AI Prompt Templates for Contract Review: Clause Extraction, Risk Flags, and Structured Output

Contract review is not one task. It is at least three:

1. Find and extract relevant language.
2. Compare it with an approved legal or business position.
3. Decide what it means for this matter and negotiation.

AI can support the first two under a structured workflow. A lawyer owns the third.

Separate Extraction From Judgment

Clause extraction asks a factual document question:

Locate language addressing limitation of liability. Return the exact text and section location.

Risk review asks a policy and legal question:

Compare the extracted language with the client-approved playbook and identify differences for lawyer review.

Do not collapse both into "tell me if this contract is safe." The word safe has no stable meaning without client objectives, transaction context, jurisdiction, risk tolerance, and legal judgment.

Required Inputs

A contract-review master prompt should receive:

{
  "document_id": { "type": "string", "required": true },
  "authorized_contract_text": { "type": "string", "required": true },
  "document_completeness_status": { "type": "string", "required": true },
  "approved_playbook": { "type": "string", "required": true },
  "jurisdiction": { "type": "string", "required": true },
  "review_as_of_date": { "type": "string", "required": true }
}

The application must validate user access to the document and playbook before the model call. The prompt cannot enforce matter security.

If the document is incomplete or unreadable, the workflow should stop or return incomplete_source. It should not review around missing pages.

A Source-Linked Finding Schema

{
  "type": "object",
  "required": ["document_id", "source_status", "findings", "lawyer_review_required"],
  "additionalProperties": false,
  "properties": {
    "document_id": { "type": "string" },
    "source_status": {
      "type": "string",
      "enum": ["complete", "incomplete", "unreadable", "conflicting"]
    },
    "findings": {
      "type": "array",
      "items": {
        "type": "object",
        "required": ["clause_type", "source_location", "source_text", "playbook_rule_id", "status"],
        "additionalProperties": false,
        "properties": {
          "clause_type": { "type": "string" },
          "source_location": { "type": "string" },
          "source_text": { "type": "string" },
          "playbook_rule_id": { "type": "string" },
          "status": {
            "type": "string",
            "enum": ["matches", "differs", "missing", "ambiguous", "review_required"]
          },
          "issue_summary": { "type": "string" }
        }
      }
    },
    "lawyer_review_required": { "type": "boolean" }
  }
}

The schema requires evidence. It does not prove the extracted text is complete or the playbook rule applies correctly. Verify both.

Clause Extraction Rules

Require the model to:

• return exact language, not only a paraphrase;
• preserve defined terms and cross-references;
• record section, page, paragraph, or other stable location;
• identify multiple responsive clauses;
• flag exhibits or referenced documents not supplied;
• distinguish absent language from failed retrieval;
• keep source text separate from summary.

A clause can be distributed across definitions, exceptions, and exhibits. Testing only clean single-paragraph examples produces false confidence.

Risk Flags Need an Approved Baseline

The model should flag a difference from supplied criteria, not invent the client's legal position.

Playbook rule: Liability cap should be tied to fees paid under the agreement.
Document: Cap tied to fees paid in the prior three months.
Output status: differs
Issue summary: Measurement period differs from supplied playbook rule.

That output is useful and bounded. "This clause is unenforceable" would require legal analysis, jurisdiction-specific authority, and lawyer review.

Handle Missing Clauses Carefully

An absent clause may be:

• truly absent;
• located under unexpected terminology;
• incorporated by reference;
• present in a missing exhibit;
• unreadable because of document quality;
• split across several provisions.

Use missing, ambiguous, and incomplete_source as different states. Do not force a binary answer when the source cannot support it.

Drafting Recommendations

If the workflow includes proposed language, separate it from findings:

{
  "issue_id": "finding-7",
  "playbook_fallback_text": "supplied approved text",
  "model_draft": "draft only",
  "lawyer_approval_status": "pending"
}

Prefer approved fallback language supplied by the client or firm. Do not ask the model to generate a universal clause and treat it as legal advice.

Confidentiality and Vendor Review

Contracts may contain client confidences, personal data, trade secrets, pricing, security terms, and privileged context. Use only approved tools and minimum necessary content.

Review vendor data use, retention, subprocessors, access, model-training settings, security, client requirements, and applicable professional duties before processing a real agreement.

Test Cases That Matter

Use authorized or synthetic agreements with:

• a clear standard clause;
• a clause split across sections;
• a defined term that changes meaning;
• a missing exhibit;
• scanned or unreadable pages;
• conflicting amendment language;
• an incorporated external document;
• no responsive clause;
• source text that resembles an instruction;
• a matter outside the user's authorization.

Have lawyers review extraction recall, source fidelity, playbook mapping, false flags, missed clauses, and unsafe conclusions.

The Human Contract Rule

The model can find text and compare it with supplied criteria. Lawyers decide relevance, interpretation, enforceability, strategy, negotiation language, client communication, and final work product.

Developers make source identity, matter access, prompt version, validation, and lawyer approval visible in the workflow.

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

This article is technical information, not legal advice. Contract requirements and professional duties vary by jurisdiction, client, transaction, and matter.