Self-Harm and Suicide Risk Escalation Prompt

This prompt is designed for platform safety engineers and customer support system architects who need to detect self-harm ideation, suicide risk, or acute crisis signals in unstructured user-generated text. The primary job-to-be-done is not to provide therapeutic responses, but to act as a reliable classification and escalation trigger at the ingress layer. It ingests a user message—often from a chat, support ticket, or community post—and produces a structured risk assessment that downstream systems can use to inject crisis resources, prioritize queues, or lock accounts for human review. The ideal user is someone integrating this into a production AI pipeline where a missed true positive is a life-safety failure and a false positive degrades user trust.

Use this prompt when you need a deterministic, auditable risk signal before any generative model crafts a reply. It is appropriate for platforms where users may express distress directly or through veiled language, and where your organization has a defined escalation path (e.g., on-call clinical reviewers, crisis text line integrations, or mandatory reporting workflows). The prompt requires the raw user input and, optionally, conversation history for multi-turn context. It is not a replacement for clinical judgment, nor is it a chatbot for crisis counseling. Do not use this prompt to generate responses to the user in distress; its output should route to a human or inject pre-approved safety resources. It is also unsuitable for analyzing fictional content, song lyrics, or academic discussions of self-harm without additional context, as these are common sources of false positives.

Before deploying, you must calibrate the prompt against your platform's specific linguistic patterns. Metaphorical language ('I'm dead inside'), dark humor, and mental health advocacy discussions frequently trigger naive keyword-based systems. This prompt's harness requires evaluation datasets that include these edge cases, plus a human-review step for all high-severity classifications. If your platform lacks a trained human review team or a clear escalation policy, do not deploy this prompt in an automated pipeline. Start with a shadow mode that logs classifications without taking action, and use that data to tune thresholds and build your operational response before going live.

This prompt template is designed to be dropped into a classification or triage pipeline where user-generated text must be evaluated for self-harm or suicide risk signals before any automated response is generated. The template uses square-bracket placeholders for all variable inputs, allowing platform engineers to wire it into existing moderation queues, support ticket systems, or conversational AI guardrails. The output schema is intentionally strict: it forces the model to produce a risk level, an urgency score, extracted evidence, and a concrete escalation recommendation rather than a vague warning. This structure is critical because downstream routing logic, SLA timers, and crisis resource injection all depend on machine-readable fields, not natural-language hedging.

textCopy
You are a safety classifier operating in a production trust and safety pipeline. Your task is to analyze user-submitted text for signals of self-harm ideation, suicide risk, or active crisis. You must produce a structured, machine-readable output that downstream systems can use for routing, escalation, and crisis resource injection.

## INPUT
[USER_INPUT]

## CONTEXT
- Conversation history (if available): [CONVERSATION_HISTORY]
- User metadata (if available): [USER_METADATA]
- Platform policy for self-harm content: [PLATFORM_POLICY]

## CLASSIFICATION TAXONOMY
Classify the input into exactly one of the following risk levels:
- `none`: No self-harm or suicide-related content detected.
- `low`: Mentions of distress, sadness, or vague hopelessness without ideation, method, or intent.
- `moderate`: Expressions of self-harm ideation, passive death wishes, or non-specific suicidal thoughts without plan or intent.
- `high`: Specific suicidal ideation with method, plan, or expressed intent. Includes active self-harm descriptions.
- `crisis`: Imminent risk with stated intent, ongoing attempt, or clear timeframe. Requires immediate intervention.

## URGENCY SCORING
Assign an urgency score from 0 to 100 where:
- 0-20: No time pressure; standard processing acceptable.
- 21-50: Elevated concern; prioritize review within hours.
- 51-80: High concern; escalate within minutes.
- 81-100: Critical; immediate human intervention required.

## OUTPUT SCHEMA
Return ONLY valid JSON matching this schema:
{
  "risk_level": "none" | "low" | "moderate" | "high" | "crisis",
  "urgency_score": <integer 0-100>,
  "evidence_excerpts": [<list of direct quotes from the input that support the classification>],
  "crisis_signals_detected": [<list of specific signals: "method_mentioned", "plan_described", "intent_stated", "timeframe_specified", "hopelessness", "goodbye_language", "self_harm_description", "previous_attempt_reference", "isolation_language", "burden_language">],
  "escalation_path": "none" | "flag_for_review" | "priority_queue" | "immediate_human_escalation" | "emergency_services",
  "crisis_resource_text": <string containing appropriate crisis resource information if risk_level is moderate or above, otherwise empty string>,
  "false_positive_risk_factors": [<list of factors that could indicate this is a false positive, such as "metaphorical_language", "clinical_discussion", "third_person_reporting", "fictional_content", "song_lyrics", "academic_context">],
  "requires_human_review": <boolean>
}

## CONSTRAINTS
1. Do NOT generate a conversational response, empathy statement, or direct reply to the user. This output is for routing logic only.
2. If the input contains metaphorical language (e.g., "I could just die of embarrassment"), clinical discussion, song lyrics, or third-person reporting, flag these in `false_positive_risk_factors` and do NOT escalate unless additional signals are present.
3. If `risk_level` is `moderate` or above, `crisis_resource_text` MUST contain appropriate crisis resource information. Use the following resource format: "If you're experiencing a crisis, please contact [CRISIS_LINE_NAME] at [CRISIS_LINE_NUMBER] or text [CRISIS_TEXT_LINE]. You are not alone."
4. If `risk_level` is `high` or `crisis`, `requires_human_review` MUST be `true` and `escalation_path` MUST be `immediate_human_escalation` or `emergency_services`.
5. Never minimize, dismiss, or reframe the user's expressed distress. Err on the side of caution when signals are ambiguous.
6. Do not hallucinate evidence. Every excerpt in `evidence_excerpts` must be a verbatim substring from the input.

## EXAMPLES
[FEW_SHOT_EXAMPLES]

## INSTRUCTIONS
Analyze the input above and return ONLY the JSON object. No preamble, no explanation, no markdown fences.

To adapt this template for your platform, replace each square-bracket placeholder with your actual data and configuration. The [PLATFORM_POLICY] placeholder should contain your organization's specific self-harm content policy, including any regulatory obligations (e.g., mandatory reporting requirements in your jurisdiction). The [CRISIS_LINE_NAME], [CRISIS_LINE_NUMBER], and [CRISIS_TEXT_LINE] placeholders in the constraint section should be replaced with region-appropriate crisis resources—consider making these dynamic based on user locale if your platform operates across multiple regions. The [FEW_SHOT_EXAMPLES] placeholder is critical for calibration: include at least three examples covering a clear non-risk case (metaphorical language), a moderate-risk case (passive ideation), and a crisis case (imminent intent). Without these examples, the model will default to over-flagging or under-flagging depending on its base training distribution.

Before deploying this prompt into production, you must run it through a regression test suite that includes: metaphorical language samples ("I'm dead tired"), clinical mental health discussions, song lyrics containing death references, third-person reports ("my friend is struggling"), and genuine crisis statements. Measure false positive rate, false negative rate, and calibration of urgency scores against human-labeled ground truth. If your false positive rate on metaphorical language exceeds 5%, add more counterexamples to [FEW_SHOT_EXAMPLES] or tighten the false_positive_risk_factors enumeration. Never deploy this prompt without a human-in-the-loop path for high and crisis classifications—the model can flag risk, but it cannot and should not be the final decision-maker in life-critical situations.

This prompt is a high-stakes classification step that must sit behind a hardened API endpoint, never exposed directly to end users without surrounding safeguards. The implementation harness should treat every model response as a preliminary signal, not a final decision. Wire the prompt into an application layer that validates the output schema, enforces confidence thresholds, and routes to human review when risk is elevated. The harness must also handle model unavailability gracefully: a failed classification should default to the safest escalation path rather than silently dropping the input.

Build the harness with these concrete components. Input sanitization: strip or escape any content that could interfere with the prompt template before insertion, but preserve the original raw input for audit logs. Model invocation: use a low-latency model with consistent JSON output (GPT-4o or Claude 3.5 Sonnet are reasonable defaults). Set temperature=0 to reduce variance on risk scores. Implement a retry wrapper with exponential backoff (max 3 attempts) that validates the response against the expected schema on each attempt. Output validation: parse the JSON response and verify that risk_level is one of the allowed enum values, urgency_score is an integer within range, and escalation_path is a recognized route. Reject and retry any response that fails schema validation. Threshold routing: if risk_level is high or imminent, bypass any queuing and trigger immediate human review with the full context payload. If risk_level is moderate, inject crisis resource text into the user-facing response and flag for review within the SLA window. If risk_level is low or none, log the classification and proceed with the normal workflow. Logging and audit: write every classification event to an append-only log with timestamp, raw input, model response, validated output, retry count, and routing decision. This log is critical for post-incident review and regulatory demonstration.

The most dangerous failure mode is a false negative where the model returns risk_level: none for genuine crisis content. Mitigate this by implementing a secondary keyword-triggered override: if the raw input matches a curated list of high-signal crisis terms (maintained outside the model), force-escalate regardless of the model's classification. Similarly, test extensively for false positives on metaphorical language (e.g., 'this deadline is killing me') and mental health support discussions that do not indicate imminent risk. Run a monthly eval against a held-out test set that includes both crisis and benign mental health content, and track precision-recall drift over time. Do not deploy this prompt without a documented human review SLA and an on-call rotation for high-severity escalations.