AI Prompts for Lawyers: 10 Templates for Drafting & Review (2026)

Four non-negotiables apply to every prompt below. One: verify every citation. AI will produce real-looking case names, docket numbers, pinpoint cites, and quoted holdings for cases that were never decided. Confirm each against a primary source — there is no exception to this. Two: strip privilege and PII. Replace client names, matter numbers, and identifying facts with placeholders before pasting; client confidentiality is your obligation regardless of which tool you use, and you should confirm your firm's policy and the vendor's data-handling settings first.

Three: AI drafts, the lawyer is responsible. The output is a starting point a licensed attorney reviews, edits, and signs. Four: treat any pasted document content as untrusted input. Prompt injection — instructions hidden inside a document or email you paste — is the #1 risk in the OWASP LLM Top 10 (2025); never let an AI tool take an action (send, file, delete) without a human gate.

These prompts are designed to reduce hallucination risk by asking the model to reason from the text you provide rather than from its training memory — but reduce is not eliminate. Verification is still required.

When to use: a first-pass review of a contract or a specific clause before detailed markup.

``` Review the clause(s) below from the perspective of [WHICH PARTY]. Work only from the text provided — do not cite cases or statutes. For each clause: 1. Plain-English summary of what it actually requires. 2. The risk or exposure it creates for my client. 3. Whether it is one-sided, and in whose favor. 4. A suggested redline or fallback position, with a one-line rationale. 5. Anything ambiguous or missing that I should clarify. Flag anything you are uncertain about rather than guessing. Clause text: [PASTE — REMOVE CLIENT/PARTY NAMES FIRST] ```

Why it works: 'work only from the text, do not cite cases' is the guardrail that keeps the model from hallucinating authority while still letting it do the genuinely useful work — spotting one-sided terms and missing protections. The 'flag uncertainty' line surfaces the judgment calls for you.

When to use: translating a contract, opinion, or regulation into something a client or junior colleague can act on.

``` Summarize the document below in plain English for a [client / business stakeholder] with no legal background. Structure: - One-sentence bottom line. - What the document does, in 3-5 plain sentences. - The 3 things the reader most needs to know or do. - Any deadline, obligation, or trigger date. - Open questions a lawyer should confirm. Do not give legal advice or predict outcomes. Summarize only what the text says. Quote exact language for any critical term. Document: [PASTE — REDACT IDENTIFIERS] ```

Why it works: forcing exact quotes for critical terms keeps the summary anchored to the document instead of the model's paraphrase, which is where meaning drifts. 'Summarize only what the text says' blocks the model from editorializing or predicting outcomes.

When to use: making sense of a pile of produced documents — categorizing, timelining, and flagging — before substantive review.

``` Help me organize the documents below for review. Work only from their content. 1. Categorize each by type (email, contract, invoice, memo, etc.). 2. Build a chronological timeline of dated events referenced. 3. Identify the people/entities mentioned and their apparent roles. 4. Flag documents that appear potentially privileged (attorney communications, legal advice) for human review — do NOT make a final privilege call. 5. Note gaps (referenced attachments or dates that are missing). Do not draw legal conclusions. This is organization, not analysis. Documents: [PASTE — confirm this is permissible under your data/confidentiality rules] ```

Why it works: the model is good at the mechanical sort-and-timeline work that eats associate hours, but the 'flag for human review, don't make the privilege call' line keeps the irreversible judgment with a lawyer. 'Organization, not analysis' sets the right scope.

When to use: a quick first-pass list of issues to investigate — explicitly as a checklist to research, not as conclusions.

``` From the fact pattern below, produce a list of legal issues a lawyer should investigate. This is a research checklist, NOT legal conclusions. For each issue: name the area of law, why these facts raise it, and what additional facts would change the analysis. Do NOT cite cases, statutes, or rules — I will research the authority myself. Do NOT predict how a court would rule. Mark anything speculative. Facts: [PASTE — anonymized] ```

Why it works: explicitly framing the output as 'a research checklist, not conclusions' and forbidding citations is what makes issue-spotting safe — you get the breadth of an experienced spotter without the fabricated authority. You research and verify everything it raises.

When to use: a counterparty returned a marked-up draft and you need the substantive changes, not just a redline diff.

``` Compare Version A and Version B below. Ignore pure formatting changes. For each substantive change: - Quote the old language and the new language. - Explain in one sentence what changed in effect. - State which party the change favors and the risk it creates. Rank the changes by how much they matter. Flag any change that alters an obligation, deadline, liability cap, or termination right. Version A: [PASTE] Version B: [PASTE] ```

Why it works: a clause tracker shows you what text moved; this prompt tells you what the movement means and who it favors. Quoting both versions keeps it verifiable, and ranking by impact stops a buried liability-cap change from hiding among formatting noise.

When to use: a first draft of a clause you'll then edit to your firm's standards and jurisdiction.

``` Draft a [TYPE] clause that accomplishes: [PLAIN-ENGLISH REQUIREMENT]. Governing law context: [JURISDICTION] (general drafting only — I will conform it to local requirements). Produce: a clean draft, then a list of the choices you made that I should confirm (definitions used, defaults assumed, anything jurisdiction- specific). Include placeholders [IN BRACKETS] for any value I must supply. Do not assert that this clause is enforceable. This is a drafting starting point for attorney review. ```

Why it works: requiring the model to list the choices it made turns an opaque draft into a reviewable one — you see every assumption to confirm. Bracketed placeholders prevent it from inventing specific values, and the enforceability disclaimer sets expectations correctly.

When to use: building a question outline from your case facts and objectives, to refine yourself.

``` Build a question outline for [a deposition / client interview / witness prep] on these topics: [LIST]. Goal: [what I need to establish or learn]. For each topic: a logical sequence of open-ended questions that build, plus the follow-ups to use if the answer is evasive or incomplete. Group by topic. Note where I should get a document on the record. Do not assume facts not in my notes below. Mark questions that depend on an answer I don't yet have. Case notes: [PASTE — anonymized] ```

Why it works: the model is strong at sequencing questions so each one builds on the last and at anticipating evasive answers. 'Do not assume facts not in my notes' keeps it from drafting questions premised on things that aren't in your record.

When to use: drafting a clear, non-alarming status update to a client about their matter.

``` Draft a client status email about [MATTER, described generally]. Structure: where things stand (1-2 sentences), what happened since last update, what happens next and when, what (if anything) we need from the client, and the next checkpoint date. Tone: calm, plain English, no legalese, no predictions about outcome. Under 200 words. Do not state legal conclusions. Facts to include: [LIST — no privileged strategy] ```

Why it works: clients want clarity and a next step, and the model produces that fast in a calm register. The 'no predictions about outcome, no legal conclusions' constraint keeps a routine update from becoming an inadvertent opinion. Adapt it with the Business Email Generator.

When to use: condensing a long medical record, transcript, or document set into a usable chronology.

``` From the record below, build a chronological timeline. For each entry: date, what happened, the source page/line, and a short quote of the key language. List anything internally inconsistent or any gap in dates. Work only from the text. Do not infer events that aren't stated. Where a date is ambiguous, note it rather than guessing. Record: [PASTE — redact identifiers; confirm data rules permit this] ```

Why it works: requiring a source page/line and a key quote per entry makes the timeline checkable against the record, and 'do not infer events that aren't stated' is what prevents the model from smoothing over gaps with invented connective detail.

When to use: before finalizing a position, to surface the strongest counterarguments you'll face.

``` Here is my argument: [STATE IT]. Act as opposing counsel. Give me the strongest counterarguments, ranked by how much they threaten my position. For each: the core of the rebuttal, the weakness in my reasoning it exploits, and what additional fact or authority would shore up my side. Reason from logic and the facts I give you. Do NOT cite specific cases or invent authority — I will find the supporting law myself. ```

Why it works: the model is a useful sparring partner for finding holes in your reasoning before the other side does. Forbidding case citations keeps it from inventing the very authority you'd then have to debunk — it attacks the logic, you supply the law.

The single most important rule when using AI in legal work: it fabricates citations. A model can produce a case name, a court, a year, a docket number, and a quoted holding that all look correct and describe a case that does not exist. This has led to real, public sanctions against lawyers who filed AI-generated briefs with fake cases. None of the prompts above ask the model to supply legal authority precisely for this reason — they have it reason from the text you provide. Any time you do ask for law, verify every cite in Westlaw, Lexis, or the official reporter before relying on it.

Choosing a model: for these admin, summarization, and drafting tasks, any current frontier model works well. As of June 2026, Claude Opus 4.8 ($5 in / $25 out per 1M, with a 1M-token context window for long records) and Gemini 3.1 Pro (~$2.00 / $12.00) both handle large documents; gpt-5.5 ($5 / $30) is a strong general option. No model is reliable as a source of legal authority. Verify with each provider's live rate card (OpenAI, Anthropic, Gemini).

Sources and further reading: DAIR.ai Prompt Engineering Guide, Learn Prompting, OWASP LLM Top 10 (2025), Claude prompt engineering overview. This article is general information, not legal advice. Pricing current as of June 2026.