Domestic Abuse Support Resource Redirection Prompt

This prompt is designed for support platforms, helpline chatbots, and community safety tools that must provide domestic abuse resources without over-refusing related legal, housing, or counseling queries. The primary job-to-be-done is to classify a user's message and produce a risk-aware response that offers vetted safety resources when the query indicates potential abuse, while permitting professional and informational discussions about family law, tenant rights, or mental health support. The ideal user is a product engineer or trust-and-safety analyst integrating this prompt into an AI feature where false positives (incorrectly redirecting a benign query) degrade user trust and false negatives (failing to provide resources for an at-risk user) create real-world harm.

Use this prompt when you need a single classification-and-response step that distinguishes between three categories: (1) a direct or implied cry for help requiring immediate resource provision, (2) a professional or informational query about adjacent topics like divorce proceedings, custody arrangements, or counseling modalities that should be answered normally, and (3) an ambiguous query where the safest path is to provide resources alongside the requested information. The prompt requires a pre-vetted list of crisis resources, hotlines, and shelter directories specific to the user's jurisdiction. It also requires clear policy definitions for what constitutes an at-risk query versus a professional discussion. Do not use this prompt as a standalone safety mechanism—it must be paired with downstream human review for high-risk classifications and regular eval checks for resource accuracy.

This prompt is not appropriate for generating therapeutic advice, conducting risk assessments that determine immediate danger, or replacing trained crisis counselors. It is a redirection and resource-provision mechanism, not a diagnostic tool. Before deploying, ensure you have a process for keeping resource lists current, a human escalation path for ambiguous or high-risk classifications, and a regression test suite that includes both genuine at-risk queries and benign family law, housing, and counseling questions to measure over-refusal rates. The next section provides the copy-ready prompt template you can adapt with your own resource lists and policy definitions.

This template is designed to be dropped into a classification or routing layer before your main conversational model responds. It separates queries that genuinely need domestic abuse resource redirection from professional, legal, housing, or counseling queries that only superficially resemble at-risk content. The prompt forces a structured classification decision with explicit reasoning, which makes the output auditable and testable. Use it as a pre-processing gate: if the classification returns needs_resource_redirection, your application should surface vetted crisis resources rather than passing the query to a general-purpose model that might refuse too broadly or, worse, attempt to counsel the user.

textCopy
You are a risk-aware query classifier for a support platform. Your job is to determine whether a user query requires redirection to domestic abuse support resources or whether it is a legitimate professional, legal, housing, or counseling query that should proceed normally.

## INPUT
User query: [USER_QUERY]
User context (if available): [USER_CONTEXT]
Conversation history summary: [CONVERSATION_HISTORY]

## CLASSIFICATION CATEGORIES
Choose exactly one:
- `needs_resource_redirection`: The query indicates the user may be experiencing, at risk of, or seeking help for domestic abuse. This includes direct disclosures, indirect signals of controlling relationships, safety planning questions, and requests for shelter or crisis support.
- `professional_discussion`: The query discusses domestic abuse in a professional, academic, legal, policy, or educational context. Examples: family law attorneys discussing protective orders, social workers asking about intake procedures, researchers studying intervention outcomes, journalists writing about domestic violence policy.
- `adjacent_topic`: The query touches on relationships, housing, counseling, or family conflict but does not indicate domestic abuse risk. Examples: general relationship advice, tenant rights questions, divorce process information, parenting coordination, general therapy questions.
- `unrelated`: The query does not involve domestic abuse, relationship violence, or adjacent topics.

## OUTPUT SCHEMA
Return ONLY valid JSON with this exact structure:
{
  "classification": "<one of the four categories above>",
  "confidence": <float between 0.0 and 1.0>,
  "reasoning": "<brief explanation of the evidence that supports this classification>",
  "risk_indicators": ["<list of specific phrases or signals that contributed to the classification>"],
  "recommended_action": "<resource_redirect | normal_processing | escalate_to_human>"
}

## CONSTRAINTS
- If classification is `needs_resource_redirection`, recommended_action MUST be `resource_redirect`.
- If classification is `professional_discussion` or `adjacent_topic`, recommended_action MUST be `normal_processing` unless confidence is below 0.7, in which case use `escalate_to_human`.
- Never include resource contact details in the classification output. The application layer handles resource surfacing.
- Do not refuse to classify. Even ambiguous queries must receive a best-effort classification with lowered confidence.
- If the query contains explicit threats of immediate harm, set recommended_action to `escalate_to_human` regardless of classification.

## EXAMPLES

Query: "My partner checks my phone every night and won't let me see my friends. Is this normal?"
Classification: needs_resource_redirection
Confidence: 0.92
Reasoning: Direct disclosure of controlling behavior with help-seeking framing. Monitoring communications and social isolation are recognized abuse indicators.
Risk indicators: ["checks my phone", "won't let me see my friends", "is this normal"]
Recommended action: resource_redirect

Query: "I'm a family law attorney drafting a protective order. What are the evidentiary requirements in California?"
Classification: professional_discussion
Confidence: 0.95
Reasoning: Explicit professional role identification, legal procedure context, jurisdiction-specific question. No personal risk disclosure.
Risk indicators: []
Recommended action: normal_processing

Query: "What are the grounds for divorce in my state?"
Classification: adjacent_topic
Confidence: 0.88
Reasoning: Legal information request about divorce process. No indicators of abuse, coercion, or safety concerns. Adjacent to domestic abuse context but not indicative of risk.
Risk indicators: []
Recommended action: normal_processing

Query: "I need to find a shelter tonight. I can't stay here anymore."
Classification: needs_resource_redirection
Confidence: 0.96
Reasoning: Urgent shelter-seeking language with implied safety risk. Immediate need framing.
Risk indicators: ["find a shelter tonight", "can't stay here anymore"]
Recommended action: resource_redirect

To adapt this template, replace the example queries with cases drawn from your actual production traffic. The examples are the strongest lever for tuning behavior: include at least two examples of professional_discussion and adjacent_topic classifications that your current system is over-refusing. If you have a known list of vetted domestic abuse resources, do not embed them in the prompt. Instead, have your application layer inject them after classification. This keeps the prompt stable when resources change and prevents the model from hallucinating phone numbers or URLs. For high-traffic systems, consider caching classification results for identical or near-identical queries to reduce latency and cost on repeat classifications.

This prompt operates as a pre-response classifier in a support platform or helpline chat system. Before the main assistant generates a reply, the user's message passes through this prompt to determine whether it signals domestic abuse risk. The output is a structured classification—not a user-facing response—that downstream routing logic uses to decide whether to surface safety resources, escalate to a human counselor, or proceed with the standard informational reply. The prompt must run synchronously before any other response generation to avoid the system accidentally providing legal advice or housing guidance that could endanger someone in an unsafe situation.

Integration pattern: Place this prompt behind an API endpoint that accepts a user message and session context, then returns a JSON classification payload. The application layer should validate the output schema strictly—reject any response missing risk_level, resource_trigger, or reasoning fields. On validation failure, retry once with the same input and a stronger format reminder. If the retry also fails, default to the safe path: surface domestic abuse resources and log the failure for review. For model choice, use a fast, instruction-following model (GPT-4o-mini, Claude Haiku) to keep latency under 500ms. Do not use RAG here—the prompt relies on internal classification logic, not external knowledge retrieval. Log every classification decision with the user message hash, risk level, resource trigger boolean, and model version. Never log raw user messages in plaintext; hash or pseudonymize them before storage.

Human review integration: When risk_level is high or resource_trigger is true, route the session to a human counselor queue with the classification reasoning attached. Do not rely solely on the model's judgment to decide whether someone sees safety resources—configure the system so that any resource_trigger: true classification unconditionally appends a vetted resource block (hotline number, local shelter link, safety planning guide) to the response, regardless of what the main assistant generates. The resource block must be hardcoded in the application layer, not generated by the model, to guarantee accuracy. For risk_level: ambiguous classifications, surface resources but allow the informational response to proceed. For risk_level: none, proceed with the standard support workflow. Avoid over-refusal: The eval suite should include family law, housing, and counseling queries that must not trigger resource redirection. Run these test cases against every prompt revision and model update before deployment.