AI Prompts for Lawyers & Contract Ops — With the Generator Built In

What this hub is

Most "AI prompts for lawyers" pages hand you a list of one-line prompts and send you on your way. The prompt is never the hard part of legal work — the hard parts are checking the output against the actual clause text, knowing what your playbook position is before you start, and not leaking privileged material into a consumer chatbot while you do it. This hub is built around those three problems.

You get three task generators, each one a complete workflow rather than a single prompt:

1. NDA risk triage — paste an NDA, add your playbook positions, and get a standard / negotiable / escalate classification with a risk table that quotes the clause behind every flag.
2. Indemnification comparison — paste two indemnity clauses (or yours versus the other side's paper) and get a side-by-side table of positions, risks, and fallback language, with a verdict on which clause favours which side and why.
3. Compliance redline checklist — run a contract against your playbook across ten clause families and get a checkable list: every item anchored to a section, marked clear or flagged, with the before → after change and a recommendation.

Every generator runs 100% client-side. The contract text you paste is assembled into a prompt in your browser and never leaves it — PromptStaging has no server receiving your documents. That is a structural answer to the confidentiality problem, not a policy promise, and it matters for legal work more than for any other profession on this site (see the Guardrails section below).

The five elements of a legal prompt

Ranking legal-prompt guides agree on the anatomy, and every template on this page follows it. A legal prompt that works names:

1. Role and side. "You are reviewing for the customer" produces a different review from "for the vendor." The same indemnity clause reads as protective or dangerous depending on which side of it you sit — a prompt that doesn't state the side gets a review written for nobody.
2. Jurisdiction. Governing law changes what is enforceable. A non-solicit period, a restraint, or an indemnity formulation that is routine in one jurisdiction can be unenforceable in another. Jurisdiction is a required field on every generator here, not an optional refinement.
3. Facts. The deal context: what is being bought or sold, the value, the parties' relative leverage, what has already been agreed. Facts you don't supply, the model cannot weigh.
4. The task. One task per prompt. "Review this contract" is not a task; "classify this NDA against these five playbook positions and quote the clause behind every deviation" is.
5. The output format. Tables with named columns, a classification with a one-line reason, a decision block. If you don't specify the format, you get prose — and prose is where unverifiable claims hide. Every template below specifies a table, because a table forces the model to commit to a clause, a quote, and a severity for every claim it makes.

Who this is for

• In-house counsel and contract ops teams running first-pass review against a playbook before attorney time is spent.
• Law firm associates and paralegals who triage inbound NDAs and vendor paper at volume and need a consistent classification standard.
• Solo practitioners and founders who receive contracts they didn't draft and need to know — quickly and defensibly — what is standard, what is negotiable, and what must go to counsel.
• Sales and procurement teams who are the first humans to see an NDA and need a go / no-go read before it enters the legal queue.

If you are not a lawyer: these tools structure a first read. They do not give legal advice, and the guardrails section on every page states what must still go to a qualified attorney.

How to use the generators

1. Pick the task — triage, comparison, or checklist — from the three generators linked below.
2. Set your playbook positions first. Your standard on mutuality, term, caps, and governing law is an input, not something the model should guess. Two minutes here is the difference between a review and a horoscope.
3. Paste the clause or contract text into the generator. It stays in your browser; copy the assembled prompt out to the AI tool of your choice.
4. Run the prompt in your AI tool and read the output against the quotes. Every risk flag on these pages is required to quote the exact clause text behind it — if a flag has no quote, treat it as unverified.
5. Act on the decision block. Each output ends with what to do next: sign, negotiate the listed points, or escalate — and what a human must verify before anything is sent.

The master template (hub default)

The generator above ships with this general contract-review template. The three task pages carry specialised versions of it.

You are a senior commercial contracts reviewer assisting [YOUR ROLE — e.g. in-house counsel] acting for the [SIDE — e.g. customer] in a [CONTRACT TYPE — e.g. master services agreement] governed by the law of [JURISDICTION].

CONTEXT
- Parties: [PARTY A] and [PARTY B]
- Deal summary: [WHAT IS BEING BOUGHT/SOLD, VALUE, TERM]
- Our playbook positions: [LIST YOUR STANDARD POSITIONS — e.g. mutual indemnities only; liability capped at 12 months of fees; governing law [STATE/COUNTRY]; no assignment without consent]

TASK
Review the contract text below against our playbook positions.

INSTRUCTIONS
1. Work clause family by clause family: parties and term, payment, obligations, warranties, indemnification, liability cap, IP and confidentiality, termination, dispute resolution, general.
2. For every deviation from a playbook position, quote the exact clause text that creates it. Do not flag a risk you cannot quote.
3. For every flagged item give: severity (High / Medium / Low), why it matters to us specifically, and fallback language we could propose.
4. Where the contract is silent on a playbook position, say "Absent from the document" — do not treat silence as agreement.
5. Where a question cannot be answered from the document, write "Not determinable from the documents provided."

NEGATIVE CONSTRAINTS
- Do not paraphrase a clause and present the paraphrase as evidence. Quote first, analyse second.
- Do not invent playbook positions, facts, or clause text that are not in what I provided.
- Do not state what the law requires as fact. Where enforceability may depend on [JURISDICTION] law, flag it as "verify under [JURISDICTION] law" instead.
- Do not soften a severity rating to be agreeable. If it deviates from the playbook, rate it.

OUTPUT FORMAT
1. Summary: 3 sentences maximum — overall posture of the document versus our playbook.
2. Risk table: | Clause / section | Quoted text | Deviation from playbook | Severity | Fallback language |
3. Absent items: playbook positions the document does not address.
4. Decision block: Recommendation (Proceed / Negotiate listed points / Escalate to counsel) + the items a human must verify before signature.

CONTRACT TEXT
[PASTE CONTRACT OR CLAUSE TEXT]

Best model: Claude Sonnet 5.5 — the strongest of the current set at clause-level nuance and instruction-following on long structured reviews. For a full agreement with exhibits that strains a context window, use Gemini 4 Argon and its 1M-context option. GPT-6.1 / GPT-4o is a sound general default.

Worked example — the master template on a liability clause

Filled inputs: Role/side: in-house counsel for the customer · Jurisdiction: England and Wales · Contract type: SaaS subscription agreement · Playbook position: liability capped at 12 months of fees, no exclusion of liability for data loss · Clause pasted:

Output excerpt (what the model returned):

Decision block (excerpt): Negotiate listed points. Two High items, both on the clause the customer cares most about; neither is a walk-away point if the fallback language is accepted.

Note what the format bought you: every flag carries the sentence that creates it. You can check both rows against the pasted clause in under a minute — that is the entire point of the quote-first rule.

Which model for which legal task

• Clause nuance, redline reasoning, fallback drafting — Claude Sonnet 5.5. First choice for triage and checklist work, where the task is reading precisely and following a multi-part instruction set.
• Very long documents — Gemini 4 Argon. A full MSA with schedules and exhibits is where context runs out; use the 1M-context option rather than splitting the document, because cross-references between clauses are exactly what a split review misses.
• Fast general review and clean structured output — GPT-6.1 / GPT-4o. A reliable default for the comparison table and for high-volume first passes.

Whichever model you use, the template's constraints — quote-first, not-determinable, no invented positions — do more for output quality than the model choice does. (Models verified Oct 2026.)

Jurisdiction is an input, not a detail

A recurring failure mode in legal AI output is the jurisdiction-free answer: confident, well-formatted, and silently written for a legal system you are not in. Every generator on this hub takes jurisdiction as a required field, and every template instructs the model to flag enforceability questions as "verify under [jurisdiction] law" rather than resolve them. Treat any AI output that states what "the law requires" without a jurisdictional caveat as unverified, whoever produced it.

How the three generators chain in a real matter

The three tasks are one workflow, in the order a matter actually unfolds. An NDA lands first: the triage generator classifies it against your playbook, and only a Standard result goes straight to signature — Negotiable returns with fallback language attached, Escalate stops the clock until counsel has looked. Once the NDA is signed and the real contract arrives, its indemnity is the clause most likely to be fought over, so it goes through the comparison generator — your standard on one side, their paper on the other, a table and a verdict in one pass. Everything else in the document — and every draft that comes back afterwards — goes through the redline checklist, rerun each round so that a payment-term edit in round three can't quietly reopen the liability row you closed in round one. Same playbook positions, entered once per matter and reused at each stage: that consistency, more than any single prompt, is what a list of one-line prompts can't give you.

GUARDRAILS — read before you paste anything

1. Quote-first, always. Every template on this hub requires the model to quote the exact clause text behind every risk flag. A paraphrase-only flag is not evidence — it cannot be checked, and unchecked flags are how plausible-but-wrong AI output gets into real negotiations. If a flag arrives without a quote, discard it and re-run.

2. "Not determinable from the documents provided." The templates require the model to say this — verbatim — whenever a question cannot be answered from what you supplied, instead of filling the gap from general knowledge. A model that always answers is a model that sometimes invents. This sentence is the pressure valve; do not edit it out of the prompts.

3. Privilege and confidentiality — the serious one. Under ABA Formal Opinion 512 (July 2024), a lawyer using generative AI carries duties of competence, confidentiality, communication, and reasonable charges — which in practice means understanding where client information goes when you paste it, and obtaining informed consent where required. The risk is not theoretical: in February 2026, in United States v. Heppner (S.D.N.Y.), documents a defendant had shared with a consumer AI tool were held not privileged. Consumer AI tools are not your firm's document management system: material pasted into them can leave your control, your privilege, and your client's protection.

Practical rules that follow: never paste privileged or identifiable client material into a consumer AI tool; anonymise with [Party A] / [Party B] placeholders when a clause must be reviewed outside approved systems; and prefer tools that don't receive the document at all. That last one is why these generators run 100% client-side — the prompt is assembled in your browser, your contract text is never transmitted to PromptStaging, and what you then paste into an AI tool is your decision, made with the warning above in front of you.

4. Not legal advice. These generators structure a first-pass review. They are not legal advice, they do not create a lawyer–client relationship, and every output requires review by a qualified attorney admitted in the relevant jurisdiction before you rely on it, sign anything, or send a redline to the other side.

INTERNAL LINKS

• NDA Risk Triage Generator — https://promptstaging.site/generators/legal-contracts/nda-risk-triage.html
• Indemnification Clause Comparison — https://promptstaging.site/generators/legal-contracts/indemnification-comparison.html
• Compliance Redline Checklist — https://promptstaging.site/generators/legal-contracts/compliance-redline-checklist.html
• Prompt Generator (all professions) — https://promptstaging.site/#tool
• The Hallucination Firewall — verification patterns for AI output — https://promptstaging.site/article-ai-hallucination-firewall.html
• MCP Builder — https://promptstaging.site/mcp-builder.html
• Privacy Policy — how PromptStaging handles your data — https://promptstaging.site/privacy.html

--- ---