AI Prompts for Lawyers (Drafting Guardrails)

AI prompts for lawyers work when you draft from a source pack you control: redacted client facts, verified authorities, and a fence that blocks invented cases, statutes, and quotes. You leave with ROLE / TASK / FORMAT shapes for office memos, clause variants, and issue lists for GPT-5.5 Instant, GPT-5.6, Claude Sonnet 5, or Gemini 3.5 Flash, plus [UNKNOWN] markers and citation rules. This guide covers drafting guardrails. Case-brief summaries live in prompting AI for legal summaries. Output is a draft for a licensed lawyer to review, not legal advice or a filing. Paste redacted facts. Soft tip: PromptMake /text can turn a rough drafting ask into a labeled scaffold at https://promptmake.net/text before you paste into your chat model.

Who AI prompts for lawyers help

You need first-pass structure on a memo, a clause set, or an issue list, and you want the model to stay inside facts and authorities you supplied. The patterns fit associates who turn a partner's notes into a memo outline, in-house counsel who need clause variants from a playbook you already approved, and solo lawyers who want a clean issue list before they pull the reporter. Research assistants who sort facts into Issues / Rules / Application / Conclusion (IRAC) blocks land here too, as long as a licensed lawyer owns the legal conclusion.

This page is for drafting. Do not use it as a substitute for a legal summary workflow: extracting dates, parties, and defined terms from a long contract or opinion belongs in prompting AI for legal summaries. Do not use it for courtroom strategy, privilege-sensitive advice, or anything you would file. Treat the model as a drafting clerk with hard limits. You supply SOURCE_PACK, CITATION_LIST, audience, and banned claims. The model proposes section order, tighter verbs, and consistent headers. You reject anything that invents a case name, a statute section, a holding, or a quote.

Firm and bar rules still govern confidentiality, advertising, and competence. Check your jurisdiction before you paste client matter into a consumer chat. Redact names, docket numbers, and unique facts that would identify a person. Use Matter A / Party 1 labels. Keep the official file in your document system, not in the chat history.

Core prompt pattern for legal drafting

Strong AI prompts for lawyers give four inputs before any tone request: the source pack, the citation list, the deliverable, and the honesty fence. The source pack holds redacted facts, dates you verified, defined terms from the deal or complaint, and open questions marked with a question mark. The citation list holds authorities you already pulled: case captions with reporters, statute citations, regulation cites, and contract clause numbers from the document in front of you. The deliverable names the artifact: office memo, clause variant table, or issue list. The honesty fence forbids invented authorities, fake quotes, made-up holdings, and confident answers where SOURCE_PACK is silent.

Paste those four blocks near the top. Put format rules next. Repeat the fence at the end so it survives a long paste. GPT-5.5 Instant, GPT-5.6, Claude Sonnet 5, and Gemini 3.5 Flash follow labeled blocks. Vague "write a legal memo" prompts produce generic essays because the model has no fact pool and no reporter list. Aim for one artifact type per thread. Mixing a memo and a clause redline in the same pass muddies citation style and invents holdings to fill empty IRAC slots. Keep SOURCE_PACK and CITATION_LIST outside the chat in plain text. Update them when you pull a real authority. Feed them into every prompt that touches the draft.

Role, task, format skeleton

Copy this skeleton and fill the brackets with your material:

ROLE: You are a legal drafting assistant for [practice area]. You draft from SOURCE_PACK and CITATION_LIST only. You never invent case names, reporter cites, statute sections, regulation numbers, holdings, quotes, or procedural dates. You do not give legal advice. You produce a draft for a licensed lawyer to review.

TASK: Turn SOURCE_PACK and CITATION_LIST into a [office memo | clause variant table | issue list] for [audience]. Keep every legal proposition tied to CITATION_LIST. Keep every fact tied to SOURCE_PACK. Mark gaps with [UNKNOWN]. Mark any line that needs lawyer judgment with [LAWYER REVIEW].

FORMAT: Use the section headers listed under OUTPUT_SHAPE. Short sentences. Pinpoint cites only when present in CITATION_LIST. If a cite is missing, write [UNKNOWN: authority] and do not guess a reporter. Ban new cases and new statutes.

AUDIENCE: [partner, client-facing letter (lawyer-owned), deal team, or file memo]

SOURCE_PACK: [paste redacted facts, defined terms, dates, and open questions marked ?]

CITATION_LIST: [paste only authorities you verified: caption, cite, and a one-line holding or clause text you typed from the source]

OUTPUT_SHAPE: [list required sections or columns]

RULES: No marketing language. No "the court would likely." No invented legislative history. If a fact or cite is missing, write [UNKNOWN]. Never invent PII. This output is not legal advice.

REMINDER: Never invent cases, statutes, or quotes. Use [UNKNOWN] for gaps and [LAWYER REVIEW] for conclusions.

That reminder line stops confident fake reporters. Models love tidy Bluebook strings. Your rule forces a gap marker you can fill from Westlaw, Lexis, or the official reporter after you check the source.

Citation fences and unknown tokens

Build a short CITATION_FENCE block you paste under RULES every time:

CITATION_FENCE: Forbid new case captions, reporter citations, docket numbers, statute sections, CFR cites, restatement sections, and quoted language unless they appear in CITATION_LIST or SOURCE_PACK. Prefer [UNKNOWN: authority] over a confident guess. If SOURCE_PACK lacks a fact needed for application, write [UNKNOWN: fact] and stop inventing a timeline.

Use two tokens and keep them distinct. [UNKNOWN] means the pack has a hole: missing cite, missing date, missing party role. [LAWYER REVIEW] means the text states a legal conclusion, a risk ranking, or advice a human must own. Example rule: "Do not rank likelihood of success. If the TASK asks for risk, list issues and mark each [LAWYER REVIEW]."

For quotes, require the model to copy only language you pasted. Add: "Quote marks only around text that appears verbatim in SOURCE_PACK or CITATION_LIST. If you cannot find the sentence, paraphrase in plain words and mark [UNKNOWN: quote]." That blocks fabricated parentheticals. For defined terms, paste the term and the definition you took from the contract or complaint. Tell the model to use those spellings and to flag any new defined term it wants to create.

Drafting prompts for memos, clauses, and issue lists

Draft in this order when the matter is new: issue list, then memo body, then clause variants if the work is transactional. An issue list forces the model to name questions before it writes conclusions. A memo then fills IRAC from CITATION_LIST. Clause work comes last so deal language does not leak invented risk rankings into the memo. Reuse the same SOURCE_PACK and CITATION_FENCE. Change only TASK and OUTPUT_SHAPE. After each draft, run a short audit before you move on. Soft tip: if your rough ask is messy, shape the prompt once at https://promptmake.net/text, then fill SOURCE_PACK yourself.

Keep audience labels tight. A partner memo can use issue headings and short application. A client letter still needs a lawyer to own advice, tone, and what you may say under your ethics rules. Ask the model for a file memo first. Convert to client language in a second pass only after you approve the legal content.

Office memo and issue-list prompts

Issue list pattern:

TASK: From SOURCE_PACK, draft an issue list. Columns: Issue (one sentence, question form), Facts needed from SOURCE_PACK, Authorities in CITATION_LIST that might apply, Status ([UNKNOWN] if no authority yet). Cap at [N] issues. Do not answer the issues. Do not add cases missing from CITATION_LIST. Mark any issue that needs lawyer framing with [LAWYER REVIEW].

Office memo pattern:

TASK: Draft an office memo under [word cap] for [partner name or role]. Sections: Question presented, Short answer (must include [LAWYER REVIEW] and may not invent a prediction), Statement of facts (SOURCE_PACK only), Discussion in IRAC per issue, Next steps that are procedural (file a motion, request documents) without advising a specific outcome. Every rule sentence must cite CITATION_LIST. If a rule lacks a cite, write [UNKNOWN: authority] and skip the holding. Ban "courts in this circuit generally."

Memo audit pattern: "TASK: Audit MEMO_DRAFT against SOURCE_PACK and CITATION_LIST. List every case name, statute, quote, and date. Reply Keep, Soften, or Remove. Soften means the claim overreaches the pack. Propose a rewrite that uses [UNKNOWN] or [LAWYER REVIEW]. Never add a new authority."

Clause variant prompts

Clause pattern:

TASK: From PLAYBOOK and SOURCE_PACK, draft [N] variants of [clause type] (example: limitation of liability, termination for convenience). Table columns: Variant name (Conservative / Balanced), Operative text, Defined terms used, Open commercial points marked [UNKNOWN] or [LAWYER REVIEW]. Rules: use only defined terms from PLAYBOOK. Do not invent statutory safe harbors. Do not claim a clause is "enforceable" or "market." If PLAYBOOK lacks a fallback, write [UNKNOWN: playbook] instead of inventing a cap amount.

PLAYBOOK is your firm's approved language or the current draft clause you typed. Paste it. The model should rearrange and compare, not invent a UCC or CISG cite you did not supply. Second-pass clause audit: "TASK: Compare CLAUSE_DRAFT to PLAYBOOK. Flag any new dollar cap, survival period, indemnity scope, or statute cite missing from PLAYBOOK. Rewrite flagged lines with [UNKNOWN] or remove them."

Do not ask the model to "make this clause bulletproof" or to hide risk from a counterparty. Ask for variants the deal team can choose. A licensed lawyer picks the variant and records the commercial call.

Step-by-step lawyer drafting workflow

Use one chat thread per matter label and artifact: Matter A issue list, Matter A memo, Matter A liability clause. Dumping every client into one long thread mixes facts and invents shared citations. The loop below keeps a stable SOURCE_PACK while you swap only TASK and OUTPUT_SHAPE. You spend free ChatGPT, Claude, Gemini, or PromptMake runs on structure, then lawyer time on reporter checks and ethics review.

Keep SOURCE_PACK in a local file: redacted facts, defined terms, dates, and open questions. Keep CITATION_LIST as a separate list you typed from primary sources you opened. The model should never be your research database or your matter management system. If your firm forbids consumer tools for client work, stop and use the approved vendor instead. The prompt shapes still apply inside that vendor.

Step 1: Build the source pack and citation list

List what you may say. Example: "Matter A (redacted). Dispute over delivery date. Contract §4.2 (pasted). Party 1 claims delay. Party 2 cites force majeure clause (pasted). CITATION_LIST: [your verified case and statute lines]. Ban: invented cases, quoted language not in pack, client names." Ugly notes beat a polished fake memo.

Mark every authority as typed-from-source or absent. In the prompt, tell the model to keep absent items as [UNKNOWN]. Guessing a Federal Reporter cite when you mean "I still need to Shepardize" creates a memo a partner cannot trust and a possible competence problem under your bar rules.

Step 2: Outline, then draft one artifact

Name the artifact and audience before you paste long notes. Ask for an outline first when the memo is new:

TASK: From GOAL and AUDIENCE, propose an outline with headers only. Do not draft body text yet. Flag any header that needs facts or authorities missing from SOURCE_PACK or CITATION_LIST.

Use that outline to decide what to paste next. Partial facts get [UNKNOWN] subsections. Missing research stays out of the Short answer until you pull the reporter. Then run the full skeleton with OUTPUT_SHAPE matching the outline you approved.

Step 3: Audit cites, then lawyer review

Take the draft into a second message:

TASK: Audit DRAFT against SOURCE_PACK, CITATION_LIST, and CITATION_FENCE. For each case, statute, quote, date, and defined term, reply Keep, Soften, or Remove. Soften means the claim overreaches the pack. Propose a safer rewrite. Never add new authorities or holdings.

Optional scaffold: open https://promptmake.net/text, describe "lawyer drafting prompts for memos, clauses, and issue lists with citation fences, [UNKNOWN] gaps, and [LAWYER REVIEW] flags," generate once, then paste your SOURCE_PACK into the returned structure. Guests get about three runs per day; free accounts get about five. Use a run to shape the prompt, then finish in your chat model.

Save the winning prompt next to the artifact name and model label so the next memo reuses the same wrapper. Print or send only after a licensed lawyer checks every cite against the primary source and owns the legal conclusion.

Mistakes that wreck lawyer drafts

Mistake 1: Asking the model to "write the memo" with no SOURCE_PACK or CITATION_LIST. The model invents cases that sound real and fail the first reporter check. Paste controlled notes instead.

Mistake 2: Allowing invented citations and quotes. If you did not type the authority, ban it in FORMAT. Fake Bluebook strings create malpractice risk. Require [UNKNOWN: authority].

Mistake 3: Treating the draft as legal advice or a filing. Put "This is a draft for lawyer review. Not legal advice." in ROLE and in the memo header the model must output. A human still signs the work.

Mistake 4: Using GPT-5.5 Instant or Gemini 3.5 Flash as the final cite checker. Fast models fit outlines and first-pass issue lists. Route the citation audit to GPT-5.6, Claude Opus 5, or Gemini 3.1 Pro, then still open the reporter yourself.

Mistake 5: Pasting confidential client files, privilege logs, or unredacted PII into a consumer chat without your firm's AI policy check. Redact. Use matter labels. Keep sensitive fields offline.

Mistake 6: One mega-prompt that asks for a memo, a motion, and a clause redline together. Split artifacts. Reuse SOURCE_PACK; change TASK and FORMAT.

Mistake 7: Asking how to hide facts, evade a statute, or "beat" a court process. Refuse that use. Prompt for issue lists and clause variants the lawyer can defend.

Mistake 8: Skipping the audit after a long paste. Narrative memos hide invented parentheticals. Prompt for Keep / Soften / Remove against CITATION_LIST every time.

Model notes for legal drafting (mid-2026)

ChatGPT often defaults to GPT-5.5 Instant for fast chat. Instant fits issue-list outlines, first-pass memo headers, and clause tables when you already locked CITATION_LIST. Keep prompts short: ROLE, TASK, FORMAT, SOURCE_PACK, CITATION_LIST, RULES. Skip long chain-of-thought slogans.

GPT-5.6 (Sol in API naming as of mid-2026) fits harder edit passes: cite audits against CITATION_LIST, contradiction checks between facts and application, and memo drafts that must not invent holdings. Give goal, constraints, and format. Drop "think step by step" padding on reasoning-class models.

Claude Sonnet 5 handles long SOURCE_PACK pastes and tidy IRAC tables well. Claude Opus 5 fits careful audits when the memo must not invent options or reporters. Claude Fable 5 is the widely released top tier when your workspace offers it; check your plan. Haiku 4.5 fits short issue-list reshuffles when latency matters more than a deep audit.

Gemini 3.5 Flash fits volume work: many outline passes and clause variant tables from an updated PLAYBOOK. Gemini 3.1 Pro fits hard reasoning over long context when you paste a thick fact pack and need conflict flags against CITATION_LIST. Hedge on exact menu names in each vendor UI. They shift. Re-check the model picker when you open a new thread.

Prompting split that holds: Instant and Flash get RTF plus a short sample of your house memo format when you need matching structure. GPT-5.6, Opus 5, and Gemini 3.1 Pro get goal + constraints + format, with an explicit "never invent" rule, a [UNKNOWN] token, and a [LAWYER REVIEW] flag. All need your SOURCE_PACK in the message. None replace a human cite check or a licensed lawyer before anyone relies on the draft.

Build lawyer drafting prompts with PromptMake

Write the rough ask in plain words: artifact type (memo, clause, issue list), citation fence, and whether you need a cite audit. Open https://promptmake.net/text and generate a structured prompt once. Expect labeled sections you can fill with SOURCE_PACK and CITATION_LIST.

Edit facts, captions, and statute numbers yourself. PromptMake cannot know your matter or your reporter. Paste the filled prompt into Instant or Flash for drafts, or GPT-5.6 / Opus 5 / Gemini 3.1 Pro for audits. Keep free-tier runs for scaffolding, not five synonym retries of the same weak ask.

Workflow that sticks: SOURCE_PACK + CITATION_LIST → PromptMake scaffold → fill redacted facts → fast model draft → reasoning model cite audit → licensed lawyer review → send or file only after that review. Store one template per artifact so you do not rewrite ROLE and RULES from scratch each matter.

FAQ

What are the best AI prompts for lawyers in 2026?

The best AI prompts for lawyers lead with ROLE and a citation fence, paste a redacted SOURCE_PACK and a CITATION_LIST you typed from primary sources, then demand FORMAT for one artifact: office memo, clause variant table, or issue list. Add a second audit prompt that marks Keep, Soften, or Remove against those packs. Match GPT-5.5 Instant or Gemini 3.5 Flash for drafts and GPT-5.6, Claude Opus 5, or Gemini 3.1 Pro for cite audits, then still open the reporter yourself.

Can AI write a legal memo from scratch?

The model can draft structure and wording from facts and authorities you supply. It should not invent cases, statutes, or holdings. Start from a SOURCE_PACK and a CITATION_LIST you built offline, and treat blank-slate "write my legal memo" prompts as high risk for fiction that fails a partner review. A licensed lawyer still owns the short answer.

How do I stop AI from inventing case citations?

State the ban in ROLE and again in CITATION_FENCE plus a REMINDER line. Forbid new captions, reporters, statute sections, and quotes, and require [UNKNOWN: authority] when a stronger claim wants a cite you did not give. Follow with an audit prompt that compares DRAFT to CITATION_LIST line by line. Then verify Keep lines in the official source.

Are AI prompts for lawyers the same as prompting AI for legal summaries?

Lawyer drafting prompts target memos, clauses, and issue lists with citation fences and [LAWYER REVIEW] on conclusions. Prompting AI for legal summaries targets first-pass extraction from a long contract or opinion: parties, dates, defined terms, clause inventory. Keep those libraries separate so a summary pass does not turn into an invented holding inside a memo.

Should I use GPT-5.5 Instant or GPT-5.6 for lawyer drafting prompts?

Use Instant for issue-list outlines, first-pass memo headers, and clause tables from a playbook you pasted. Use GPT-5.6 when you need a careful cite audit, conflict checks between facts and application, or a memo that must not invent holdings. Run the same packs through both only when you measure quality for a recurring template.

Is output from AI prompts for lawyers legal advice?

Prompt output is a draft for a licensed lawyer to review, not legal advice, so put that limit in ROLE and in the memo header. Bar rules on competence, confidentiality, and unauthorized practice still apply. Do not send a client a model draft as advice or file it. Check your firm's AI policy before you paste matter facts into a consumer tool.

Can PromptMake help with AI prompts for lawyers on the free tier?

Yes. PromptMake /text turns a rough drafting ask into a labeled prompt you can aim at ChatGPT, Claude, or Gemini. Guests get about three generations per day; registered free users get about five. Fill in your own redacted SOURCE_PACK and CITATION_LIST, then paste into Instant, Flash, or a reasoning model for the audit. PromptMake does not practice law and does not store your matter as a case file.