ChatGPT Contract Review Prompt: A 5-Part Template That Works

Lawyers ask us all the time whether ChatGPT can review a contract. Yes, for the right kind of contract, and only if you write the prompt the way you would brief a paralegal. The difference between a useful first pass and a wall of generic bullet points is the prompt itself: role, law, task, context, format. This piece gives you a copy-paste template, a short list of documents you should never send to a public chatbot, and a verification pass so you catch what the model quietly gets wrong.

Why "review this contract" doesn't work

Type "review this contract and flag the risks" into ChatGPT and you get a checklist a first-year associate could write from memory: indemnity is broad, termination is one-sided, IP ownership is unclear. Nothing wrong with it, but nothing you couldn't produce in five minutes yourself, and no numbers, sections, or quoted language you can drop into a redline.

The reason is boring. The model has no idea who you represent, which law governs, what your playbook says, or how the answer should look. It fills the gap with averages. A structured prompt closes the gap. As of September 2026, most legal AI writeups (from Spellbook to LegalOn to the ABA's own guidance) converge on the same recipe.

Our full breakdown of the four elements of a good work prompt is in How to Write Better ChatGPT Prompts for Work. For contract review specifically, we add a fifth element: law.

The five-part prompt: role, law, task, context, format

Role. Tell the model who it is. "Senior in-house counsel at a B2B SaaS company" produces a different review than "outside litigation partner." A role anchors the vocabulary and the level of caution.

Law. Name the jurisdiction and governing law. New York versus Delaware versus England & Wales changes what a reasonable liability cap looks like, whether limitation-of-liability carve-outs are enforceable, and what boilerplate you expect. Without this line, the model averages across every common-law and civil-law regime it has ever read.

Task. Say what you actually want. "Flag risks" is not a task. "Return a table of provisions that create risk for the customer, with the exact clause number, quoted language, and one suggested redline per row" is.

Context. Give it what it cannot guess: the deal type, the party you represent, the counterparty's bargaining power, your risk appetite, your playbook's non-negotiables. Two or three sentences is enough, and the difference in output is not subtle.

Format. Specify the shape of the answer. A markdown table with columns for Section, Quoted Language, Risk, and Redline works for most reviews. A short executive summary followed by tracked-changes-style edits works for a partner review. For long agreements, run a summary pass first; the method is in How to Summarize a Long PDF With AI Without Losing What Matters.

A copy-paste template you can use today

Paste this into ChatGPT (or Claude, if you need a bigger context window for a long agreement), replace the bracketed lines, and attach the contract only if the document is safe to share (see the next section).

ROLE
You are senior in-house counsel at a [B2B SaaS company / manufacturer / marketing agency].
You represent the [CUSTOMER / VENDOR / LICENSEE].

LAW
Governing law: [New York / Delaware / England & Wales].
Assume standard [common-law commercial contract] rules unless the contract says otherwise.

TASK
Review the attached [Master Services Agreement / NDA / DPA / SLA] and return:
1. A table of every provision that creates material risk for my client, with:
   - Section number
   - Exact quoted language (no paraphrase)
   - Why it is a risk (one sentence)
   - Suggested redline (concrete words, not "consider revising")
2. A list of standard protections that appear to be MISSING for my side
   (limitation of liability, mutual indemnity, notice, termination for convenience,
    IP ownership, confidentiality, data protection, assignment, dispute resolution).
3. A short executive summary: approve, negotiate, or reject, with the top 3 reasons.

CONTEXT
Deal type: [one-year pilot / three-year renewal / one-off SOW].
Counterparty bargaining power: [high / balanced / low].
My playbook non-negotiables: [mutual LoL capped at 12 months fees;
  carve-outs only for confidentiality, IP infringement, and gross negligence;
  no unilateral change of terms; 30-day termination for convenience].
My risk appetite: [conservative — I would rather lose the deal than accept unlimited liability].

FORMAT
Markdown table for item 1. Bulleted list for item 2. Two short paragraphs for item 3.
If you are unsure whether a clause is present, say so; do not guess.
Do not invent section numbers or quote language that is not in the document.

Run it once, read the output, then push back. "You quoted section 8.3 as 'unlimited liability'. Paste the full section 8.3 verbatim and re-check." That second turn catches most of the invented quotes.

For a second angle on the same document, flip the perspective:

Now re-read the contract as counsel for the OTHER SIDE.
Return the top 5 provisions they would push back on if I sent them my redline above.
For each: the argument they would make, and the fallback position I should prepare.

Two prompts, one document, about twenty minutes. That is the shape of a useful first pass.

What you should never paste into a public chatbot

Confidentiality is the one mistake you cannot undo. As of September 2026, the free, Plus, and Pro tiers of ChatGPT can use your inputs for training unless you turn that setting off; ChatGPT Business ($20 per seat per month on annual billing, per OpenAI's April 2026 pricing) and ChatGPT Enterprise (custom contract) do not, and both offer zero data retention on qualifying terms. Consumer-tier default retention is 30 days, and a court preservation order in NYT v. OpenAI required OpenAI to hold consumer conversations from mid-May to September 26, 2025, and the data preserved in that window is still held.

None of that changes your duty to the client. The ABA's Formal Opinion 512 (July 2024) folded generative AI into the existing Rules 1.1 (competence) and 1.6 (confidentiality): if the tool sees the document, your client's information has been disclosed to a third party.

Do not paste, upload, or link the following into a consumer-tier chat:

Category	Examples	Why it's a problem
Client-identifiable contracts	Signed MSAs, engagement letters, deal drafts with party names	Third-party disclosure of confidential information
NDAs where the counterparty is named	Any NDA that itself names the counterparty	The NDA usually forbids disclosure to unnamed third parties, which includes AI vendors
Personal data	Names, addresses, national IDs, salary, health data	GDPR and CCPA obligations sit with you, not the vendor
Trade secrets	Pricing models, source code, unreleased financials	Trade-secret protection requires "reasonable measures", and pasting into a public tool weakens the claim
Privileged material	Litigation strategy, internal legal memos	Waiver risk

If you must use a consumer tier, redact first: replace party names with [CUSTOMER] and [VENDOR], strip dollar figures, remove dates that identify the deal. For anything sensitive, work inside a Business or Enterprise workspace your firm has approved, or a legal-specific tool with the right data terms.

Verify every finding against the original

A confident answer that quotes a clause that does not exist is the classic ChatGPT failure mode. Mata v. Avianca (2023, sanctions for fake citations) is the case every US lawyer knows. Contract review has the same failure, quieter: invented section numbers, paraphrased "quotes" that soften a harsh clause, missing provisions the model claims are present.

A six-step pass, every time:

• Open the original contract next to the output.
• For every quoted clause, Ctrl+F the first six words in the PDF. If they are not there verbatim, delete the finding.
• For every "missing" clause the model flagged, search the document for the concept (indemn, terminat, liabil, govern law). If it is present, the model missed it.
• For every suggested redline, check the surrounding sentence still reads as a contract.
• Re-run the prompt with the instruction "quote the section number and full text before commenting" and compare to the first run.
• For anything legal-strategic, apply your own judgment. The model does not know your relationship with the counterparty.

Where this breaks

The prompt above is built for a first pass on standard commercial agreements: NDAs, MSAs, SLAs, DPAs, SOWs, standard employment agreements. It will help you triage faster and produce a redline you can defend.

It will not replace judgment on:

• Regulated agreements (financial services, healthcare, insurance, public sector) where jurisdiction-specific rules dominate.
• Novel or heavily negotiated commercial deals where the risk lives in the interaction between clauses.
• Anything the client will sign without your review.

And it will not fix a bad choice of tool. For long contracts (anything past roughly 50 pages), Claude's larger context window handles the full document better than a chunked ChatGPT session; we compared both across six office tasks in ChatGPT vs Claude vs Gemini for Work.

Take this further in the practicum

Contract review is one of the pieces we run live in AI system in two weeks. Session 3 is documents, data, and visuals: the five-part prompt on your own contracts, the redaction workflow for consumer tiers, the verify-against-the-original pass above, and what you should never upload from your firm's files. You leave with a working setup, not notes. Our page for legal professionals is at /for/lawyers; the full program and next cohorts are at /#program.