Master Prompts for Legal Teams: Contract Review, Research, and Compliance Documentation

Legal AI should make review more structured without making responsibility less clear.

A legal master prompt can extract clauses, organize supplied authorities, compare a document with an approved policy, or produce a source-linked draft. It cannot establish that a legal conclusion is correct, that a citation remains good law, or that confidential information was safe to send to a vendor.

Those decisions remain with lawyers and the organizations responsible for the matter.

Professional Duties Still Apply

The American Bar Association's Formal Opinion 512 says lawyers using generative AI must consider duties including competence, confidentiality, communication, supervision, meritorious claims, candor to tribunals, and reasonable fees. It also emphasizes appropriate independent verification and review of AI output. See ABA Formal Opinion 512.

Applicable duties and disclosure rules vary by jurisdiction, court, client agreement, and matter. A master prompt should receive the approved local rules and firm policy; it should not invent them from general model knowledge.

Good Legal Master-Prompt Workflows

Contract Clause Extraction

Extract supplied clauses into a defined schema with section location, exact source text, normalized category, and review status.

The model can identify candidate clauses. A lawyer determines meaning, risk, enforceability, and negotiation position.

Contract Review Against an Approved Playbook

Compare a supplied agreement with a client-approved review policy:

{
  "clause_type": "limitation_of_liability",
  "source_location": "section-reference",
  "source_text": "authorized excerpt",
  "playbook_rule_id": "approved-rule-id",
  "status": "review_required",
  "issue_summary": "",
  "lawyer_decision": null
}

Do not ask the model to create the organization's risk policy while applying it.

Read AI Prompt Templates for Contract Review for the full pattern.

Legal Research Organization

Use retrieval from authorized sources, preserve citation identity, distinguish quoted or summarized propositions, and require a verification state.

The output should never present an unverified model-generated citation as authority.

Read Master Prompts for Legal Research for case and citation controls.

Compliance Evidence Documentation

Organize supplied policies, controls, evidence, findings, owners, and review state. Keep legal conclusion and sign-off outside the model.

Read Legal Compliance Documentation With AI for the evidence contract.

Drafting Support

Generate a draft from attorney-approved facts, authorities, position, template, and audience. Require source references and mark every output as draft.

The lawyer reviews accuracy, strategy, tone, privilege, confidentiality, citations, and filing or communication requirements.

Legal Variables Need Matter Context

Useful variables may include:

{
  "matter_id": { "type": "string", "required": true },
  "authorized_source_text": { "type": "string", "required": true },
  "jurisdiction": { "type": "string", "required": true },
  "as_of_date": { "type": "string", "required": true },
  "approved_review_policy": { "type": "string", "required": true }
}

The application must verify that the user and vendor may access the matter data. A matter_id in the prompt does not enforce an ethical wall or client boundary.

Source Provenance Is Required

For every substantive finding, store:

• source document ID;
• page, section, paragraph, or other location;
• exact excerpt where appropriate;
• authority identifier and retrieval source;
• date retrieved or verified;
• prompt version;
• human review state.

Schema validation can require these fields. Application and lawyer review must verify that they are real and support the stated proposition.

Confidentiality Is a System Control

Before using any legal data, review:

• vendor terms and data use;
• retention and deletion;
• model training settings;
• access and authentication;
• matter and tenant isolation;
• subprocessors;
• logging and support access;
• client consent or disclosure requirements where applicable;
• firm policy and jurisdiction-specific duties.

Do not put client information into an unapproved tool because the prompt says "keep this confidential."

Use the minimum authorized data for the task. Legal confidentiality can cover more than material that is formally privileged.

The Output Needs a Human State

{
  "draft": true,
  "issues": [],
  "sources": [],
  "unverified_citations": [],
  "missing_information": [],
  "lawyer_review_required": true,
  "approved_for_use": false
}

Only authorized application actions should change approved_for_use. The model cannot approve its own legal work.

Test Legal Failure Cases

Use authorized or synthetic matters that include:

• a standard clause;
• a nonstandard clause with similar wording;
• missing governing-law information;
• conflicting authorities;
• an outdated or overruled authority in the retrieval set;
• a fabricated citation;
• incomplete document pages;
• source text that attempts to change instructions;
• confidential data from the wrong matter;
• a filing or communication with a mandatory human review step.

The system must stop cross-matter access and unverified authority before the model result reaches a client, counterparty, agency, or court.

The Human Legal Rule

Lawyers own legal judgment, client communication, confidentiality, supervision, strategy, citation verification, and final work product. Legal operations and knowledge teams own approved sources and workflows. Developers own authorization, source identity, validation, observability, and side-effect controls.

The master prompt makes the model's task reviewable. It does not make the model a lawyer.

Browse legal workflow contracts in the CyWire marketplace and use the production-ready checklist before using client or matter data.

This article is technical information, not legal advice. Consult qualified counsel and applicable professional rules for the specific jurisdiction and matter.