Master Prompts for Legal Research: Case Summarization, Precedent Analysis, and Citation Output

Legal research AI fails most dangerously when a citation looks finished.

A correctly formatted case name, reporter citation, quotation, and parenthetical can still be fabricated, misquoted, irrelevant, from the wrong jurisdiction, or no longer good authority for the proposition.

A legal research master prompt should make verification a required state, not a sentence in the disclaimer.

Retrieval Before Analysis

Do not ask a general model to supply authorities from memory.

Use an authorized legal research source or controlled retrieval system to provide candidate authorities and stable source identifiers. The prompt then summarizes and organizes only the retrieved material.

authorized research query
  -> jurisdiction and date scope
  -> controlled authority retrieval
  -> source IDs and document text
  -> versioned master prompt
  -> structured research object
  -> independent lawyer verification

Read Master Prompt vs. RAG for the retrieval contract.

Required Research Inputs

{
  "research_question": { "type": "string", "required": true },
  "jurisdiction": { "type": "string", "required": true },
  "court_scope": { "type": "string", "required": true },
  "as_of_date": { "type": "string", "required": true },
  "authorized_authorities": { "type": "string", "required": true },
  "lawyer_instructions": { "type": "string", "required": true }
}

If retrieved authority is empty, return insufficient_authority. Do not fill the gap from model memory.

A Citation-First Output Schema

{
  "type": "object",
  "required": ["research_scope", "authorities", "gaps", "lawyer_review_required"],
  "additionalProperties": false,
  "properties": {
    "research_scope": { "type": "string" },
    "authorities": {
      "type": "array",
      "items": {
        "type": "object",
        "required": ["source_id", "citation_as_retrieved", "proposition", "support_location", "verification_status"],
        "additionalProperties": false,
        "properties": {
          "source_id": { "type": "string" },
          "citation_as_retrieved": { "type": "string" },
          "proposition": { "type": "string" },
          "support_location": { "type": "string" },
          "quotation": { "type": ["string", "null"] },
          "verification_status": {
            "type": "string",
            "enum": ["unverified", "source_verified", "treatment_verified", "rejected"]
          }
        }
      }
    },
    "gaps": { "type": "array", "items": { "type": "string" } },
    "lawyer_review_required": { "type": "boolean" }
  }
}

Only the application or authorized reviewer should advance verification status. The model cannot verify its own citation by repeating it.

Case Summarization Rules

A case summary should distinguish:

• court and date from the retrieved source;
• procedural posture;
• relevant facts;
• issue addressed;
• holding;
• reasoning;
• disposition;
• quoted language and exact location;
• unresolved relevance questions;
• source and treatment verification.

Do not flatten every judicial statement into a holding. Lawyers must review whether language is controlling, persuasive, dicta, limited by posture, or relevant only under different facts.

Precedent Analysis Needs Comparison Fields

Instead of "Is this case similar?" use a comparison structure:

{
  "authority_source_id": "case-17",
  "shared_material_facts": [],
  "different_material_facts": [],
  "legal_standard_as_summarized": "",
  "procedural_differences": [],
  "jurisdiction_status": "",
  "model_relevance_note": "",
  "lawyer_analysis_required": true
}

The model identifies comparison points. A lawyer decides legal significance and how to use the authority.

Quotations Need Exact Verification

Require:

• quotation copied from supplied authority text;
• page, paragraph, or stable location;
• source ID;
• clear marking of ellipses or alterations;
• no quotation when the exact text is unavailable.

After generation, verify the quotation against the authoritative source. Schema-valid quotation fields can still contain changed words or missing context.

Good-Law and Treatment Checks Stay External

Case status and subsequent treatment change over time. Use an authoritative citator or approved research process to verify current treatment under the applicable jurisdiction and as-of date.

Do not let a master prompt infer good-law status from the case text or from the absence of negative treatment in a limited retrieval set.

The output should preserve who verified treatment, when, and through which approved source where the firm's process requires it.

Professional Responsibility

ABA Formal Opinion 512 warns that uncritical reliance on generative AI can lead to inaccurate legal advice or misleading representations and requires appropriate independent review. It addresses competence, confidentiality, supervision, meritorious claims, and candor, among other duties. See ABA Formal Opinion 512.

Courts, jurisdictions, clients, and firms may impose additional rules, disclosures, or restrictions. Verify what applies to the specific matter.

Confidentiality and Research Scope

Research questions can reveal client strategy and confidential facts even when no source document is attached.

Use approved tools, minimum necessary matter context, access controls, retention policy, and vendor review. Separate matters at retrieval and storage. Do not use one client's facts as examples for another matter without authorization.

Test the Dangerous Cases

Use authorized or synthetic research sets containing:

• a real controlling authority;
• a persuasive authority from another jurisdiction;
• a case with similar facts but different procedural posture;
• a superseded statute or rule;
• an authority with negative treatment;
• a fabricated citation in source notes;
• a real citation with a false quotation;
• an authority that does not support the proposition;
• no responsive authority;
• source text that attempts to change instructions.

The workflow fails if it creates an authority, advances verification state, hides conflicting law, or sends an unverified citation into work product.

The Human Research Rule

The retriever finds candidate sources. The master prompt structures supplied authorities. The application validates source IDs and workflow state. The lawyer verifies existence, quotation, treatment, jurisdiction, relevance, analysis, and final use.

That chain is slower than accepting fluent output and much faster than explaining a false citation to a client or court.

Read the Legal Master Prompts guide, review Legal Document Automation, and browse legal workflow contracts in the CyWire marketplace.

This article is technical information, not legal advice or a substitute for authoritative legal research and qualified counsel.