Legal AI Prompts: Templates for Research and Draft Review

A legal prompt is an instruction that gives an AI assistant a defined legal-work task, the material it may use and the checks a lawyer must perform. The strongest templates separate supplied facts, legal authority, assumptions and questions that remain unanswered.

Use the examples below to prepare work for review, rather than treating the response as advice ready for a client. Tools like fenbs can keep reusable prompts and firm-approved rules outside a single chat, so another connected assistant starts with the same instructions.

What belongs in a legal-work prompt?

State the jurisdiction, relevant date, procedural setting and intended reader before describing the output. A memo for supervising counsel needs different detail from a plain-language explanation prepared for a client. If a fact is unknown, label it unknown rather than inviting the assistant to supply a likely answer.

• Assignment: the narrow question and the purpose of the work.
• Sources: the approved excerpts or documents, each with an identifier and version.
• Boundaries: what the assistant must not infer, decide or send.
• Output: headings, length, source references and a separate uncertainty list.
• Review: the person who will check authority, reasoning and suitability.

A role instruction such as “act as an attorney” does not supply expertise or a review process. The more useful instruction describes observable behavior: identify the clause, quote its relevant words, explain the concern and state what evidence is missing. Our system prompt examples show how to separate recurring instructions from each assignment.

How can a prompt prepare a research assignment?

Start with a research plan when you have not yet collected reliable authority. Ask for questions, search terms and factual dependencies. Do not ask an assistant without verified sources to produce a finished list of controlling cases and then mistake that list for completed research.

Prepare a research plan, not a legal conclusion.
Question: [narrow issue]
Jurisdiction and relevant date: [scope]
Procedural setting: [setting]
Approved facts: [sanitized facts]
Supplied authority: [source IDs or none]
Return: issues to research, useful search terms, missing facts, and authorities to locate.
Do not invent cases, quotations or citations. Label unverified leads as leads.
Separate supplied authority from suggestions that require independent research.

For example, a hypothetical assignment about notice under a services agreement might depend on the contract language, how notice was delivered and when it arrived. A good response surfaces those dependencies. An answer that selects a jurisdiction or fills in delivery dates has exceeded the assignment, even if the resulting discussion sounds persuasive.

Use AI for legal research for the larger source-verification workflow and AI for legal documents for document-handling choices; these templates address the instructions you give the assistant.

Which prompt helps review a clause?

Give the assistant the actual clause and the approved commercial position. Without both, it may replace a negotiated choice with generic language. Ask for questions and differences before requesting a rewrite, and keep the original wording available for comparison.

Review only the supplied clause against this approved position.
Clause and document version: [text and version]
Approved position: [human-approved requirements]
Relevant definitions and cross-references: [excerpts]
Return a table: exact wording, potential mismatch, explanation, question for counsel.
Identify missing definitions. Do not assume another clause fixes a gap.
Offer optional wording separately, with each substantive change explained.
Do not state enforceability or approve the clause.

Check whether the suggested edit changes obligations, exceptions, timing or who bears a risk. A shorter sentence is not necessarily an equivalent sentence. Have the assistant list each changed obligation separately so the lawyer can review the business effect before polishing the prose.

Keep recurring review requirements in fenbs Decisions and rules, so the next assistant can read the same approved position before proposing edits.

How do you ask for a useful draft check?

Separate factual consistency, source support and readability into distinct passes. A broad request to “improve this” can silently change a qualification or turn an unresolved allegation into an established fact. Name the details that must survive: defined terms, dates, party labels, reservations and the requested action.

Check this draft against the approved source packet.
Draft: [sanitized text]
Packet: [source IDs and excerpts]
Preserve defined terms, qualifications and the distinction between allegation and fact.
List unsupported claims, inconsistent dates, missing citations and ambiguous references.
For each issue, identify the passage and the supporting source or say not found.
Suggest edits separately. Do not invent support or change the legal position.
End with questions for the reviewing attorney.

Require a location for every finding. “The dates conflict” is less useful than identifying the two passages and explaining their difference. When a draft has no source for a statement, the correct outcome is an unresolved question or deletion for review, not a newly generated citation.

Which checks remain the lawyer’s responsibility?

The ABA’s Formal Opinion 512↗ (opens in a new tab) discusses competence, confidentiality and review when lawyers use generative AI. It is guidance under the Model Rules, not a substitute for the rules that apply to your practice. Confirm local professional obligations and any relevant court requirements.

Independently open every authority relied on. Check the proposition against the text, confirm the quotation and pinpoint reference, and use an appropriate legal research service to check subsequent treatment. An assistant reporting “verified” is not evidence that these steps occurred, particularly if it could not access the source.

The ABA’s Model Rule 3.3↗ (opens in a new tab) addresses candor to tribunals, including knowingly false statements of fact or law. Treat source and factual review as a substantive part of preparing a filing. A prompt requiring accuracy cannot discharge a lawyer’s professional duties.

How do you protect information and test a template?

Use only systems and data handling your firm has approved. Review the intended use, access and retention of information before entering client material. Removing a name may still leave a matter identifiable through unusual facts, dates or transaction details. Reusable prompt libraries should contain instructions and fictional examples rather than live client narratives.

Test a template with a small fictional packet containing a missing definition, a contradictory date and an unsupported proposition. Define the expected findings before running it. Then check whether the assistant identifies the problems without inventing a fix. Repeat with a clean packet to see whether it raises unsupported objections.

Save the prompt version, test material, observed errors and reviewer feedback. Change one instruction at a time so you can see which change helped. The wider AI prompts collection offers templates for other work, while project-management prompts address the follow-up tasks a review creates.

Keep the reusable version in fenbs Skills and prompts, with the source requirements and review limits beside it. A later assistant can read those instructions without needing the confidential material from the assignment that exposed the problem.

How to do this with fenbs

1. Create a free account, free for one person with no credit card, and add a project for your prompt library.
2. Save a sanitized template in Skills and prompts. Include its purpose, required sources, output and review checks.
3. Record the firm’s human-approved instructions in Decisions and rules, including what information must stay in approved client systems.
4. Connect a supported assistant and ask it to read the project context before using the template. Keep the attorney’s final review separate from the generated draft.
5. Record lessons from fictional tests and save progress with the next review step, so another session can continue without guessing.

fenbs does not verify case law, provide legal advice or replace a matter-management system. Keep client records and privileged work in systems your firm has approved; use the shared prompt library for reusable instructions that are appropriate to store there.