerous activity. It is not a detection mechanism. The prompt is designed to be the final response layer in a multi-stage safety architecture, operating on the assumption that an upstream classifier has already flagged the input as high-risk and routed it here. The core value is a refusal that doesn't just say 'no' but simultaneously redirects the user to de-escalation resources and crisis support, turning a safety enforcement moment into a potential intervention point.

Integrate this prompt when your system requires a deterministic, high-assurance refusal posture for the most severe category of unsafe content. The prompt is built for scenarios where a soft refusal, a warning, or an educational alternative is insufficient and could be interpreted as endorsement or facilitation. It expects a specific input context, including the original user query and a pre-determined [RISK_LEVEL] and [THREAT_CATEGORY], to tailor the redirection resources appropriately. The output is a structured refusal that must be logged, monitored, and subject to human review in a subset of cases to ensure the redirection resources remain accurate and the refusal tone hasn't drifted.

Do not use this prompt for low-severity policy violations, academic discussions of violence in historical or literary contexts, or legitimate security research inquiries. Using a non-negotiable refusal for a borderline or misclassified input will create a severe user experience failure, eroding trust and potentially blocking legitimate work. Those scenarios require different refusal postures—such as a warning with an educational alternative or a clarification request—which are covered in sibling playbooks like the 'Safe Alternative and Redirection Prompt Template' and 'Cybersecurity Misuse Redirection Prompt.' Before deploying this prompt, ensure your upstream classification and routing logic has a high precision for true high-risk threats to avoid over-refusal and unnecessary crisis resource redirection.

This template is designed to be a self-contained safety instruction for your AI system. It handles the detection of requests related to violence, threats, or planning of dangerous activities, and produces a structured refusal that prioritizes user safety and redirects to appropriate support. The prompt is structured to be strict on the core violation while remaining helpful in its redirection, a balance that is critical for maintaining user trust and safety in production systems.

textCopy
You are a safety-classification and redirection layer. Your task is to analyze the user's request and determine if it involves threats of violence, planning of violent acts, detailed instructions for causing physical harm, or glorification of dangerous activities against oneself or others.

If the request does NOT violate this policy, respond with the single word: SAFE.

If the request DOES violate this policy, you must NOT fulfill the request. Instead, generate a structured refusal using the following template. Do not deviate from this structure.

REFUSAL TEMPLATE:
---
I can't provide instructions or support for that request, as it goes against my safety policy on violence and harm. If you are in immediate danger or are having thoughts of harming yourself or others, please contact emergency services or a crisis support line.

Here are some resources that can provide immediate, confidential support:
- **Crisis Text Line:** Text HOME to 741741 from anywhere in the US, anytime.
- **National Suicide Prevention Lifeline:** Call or text 988 in the US.
- **The Trevor Project:** 1-866-488-7386 for LGBTQ youth in crisis.

If you are not in crisis, I can help you with information on non-violent conflict resolution, stress management, or finding a local mental health professional. Would you like to explore one of these safer alternatives?
---

[INPUT]:
"""
[USER_REQUEST]
"""

To adapt this template, replace the [USER_REQUEST] placeholder with the actual user input in your application logic. The resource list should be customized based on your user's geographic location, which can be determined from session metadata or IP geolocation. The final offer of safer alternatives can also be extended to include specific, relevant pathways your application supports, such as linking to a help center article on conflict resolution.

Before deploying this prompt, you must build a validation layer around it. A simple string match for SAFE is insufficient. Implement a post-processing check that confirms the refusal template's structure is intact if a violation is detected. Log all violations for audit and safety review. Crucially, this prompt is a first line of defense; for high-risk applications, the detected violation should also trigger an internal escalation to a human trust-and-safety team for review, independent of the user-facing redirection.

This prompt is designed to sit inside a safety screening layer that evaluates user input before it reaches a generative model or agent. The implementation harness must classify the request, invoke the redirection prompt when a violence or harm threshold is met, validate the output, and decide whether to return the redirection message, escalate to human review, or both. Because the domain involves threats, violence planning, and dangerous activity requests, the harness must treat every positive classification as a potential high-severity incident until proven otherwise. The prompt itself is not a standalone safety solution—it is one component in a pipeline that includes upstream classification, downstream validation, and mandatory logging for audit and review.

Wire the prompt into an application by first running a violence and harm classifier on the user input. Use a separate classification prompt or a fine-tuned model that returns a structured risk object with fields for severity (low, medium, high, critical), category (threat, planning, incitement, dangerous activity, other), and confidence (0.0–1.0). If severity is medium or above, route the input to the Violence and Harm Prevention Redirection Prompt with the [INPUT] placeholder set to the original user message, [RISK_LEVEL] set to the classified severity, and [CONTEXT] populated with the category and confidence. The prompt returns a refusal message and a set of de-escalation or crisis resources. Validate the output by checking that: (1) the response does not repeat, summarize, or engage with the harmful content; (2) at least one resource is present and matches the risk category; (3) the tone is firm but not aggressive. If validation fails, retry once with an explicit instruction to correct the specific failure. If the second attempt also fails, escalate to a human reviewer and log the failure for prompt improvement. For critical severity classifications, bypass the prompt entirely and route directly to human review with the classified risk object and original input.

Model choice matters here. Use a model with strong instruction-following and safety alignment—GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro are reasonable defaults. Avoid smaller or less-aligned models that may leak harmful content or produce inconsistent refusal tone. Set temperature to 0 or very low (0.0–0.2) to maximize output determinism and reduce the risk of creative deviations from the refusal script. Log every invocation: store the original input hash, classification result, prompt version, model response, validation result, and final action (returned, escalated, or both). Do not log the raw user input if it contains explicit threats or harmful content—use a one-way hash for traceability without storing dangerous material in plaintext. This log becomes your audit trail for safety incidents and your dataset for improving classification thresholds and prompt performance over time. The next step is to integrate this harness into your existing safety pipeline and run a regression suite of known violent and benign inputs to confirm that classification thresholds, prompt behavior, and escalation rules work as expected before production deployment.