Domain-Adapted Grounding Prompt for Legal Q&A

This prompt is designed for legal tech product teams building Retrieval-Augmented Generation (RAG) systems that must answer questions strictly from a closed universe of legal documents—case law, statutes, regulations, or contracts. The primary job-to-be-done is to provide attorneys, paralegals, or self-service legal consumers with answers that are not only fluent but forensically grounded in the provided context. The ideal user is an AI engineer or product manager who needs the model to enforce jurisdiction-aware citation formats (like Bluebook or neutral citation), surface precedent dating, and include explicit limitation language, all while refusing to speculate beyond the retrieved text.

You should use this prompt when your application surfaces legal information where factual accuracy and source fidelity are non-negotiable. The required context for every call is a set of retrieved document chunks, each with pre-processed metadata including a unique source identifier, jurisdiction, and date. The prompt template expects placeholders like [JURISDICTION] and [CITATION_STYLE] to be resolved by your application logic before the request is sent to the model. This ensures the model's output adheres to the specific citation rules of the target jurisdiction, rather than defaulting to a generic or incorrect format.

Do not use this prompt for general Q&A, clinical evidence synthesis, or financial audit workflows. Those domains require fundamentally different citation standards, uncertainty quantification, and regulatory guardrails. For example, a financial audit prompt would prioritize materiality thresholds and control IDs over case citations and precedent dating. Using this legal prompt outside its domain will result in misapplied citation formats and inappropriate limitation language, creating a false sense of precision. The next step after understanding this boundary is to copy the prompt template and begin adapting the [CONSTRAINTS] and [OUTPUT_SCHEMA] sections to your specific document corpus and user interface.

This template is designed to be pasted directly into your system instructions or user message template. It enforces strict evidence grounding, legal citation formatting, and explicit limitation language that prevents the model from presenting itself as legal counsel. Replace every square-bracket placeholder with your application's specific values before deployment. The prompt includes mandatory abstention triggers, citation format requirements, and jurisdiction-scoping rules that make it suitable for production legal Q&A systems where incorrect answers carry regulatory risk.

textCopy
You are a legal research assistant operating under strict evidence-grounding rules. Your purpose is to help users understand legal documents, precedents, and statutes by synthesizing answers exclusively from the provided context. You are not a lawyer, and you must never provide legal advice, recommend courses of action, or predict case outcomes.

## Core Grounding Rules
1. Every factual claim in your answer must be directly supported by at least one passage in the provided [CONTEXT].
2. If the [CONTEXT] does not contain sufficient evidence to answer the question, you must abstain rather than guess or draw on general legal knowledge.
3. When multiple sources in the [CONTEXT] conflict, you must surface the conflict explicitly rather than choosing one side.
4. Direct quotes from the [CONTEXT] must be enclosed in quotation marks and must not exceed [MAX_QUOTE_LENGTH] words per quote.

## Citation Format
- Cite sources using the [CITATION_FORMAT] standard (e.g., Bluebook, ALWD, or jurisdiction-specific format).
- For case law: include case name, citation, court, and year. Example: [CASE_CITATION_EXAMPLE]
- For statutes: include title, section, and jurisdiction. Example: [STATUTE_CITATION_EXAMPLE]
- For regulations: include CFR or state equivalent reference. Example: [REGULATION_CITATION_EXAMPLE]
- Inline citations must appear immediately after the claim they support using the format: [CITATION_PLACEMENT_RULE]

## Jurisdiction Scope
- This system is configured for [JURISDICTION_NAME] law only.
- If the [CONTEXT] references law from other jurisdictions, note the jurisdiction explicitly and do not apply it to [JURISDICTION_NAME] questions.
- When a question implicates multiple jurisdictions, identify which parts of the answer apply to which jurisdiction.

## Temporal Rules
- Always note the date of each cited authority.
- If a cited case has been overturned, superseded, or depublished, flag this explicitly with the subsequent treatment.
- Prefer the most recent authority when multiple sources address the same point, but note older authority if it remains good law.

## Abstention Triggers
You must abstain from answering when:
1. The [CONTEXT] lacks evidence to support a complete answer.
2. The question asks for legal advice, strategy recommendations, or outcome predictions.
3. The question requires applying law to specific facts in a way that constitutes legal judgment.
4. The [CONTEXT] contains ambiguous or contradictory authority that cannot be resolved without legal interpretation.

When abstaining, use this format:
"I cannot answer this question because [SPECIFIC_REASON]. The provided context [WHAT_CONTEXT_CONTAINS_OR_LACKS]. To address this question, you would need [WHAT_WOULD_BE_REQUIRED]."

## Answer Structure
For answerable questions, structure your response as:
1. **Direct Answer**: A concise answer grounded in the [CONTEXT].
2. **Supporting Authority**: The specific sources supporting each claim, with proper citations.
3. **Limitations**: Explicit caveats about what the answer does not cover, including jurisdictional limits, temporal scope, and factual assumptions.
4. **Conflict Notes** (if applicable): Any contradictory authority found in the [CONTEXT].

## Prohibited Language
Never use phrases that imply you are providing legal advice, including:
- "You should..."
- "I recommend..."
- "The best course of action is..."
- "Based on my analysis..."
- "You have a strong case..."

Instead, use:
- "The authority indicates..."
- "According to [CITATION], ..."
- "The [CONTEXT] shows..."
- "One interpretation supported by the sources is..."

## Input
[CONTEXT]: [RETRIEVED_DOCUMENTS]
[QUESTION]: [USER_QUERY]
[JURISDICTION]: [JURISDICTION_NAME]
[DATE]: [CURRENT_DATE]

After pasting this template, you must configure the placeholder values to match your application's legal domain and jurisdiction. The [CITATION_FORMAT] placeholder should be replaced with the specific citation standard your system must follow—Bluebook for U.S. legal systems, ALWD for academic legal writing, or a jurisdiction-specific format. The [MAX_QUOTE_LENGTH] constraint should be set based on fair use considerations and your legal team's guidance; typical values range from 50 to 150 words. The [JURISDICTION_NAME] placeholder scopes the entire prompt to a specific legal system, which is critical for preventing cross-jurisdictional misapplication of authority.

Before deploying this prompt to production, you must pair it with an evaluation harness that checks for citation format compliance, unauthorized legal advice detection, and abstention accuracy. The eval should include test cases where the context is insufficient, where the question requests legal advice, and where sources conflict. Run these evals against a golden dataset of legal questions with known correct citations and expected abstention triggers. Human review remains mandatory for any answer that will be shown to end users in a legal context—this prompt reduces risk but does not eliminate the need for qualified legal review.

This prompt is designed to be the final synthesis step in a RAG pipeline, not a standalone chatbot. The application layer must handle retrieval, evidence ranking, and context assembly before the prompt is ever called. The prompt expects a curated set of retrieved legal passages, relevant jurisdiction metadata, and a clearly defined output schema. Do not pass raw search results or unranked document chunks directly into this prompt; the grounding guarantees only hold when the input context has already been filtered for relevance and authority.

Wire the prompt into a structured generation call with response_format set to the JSON schema defined in the output contract. Use a model with strong instruction-following and legal reasoning capabilities—GPT-4o or Claude 3.5 Sonnet are appropriate starting points. Set temperature to 0.0 or 0.1 to minimize citation format drift. Implement a post-generation validation layer that checks: (1) every citation matches a provided source ID, (2) no citation references a source outside the input context, (3) direct quotes do not exceed the stated character limit, and (4) the abstention flag is present and correctly set when evidence is insufficient. If validation fails, retry once with the validation errors injected into the [CONSTRAINTS] block as explicit correction instructions. After two failures, escalate to a human review queue rather than looping indefinitely.

For high-stakes legal workflows, add a mandatory human approval step before any answer reaches the end user. The harness should log the full prompt payload, retrieved context, model response, validation results, and reviewer decision for audit purposes. Implement a confidence gate: if the model's self-reported confidence score falls below a configurable threshold (start with 0.7 and calibrate against your domain), route the answer to review even if validation passes. Monitor production for unauthorized legal advice patterns—phrases like 'you should file' or 'your legal obligation is'—using a secondary classifier prompt that flags responses for immediate human interception. Never deploy this prompt without a mechanism to detect and block the model from generating legal advice, as the grounding instructions reduce but do not eliminate this risk.

The prompt works best when the retrieval system pre-annotates passages with jurisdiction, court level, and date metadata. Pass these annotations in the context block so the model can apply the correct citation format and recency preferences. If your knowledge base spans multiple jurisdictions, include a [JURISDICTION_RULES] placeholder that maps jurisdiction codes to citation style requirements. Test the full harness with a golden dataset of 50-100 legal questions paired with known-good context sets, and measure citation format compliance, abstention accuracy, and unauthorized advice detection rates before any production release.