25 ChatGPT-5.5 and Codex Prompts for Legal Drafting: Matter Context, Source Hierarchy, Numbered Structure, Issue Coding, Citation Checks, and Lawyer Review

What this prompt set is for—and what it is not

This article provides 25 legal-drafting prompts for teams that want ChatGPT, Codex, or comparable OpenAI-powered workflows to help organize authorized matter context, structure drafts, check source locations, flag uncertainty, and prepare materials for qualified lawyer review. The prompts are designed for lawyers, legal-operations teams, knowledge-management leaders, legal-technology professionals, compliance teams, educators, and advanced users who already understand that legal drafting is a high-stakes activity where a model output is not a substitute for professional judgment.

This is educational content only and is not legal advice. Nothing in this article creates an attorney-client relationship, interprets law for a particular person or matter, recommends a litigation or transactional strategy, determines whether a filing is sufficient, or authorizes anyone to send, file, sign, publish, negotiate, or rely on a legal document. A qualified lawyer with responsibility for the matter must review all outputs, verify all sources, assess professional obligations, and approve any consequential action.

The prompt set is inspired by a documented OpenAI customer story about Harvey, a legal AI company. OpenAI reports that Harvey uses GPT-6 Astra to help lawyers analyze, synthesize, and draft from court information, law-firm documents, case-law research, and other matter context. OpenAI also reports that Harvey’s memory panel can bring individual preferences into the workflow, including preferences such as numbered lists, prioritizing EDGAR as a source, and color-coding issues by priority. Those are OpenAI-reported product and customer-story claims; they are not independent legal-quality benchmarks and they do not prove that a general prompt list can reproduce Harvey’s product controls, security posture, review workflows, or legal correctness.

The transferable practice is narrower and more practical: legal teams can improve drafting discipline by telling the model what sources are authorized, how sources rank, what jurisdiction and date boundaries apply, what structure is required, which issues should be coded by priority, where citations and quotations must be checked, and when the model must stop and flag uncertainty. OpenAI’s prompt-engineering guidance supports explicit instructions and structured context, while OpenAI’s safety guidance emphasizes constrained inputs and human-in-the-loop review for high-stakes domains. This article turns those general practices into reusable legal-drafting prompts with conservative review gates.

This article provides 25 ChatGPT-5.5 prompts for defensible deep research workflows using source boundaries, domain filters, live steering, citation checks, and decision artifacts. The 25 ChatGPT-5.5 Prompts for Source-Controlled Deep Research: Plan Review, Domain Filters, Live Steering, Citation Checks, and Decision Artifacts article is a focused companion for Source Controlled Legal Research because it directly matches the marker’s focus on source-controlled research and citation-grounded outputs, making it the strongest companion for legal drafting prompts that depend on verified source hierarchy.

The non-negotiable legal AI operating rules

Before using any prompt in this set, establish an operating rule that the model may work only with authorized sources and approved systems. Authorized sources may include public court materials, licensed legal research results that the user is permitted to use in the workflow, client documents approved for that system, public company filings, matter chronologies, internal templates, and lawyer-approved summaries. If the team has not confirmed authorization, the correct instruction is not “analyze anyway”; it is “stop and identify what authorization or redaction is required.”

Confidentiality and privilege controls must come before drafting convenience. Do not paste privileged communications, confidential client material, sealed records, regulated personal information, proprietary transaction data, trade secrets, settlement communications, juvenile records, health information, financial account details, credentials, or personal identifiers into a model unless the responsible organization has approved the system, the user is authorized, and the minimum-necessary information has been selected. When possible, use redacted excerpts, synthetic examples, neutral matter labels, source IDs, or lawyer-prepared summaries instead of raw sensitive records.

Jurisdiction and currentness are mandatory drafting variables, not optional metadata. A prompt that asks for a motion outline, contract clause, diligence memo, regulatory summary, or board note must state the governing jurisdiction, court or agency if relevant, document type, relevant time period, and date through which authorities have been checked. If the source packet is not current, the model must say so plainly and should not imply that it has checked new cases, regulations, filings, guidance, or docket activity beyond the provided material.

Source-location references are required for consequential legal drafting. The model should be instructed to tie each material proposition to a source ID and pinpoint location, such as a page, paragraph, section, clause, exhibit, docket entry, transcript line, EDGAR filing section, regulation subsection, or research memo heading. If the model cannot locate support in the provided sources, it should mark the statement as unsupported, tentative, or requiring lawyer research rather than smoothing the gap into persuasive prose.

No-fabrication rules must be explicit because legal drafts often contain names, dates, docket numbers, quotations, case holdings, statutory text, regulatory requirements, defined terms, and procedural deadlines that can cause serious harm if invented. Every prompt in this article assumes the model must not fabricate, silently complete, or imply verification of authorities, citations, quotations, dates, dollar amounts, client instructions, document language, party names, court rules, or filing requirements. If a citation, quotation, or factual proposition is not in the authorized material, the output must flag the absence.

No prompt in this article authorizes external action. The model must not file a pleading, submit a form, send a client email, contact a court, communicate with opposing counsel, publish a document, update a docket, approve a settlement, change permissions, make a payment, sign a document, launch a campaign, or take any other consequential step. Drafting assistance ends at a reviewable work product that a qualified human can inspect, revise, verify, and approve under applicable professional rules and organizational policy.

This legal research and contract-analysis prompt collection provides fifteen GPT-5.5 workflows that can be compared against the current article’s stricter source, confidentiality, verification, and lawyer-review controls. The 15 GPT-5.5 Prompts for Legal Research and Contract Analysis article is a focused companion for Legal Prompt Governance because a legal-specific prompt collection is a more honest companion for legal prompt governance than a broad tooling-stack comparison; the bridge explicitly adds the governance boundary rather than attributing controls the target title does not claim.

How to read OpenAI’s Harvey story without overclaiming

OpenAI’s Harvey customer story is useful because it describes a legal-work pattern that many teams recognize: lawyers need to combine matter-specific materials, legal research, court information, and drafting preferences into structured outputs. The story also highlights preference examples that are easy to operationalize, such as numbered lists, EDGAR prioritization, and issue color-coding. Those examples support a practical lesson: the more precisely a team states source hierarchy, formatting requirements, and review criteria, the easier it is to inspect the resulting draft.

The story should not be read as a promise that any user can paste legal materials into a general chat and receive reliable legal work. OpenAI reports improvements in formatting, context awareness, completeness, and structure compared with other models, but the source notes do not provide independent benchmark values, jurisdiction-specific legal accuracy rates, professional-responsibility analysis, or a guarantee that a generated draft is correct. Legal correctness depends on facts, sources, governing law, procedural posture, currentness, adversarial context, ethics obligations, and lawyer review.

The prompts below therefore use Harvey’s documented pattern as inspiration, not as a claim of replication. They ask for source packets, drafting preferences, structured outputs, citation ledgers, uncertainty registers, adverse-authority checks, and lawyer-review checkpoints. They do not claim to reproduce Harvey, access Harvey’s memory panel, implement Harvey’s controls, or produce an equivalent legal work product. They are practical prompt templates for controlled workflows, not a legal AI product evaluation.

Source-grounded point	Safe transferable practice	Unsafe overclaim to avoid
OpenAI reports that Harvey uses GPT-6 Astra with court information, law-firm documents, case-law research, and other matter context.	Prepare an authorized source packet with source IDs, matter scope, jurisdiction, date boundaries, and access controls.	Claiming that any general model session can safely process all client documents or replace a lawyer’s analysis.
OpenAI describes preference examples such as numbered lists, EDGAR prioritization, and issue color-coding.	State formatting, source-hierarchy, and issue-priority rules directly in the prompt before asking for a draft.	Assuming the model knows a lawyer’s preferences or firm style without explicit instructions and review.
OpenAI reports improvements in formatting, context awareness, completeness, and structure compared with other models.	Use evaluation checklists, side-by-side review, citation verification, and lawyer sign-off to test outputs locally.	Treating source-reported improvements as independent proof of legal accuracy in a specific jurisdiction or matter.

The source hierarchy that every legal-drafting prompt should declare

A legal-drafting prompt should define a source hierarchy before asking for analysis. In litigation, that hierarchy may prioritize controlling statutes, rules, binding appellate authority, trial-court orders in the same matter, record citations, deposition testimony, exhibits, and lawyer-approved research notes. In corporate or securities work, it may prioritize governing agreements, board materials, public filings, statutes, regulations, agency guidance, disclosure schedules, and lawyer-approved diligence summaries. The hierarchy tells the model which sources control, which sources merely inform, and which sources require human confirmation.

For example, a securities-law prompt might state: “When discussing company facts, prefer the provided EDGAR filings over press materials; cite the filing type, date, section heading, and page or item when available; flag any inconsistency between the 10-K, 10-Q, 8-K, proxy statement, and deal documents.” This mirrors the transferable value of OpenAI’s Harvey example about prioritizing EDGAR, but it remains a user-created workflow rule that still requires lawyer verification.

A litigation prompt might state: “For procedural posture, prioritize the docket sheet and court orders; for facts, prioritize admitted evidence and deposition excerpts over attorney argument; for law, prioritize binding authorities within the stated jurisdiction; treat secondary sources as explanatory only.” This instruction reduces the risk that a polished draft will blur evidence, argument, and authority. It also gives the reviewing lawyer a clear way to test whether the model respected the hierarchy.

Recommended source-packet format:
Matter name or neutral matter ID:
User authorization confirmed: yes/no
Approved system or workspace:
Jurisdiction:
Court, agency, or governing forum:
Currentness date:
Document type requested:
Source hierarchy:
  1. Controlling statutes, rules, regulations, or governing agreements
  2. Binding case law or controlling orders
  3. Matter record: pleadings, testimony, exhibits, docket entries
  4. Public filings or official agency materials
  5. Lawyer-approved research notes or internal templates
  6. Secondary sources, only if authorized and identified
Excluded sources:
Confidentiality or privilege constraints:
Required output structure:
Required citation format:
Actions expressly prohibited:
Reviewing lawyer:

Drafting with uncertainty instead of hiding it

Good legal drafting often requires confident prose, but good AI-assisted legal drafting requires visible uncertainty. A model should be instructed to distinguish verified propositions from unsupported statements, ambiguous facts, missing authorities, stale research, unresolved conflicts, and assumptions made only for drafting convenience. If the prompt does not demand uncertainty flags, the output may appear more complete than the source packet supports.

Use an uncertainty register whenever a draft could affect rights, obligations, negotiations, litigation strategy, regulatory submissions, employment actions, financial commitments, or client communications. The register should identify the uncertain point, why it matters, what source is missing, who should resolve it, and whether drafting can proceed with a bracketed placeholder. This turns uncertainty into a review workflow rather than a hidden drafting defect.

Uncertainty type	How the prompt should handle it	Lawyer-review implication
Missing authority	Flag the proposition as unsupported and request legal research rather than inventing a citation.	A lawyer must research and approve the proposition before use.
Ambiguous fact	List competing interpretations and cite the source locations for each version.	A lawyer must decide whether the fact can be asserted and how to qualify it.
Currentness gap	State the latest source date and warn that later authorities or filings may exist.	A lawyer or authorized researcher must update the research before reliance.
Privilege or confidentiality concern	Mark the text for privilege/confidentiality review and avoid unnecessary reproduction of sensitive content.	A responsible lawyer must determine whether the material may be used or disclosed.
Professional-responsibility issue	Stop short of advice and identify the issue category for supervisory review.	A qualified lawyer must analyze duties such as competence, confidentiality, conflicts, candor, and communication.

The review gates that make these prompts usable in real workflows

Each prompt in this article includes a verification checkpoint because the useful output is not the model’s first draft; it is a draft plus a review map. A reviewer should be able to trace every important proposition back to a source location, see which statements are uncertain, identify missing authority, confirm the requested structure, and decide whether further research is required. If the output cannot be audited, it should not be used for client advice, court filing, negotiation, disclosure, or internal decision-making.

OpenAI’s safety guidance supports human-in-the-loop review for high-stakes domains, and legal work plainly falls within a category where errors can have significant consequences. Human review should be more than a final skim. It should include source verification, quotation checking, citation validation, jurisdiction and currentness review, adverse-authority assessment, privilege and confidentiality screening, defined-term consistency, client-instruction confirmation, and professional-responsibility review.

For Codex users, the same principle applies when prompts produce drafting tools, citation ledgers, document parsers, redline utilities, or matter-summary scripts. Code that touches legal documents should be reviewed before use, tested on synthetic or approved sample material, and blocked from destructive actions or external communications unless a qualified human has approved the workflow. A tool that organizes sources can be helpful; a tool that silently alters filed language, emails opposing counsel, or updates a production matter system without approval creates unacceptable risk.

Operational rule: treat every AI-assisted legal draft as unverified until a qualified lawyer confirms authorization, source support, jurisdiction, currentness, confidentiality, privilege, conflicts, professional responsibility, client instructions, and final wording. The model may assist with organization and drafting, but it does not approve legal conclusions or actions.

The contract used by all 25 prompts

All 25 prompts below assume the same baseline contract. The user confirms that the source materials are authorized for the system being used, that confidential or privileged material has been handled under applicable controls, and that the request is for drafting support rather than final legal advice. If those conditions are not met, the model should ask for safer inputs, a narrower task, redacted materials, public sources, synthetic examples, or lawyer-approved summaries.

The prompts also require the model to avoid silent completion. If a case name is incomplete, a statute section is missing, a quote lacks a page reference, a defined term conflicts across documents, or a filing deadline is not provided, the model should not “clean it up” by guessing. It should flag the gap and ask for the missing source or mark the passage for lawyer review. This is especially important in legal work because a plausible but false citation or quotation can be more dangerous than an obvious blank.

Finally, each prompt ends with a verification checkpoint. The checkpoint is not decorative; it is the handoff from model-assisted drafting to professional review. The checkpoint tells the human reviewer what to inspect, what source locations to confirm, what uncertainties remain, and what action remains prohibited until lawyer approval. Teams can adapt the language to their document-management system, legal research platform, or matter workflow, but they should not remove the approval gate.

Baseline instruction to prepend when needed:
You are assisting with legal drafting support only. Use only the authorized sources and facts I provide or specifically identify as approved. Do not fabricate or imply verification of facts, authorities, quotations, citations, dates, docket numbers, defined terms, client instructions, or legal conclusions. Identify jurisdiction, currentness limits, source locations, uncertainty, missing evidence, and conflicts. Do not provide final legal advice and do not file, send, publish, sign, contact anyone, update any external system, or take any consequential action. Prepare reviewable work product for a qualified lawyer.

How to adapt these prompts safely

When adapting a prompt, change the matter-specific fields first: jurisdiction, document type, source hierarchy, currentness date, confidentiality limits, and required output format. Do not begin by adding broader authority to the model. A narrower prompt with better sources is usually safer than a broad prompt that asks the model to “research everything” without source boundaries.

For law firms and legal departments, a practical implementation pattern is to maintain approved prompt versions, test them against synthetic or previously closed-and-approved matters, record known failure modes, and require reviewer sign-off before use in active matters. OpenAI’s evaluation guidance supports testing and evaluation as part of production-quality workflows. In legal settings, evaluation should include whether the prompt caused the model to invent citations, overstate authority, omit adverse facts, mishandle confidentiality, ignore jurisdiction limits, or blur legal advice with drafting assistance.

For educators and students, the prompts can be used to teach issue spotting, source hierarchy, citation discipline, and draft critique, but they should not be used to submit work in violation of academic rules or professional-training requirements. In a classroom, safer inputs include public cases, hypothetical facts, redacted sample contracts, and instructor-approved source packets. The educational value is highest when students compare model output against source documents and identify errors, gaps, and unsupported claims.

For parents and guardians supporting young learners, the safest role for these prompts is explaining structure and source-checking discipline, not generating legal advice or documents for real disputes. If a family faces an actual legal issue involving housing, immigration, custody, school discipline, criminal allegations, employment, debt, benefits, safety, or health, they should seek qualified legal help through licensed counsel, legal aid, a court self-help center, or another appropriate real-world resource.

The rest of the article provides the 25 prompts in sequence. Each prompt is written to be copied, narrowed, and reviewed, not blindly trusted. The most important habit is to keep the source packet visible: what went in, what the model did with it, what remains unsupported, and what a qualified lawyer must decide before anything leaves the draft environment.

Prompts 1–9: Build the Matter Record Before Drafting

These first nine prompts create the legal-drafting control layer: authorization, source inventory, jurisdictional boundaries, authority ranking, fact chronology, issue mapping, argument testing, numbered structure, and source preference. OpenAI reports that Harvey uses GPT-6 Astra with court information, law-firm documents, case-law research, and other matter context, and that preference examples can include numbered lists, prioritizing EDGAR, and color-coding issues by priority. The prompts below adapt that documented pattern into a conservative workflow; they do not reproduce Harvey, guarantee legal correctness, or remove the need for lawyer supervision.

Use each prompt with only authorized, properly handled material. The OpenAI prompt-engineering guidance emphasizes explicit instructions and structured context, while OpenAI’s safety guidance supports human-in-the-loop review in high-stakes work. For legal teams, that means every output must remain a draft work product for qualified review, not a filing, client instruction, negotiation position, or legal opinion.

This independent-assessment evidence-room guide addresses controlled access, methods, confidentiality, and remediation for sensitive evidence, providing a transferable governance pattern for authorized legal source packets without claiming legal-specific controls. The Prepare a Safety-Case Evidence Room for Independent AI Assessment: Claims, Access, Methods, Confidentiality, and Remediation article is a focused companion for Confidential Document Handling because the target directly covers confidentiality and controlled evidence access, making it more defensible than a generic bulk-document processing article.

Prompt 1: Authorization and matter intake

Purpose

This prompt establishes whether the user is authorized to use the matter materials, what confidentiality controls apply, who the supervising lawyer is, and what the model is allowed to do. It is the intake gate before any drafting, summarization, extraction, or analysis occurs.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with a legal-drafting workflow. Before analyzing or drafting, create an authorization and matter-intake checklist.

LOCKED LEGAL SAFETY AND LAWYER-REVIEW CONTRACT:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
1. Ask only for minimum necessary matter metadata, not confidential substance.
2. Confirm whether the user has authorization to use each source category.
3. Identify confidentiality, privilege, protective-order, sealed-record, privacy, or regulated-data issues that require human review.
4. Define the permitted output type: intake checklist only, research plan, draft outline, memorandum draft, redline explanation, or verification ledger.
5. List actions that must remain reserved for a qualified lawyer.

Return:
- Matter intake table
- Authorization status by source category
- Missing approvals
- Professional-responsibility questions
- Safe next step recommendation

Required inputs

• Matter name or neutral matter code, without unnecessary client identifiers.
• Supervising lawyer or review owner role.
• Source categories expected to be used, such as pleadings, contracts, discovery, statutes, regulations, cases, SEC filings, public records, or internal research notes.
• Jurisdiction, forum, or transaction context if known.
• Any known protective order, confidentiality designation, sealed material, or client instruction relevant to AI use.

Expected output

The output should be a concise intake matrix that separates permitted sources from sources requiring additional approval. It should also identify whether the requested task is safe to begin, needs a human authorization decision, or should be stopped until controls are clarified.

Verification checkpoint

A qualified lawyer or authorized matter manager should confirm the matter code, source permissions, confidentiality status, and intended use before any substantive legal analysis is performed. If authorization is unclear, do not paste documents or ask the model to infer permission from context.

Prompt 2: Source packet inventory

Purpose

This prompt turns a disorganized collection of legal materials into a source packet inventory. The goal is not to analyze the merits yet; it is to create a manifest showing what exists, where it came from, what it can support, and what requires verification.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are organizing a legal source packet for later drafting. Do not analyze the merits yet. Build a source inventory and verification plan.

LOCKED LEGAL SAFETY AND LAWYER-REVIEW CONTRACT:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
Create a source packet inventory from the materials described below.

For each source, capture:
1. Source ID
2. Source type
3. Origin or custodian
4. Date or date range
5. Jurisdiction or governing-law relevance
6. Confidentiality or privilege concern
7. Whether it is primary authority, secondary authority, factual record, party document, public filing, or attorney work product
8. Specific source-location format to use later, such as page, paragraph, section, exhibit, docket entry, line, timestamp, or accession reference
9. Verification status: verified, needs verification, incomplete, superseded risk, or unknown

Return a table and a short list of missing materials that may be needed before drafting.

Required inputs

• A list of available sources by filename, document title, docket entry, database result, public filing name, or internal source ID.
• Known dates and jurisdictions for each source.
• Known confidentiality labels or privilege concerns.
• Preferred citation-location convention for the matter team.

Expected output

The output should be a structured manifest that enables later source-linked drafting. A useful inventory will prevent the model from treating attorney notes, adverse pleadings, public filings, and binding authority as if they have equal weight.

Verification checkpoint

Before drafting, compare the inventory against the actual document repository or research folder. Confirm that each source ID points to the correct version and that no privileged, sealed, or unauthorized item has been included by mistake.

Prompt 3: Jurisdiction and date boundary

Purpose

This prompt defines the legal universe for the task: jurisdiction, forum, governing law, procedural posture, and currentness date. It reduces the risk of mixing federal and state rules, using superseded authority, or applying law from the wrong time period.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are setting jurisdiction and date boundaries for a legal-drafting task. Do not provide final legal conclusions. Identify the scope limits that must govern later analysis.

LOCKED LEGAL SAFETY AND LAWYER-REVIEW CONTRACT:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
Using the matter description and source inventory, create a jurisdiction and date-boundary worksheet.

Include:
1. Forum or expected forum
2. Governing law if known
3. Procedural posture or transaction stage
4. Relevant date range for facts
5. Legal-currentness cutoff date
6. Authorities that must be checked for current validity
7. Jurisdiction conflicts or open questions
8. Whether federal, state, local, foreign, arbitral, agency, exchange, or contractual rules may apply
9. Items requiring lawyer confirmation before drafting

Flag every uncertain boundary. Do not resolve contested legal questions without source support and lawyer review.

Required inputs

• Forum, court, agency, arbitral body, or transaction setting.
• Governing-law clause or jurisdictional facts if available.
• Relevant dates for conduct, filing, contract execution, amendments, notices, or agency action.
• Research cutoff date or required currentness standard.

Expected output

The result should be a boundary worksheet that later prompts can reuse. It should make uncertainty visible, especially when the user has provided facts suggesting more than one jurisdiction or an unsettled currentness point.

Verification checkpoint

The supervising lawyer should approve the jurisdiction and date scope before anyone relies on a draft. If the boundary depends on facts not yet verified, mark the entire downstream analysis as provisional.

Prompt 4: Authority hierarchy

Purpose

This prompt creates a ranking system for authorities before the model writes argument text. It helps distinguish binding law from persuasive cases, agency guidance, secondary sources, party assertions, public filings, and internal notes.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are creating an authority hierarchy for a legal-drafting assignment. Do not invent authorities or treat unverified sources as controlling.

LOCKED LEGAL SAFETY AND LAWYER-REVIEW CONTRACT:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
Build an authority hierarchy for the matter.

Use these categories unless the user supplies a different lawyer-approved hierarchy:
1. Binding constitutional, statutory, regulatory, or rule authority
2. Binding appellate or trial authority in the relevant forum, if applicable
3. Persuasive authority from other courts or agencies
4. Official agency releases, orders, interpretations, or no-action materials, if relevant
5. Public filings and official records
6. Contracts, pleadings, discovery, exhibits, and other matter facts
7. Secondary sources and treatises
8. Internal research notes or attorney work product

For each authority, list:
- Source ID
- Rank
- Proposition it may support
- Currentness or validity check needed
- Adverse or limiting treatment to check
- Whether it can be cited directly in the contemplated document

Return a table and a warning list for any source being over-weighted.

Required inputs

• Source inventory from Prompt 2.
• Jurisdiction and date worksheet from Prompt 3.
• Known forum citation rules or lawyer-approved citation conventions.
• Draft type, such as internal memorandum, client alert, motion, contract analysis, board memo, or regulatory summary.

Expected output

The output should give the drafting team a source-weighting map. For example, a party’s complaint may support what that party alleged, but it should not be treated as proof of the underlying fact unless another verified source supports it.

Verification checkpoint

Before using the hierarchy, a lawyer should confirm that the ranking fits the forum and document type. Any authority marked as binding, controlling, current, or citable should be independently checked in an approved legal research system.

Prompt 5: Fact chronology

Purpose

This prompt extracts a chronology of facts while separating verified facts, allegations, inferences, disputed points, and missing evidence. It is especially useful before drafting a statement of facts, background section, transaction timeline, or litigation chronology.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are building a fact chronology for a legal-drafting task. Separate facts from allegations, inferences, and legal conclusions.

LOCKED LEGAL SAFETY AND LAWYER-REVIEW CONTRACT:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
Create a fact chronology from the authorized source packet.

For each entry, include:
1. Date or date range
2. Event description
3. Source ID and source location
4. Actor
5. Fact status: verified, alleged, disputed, inferred, unknown, or privileged-review-needed
6. Relevance to issue or claim
7. Missing evidence or corroboration needed
8. Whether the entry is safe for external use only after lawyer review

Rules:
- Do not fill gaps with assumptions.
- Do not turn allegations into facts.
- Do not include privileged mental impressions in a public-facing chronology.
- Flag inconsistent dates, names, amounts, docket numbers, and document versions.

Required inputs

• Authorized source packet or source summaries with exact location references.
• Relevant factual date range.
• Known claims, defenses, issues, transaction milestones, or regulatory questions.
• Any instruction about excluding privileged communications or attorney mental impressions.

Expected output

The output should be a table suitable for lawyer review, not a polished narrative. The best chronology will make evidentiary gaps obvious and preserve dispute labels rather than smoothing them into a misleading story.

Verification checkpoint

Spot-check every critical chronology entry against the cited source. Do not use the chronology in a filing, client communication, negotiation, or public statement until a qualified lawyer confirms both accuracy and strategic appropriateness.

Prompt 6: Issue map

Purpose

This prompt converts the matter record into a map of legal and factual issues. It helps teams assign issue codes, identify elements, connect facts to propositions, and mark unresolved questions before drafting an argument or memorandum.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are creating an issue map for a legal-drafting workflow. Do not provide final legal advice or final conclusions.

LOCKED LEGAL SAFETY AND LAWYER-REVIEW CONTRACT:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
Create an issue map using the source inventory, jurisdiction worksheet, authority hierarchy, and fact chronology.

For each issue, include:
1. Issue code
2. Short issue label
3. Governing source or rule to verify
4. Elements or analytical factors, if source-supported
5. Supporting facts with source locations
6. Disputed or missing facts
7. Counterarguments or adverse authority to research
8. Priority: high, medium, low, or unknown
9. Drafting destination: facts, legal standard, analysis, risk section, open questions, or appendix

Do not create elements or legal tests unless the source packet supports them or marks them as research-needed.

Required inputs

• Source inventory.
• Jurisdiction and date boundary worksheet.
• Authority hierarchy.
• Fact chronology.
• Known claims, defenses, contract provisions, regulatory obligations, or client questions.

Expected output

The result should be an issue-coded table that allows a lawyer to decide what should be drafted, researched further, or excluded. Priority labels should reflect the matter team’s instruction or visible uncertainty, not an unsupported prediction about case outcome.

Verification checkpoint

Confirm that each issue code maps to an actual claim, defense, transaction question, or lawyer-approved research question. Remove issues based solely on speculation unless the supervising lawyer wants them preserved as research leads.

Prompt 7: Argument and counterargument

Purpose

This prompt tests both sides of an issue without pretending that the model has verified the law. It is useful for internal memoranda, motion planning, negotiation preparation, board-risk summaries, and adversarial review of a draft position.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are preparing a balanced argument and counterargument matrix for lawyer review. Do not write final legal advice.

LOCKED LEGAL SAFETY AND LAWYER-REVIEW CONTRACT:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
For each selected issue code, create an argument and counterargument matrix.

For each issue, include:
1. Issue code and label
2. Draft position
3. Source-supported propositions
4. Supporting facts with source locations
5. Authorities to verify before use
6. Counterarguments
7. Adverse or limiting authority to research
8. Factual weaknesses
9. Professional-responsibility or fairness concern
10. Uncertainty flag
11. Lawyer-review question

Rules:
- Do not strengthen an argument by inventing facts or authorities.
- Do not suppress weaknesses.
- Do not state that an authority is controlling unless verified.
- Do not recommend filing, sending, or making the argument externally.

Required inputs

• Issue map with priority labels.
• Authority hierarchy and currentness notes.
• Fact chronology with source locations.
• Draft posture, such as plaintiff, defendant, buyer, seller, issuer, regulator-facing, employer, employee, board, or internal risk review.

Expected output

The output should show where the draft position is source-supported, where it is vulnerable, and what must be checked before a lawyer can rely on it. A high-quality result will include counterarguments even when the initial position seems strong.

Verification checkpoint

A lawyer should review whether the counterarguments are complete enough for the forum and strategy. Authorities, quotations, and factual propositions must be checked in primary or approved sources before they appear in a memorandum or brief.

Prompt 8: Numbered memorandum structure

Purpose

This prompt applies a numbered structure preference to a legal memorandum or analysis outline. OpenAI’s Harvey story identifies numbered lists as an example of a user preference, but this prompt uses that idea only as a drafting preference for review, not as evidence of correctness.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are creating a numbered legal memorandum structure for lawyer review. Produce an outline, not final legal advice.

LOCKED LEGAL SAFETY AND LAWYER-REVIEW CONTRACT:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
Create a numbered memorandum structure using the matter materials and issue map.

Use this format unless the user supplies a lawyer-approved alternative:
1. Question presented or assignment scope
2. Short answer or preliminary assessment, clearly marked draft
3. Facts and assumptions
4. Governing law or source framework
5. Analysis by issue code
   5.1 Issue
   5.2 Rule or source framework
   5.3 Supporting facts
   5.4 Counterarguments and adverse authority
   5.5 Uncertainty and verification needs
6. Practical implications for lawyer review
7. Open questions
8. Verification ledger
9. Privilege/confidentiality and professional-responsibility review notes

For each section, list the source IDs that should support it and mark any section that cannot yet be drafted safely.

Required inputs

• Issue map.
• Argument and counterargument matrix.
• Authority hierarchy.
• Fact chronology.
• Preferred memorandum audience, such as supervising partner, client legal team, business stakeholder, board committee, or internal research file.

Expected output

The expected output is a numbered outline that prevents the memorandum from blending facts, law, assumptions, and recommendations. The outline should identify which sections are ready for drafting and which need additional research or lawyer judgment.

Verification checkpoint

Review whether the structure matches the assignment and audience. Do not treat a neat outline as a verified analysis; it is only a container for later source-checked drafting.

This guide explains the Structured Prompting Framework, which divides prompts into role, context, instructions, examples, input, and output schema for consistent AI behavior. The The Structured Prompting Prompting Framework: Complete Guide for 2026 article is a focused companion for Structured Drafting Templates because the framework closely supports structured legal drafting templates because it focuses on repeatable prompt sections and defined output schemas.

Prompt 9: EDGAR/source preference

Purpose

This prompt sets a source-preference rule for matters involving public-company filings, securities disclosures, deal documents, governance materials, or financial-reporting context. OpenAI’s Harvey story includes prioritizing EDGAR as an example preference, but teams still must verify filings, accession details, exhibits, dates, amendments, and relevance through approved systems.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are setting source preferences for a legal-drafting task involving public-company, securities, transactional, governance, or disclosure materials. Do not retrieve, invent, or imply verification of filings unless source details are provided and checked.

LOCKED LEGAL SAFETY AND LAWYER-REVIEW CONTRACT:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
Create an EDGAR/source-preference plan for the matter.

When public-company information is relevant, prefer official SEC EDGAR filings or other official source documents over summaries, news reports, or internal paraphrases, unless the supervising lawyer instructs otherwise.

For each needed public-company source, list:
1. Company or registrant name, if already provided
2. Filing type or exhibit type to verify
3. Filing date or period to verify
4. Accession number or official identifier, if already provided
5. Exact proposition the source may support
6. Whether an amended filing, exhibit, incorporated document, or later filing may change the analysis
7. Source-location format to capture later
8. Verification owner
9. Warning if the information is not yet verified

Do not create accession numbers, filing dates, exhibit numbers, quotations, or disclosure language. Mark all missing items as verification-needed.

Required inputs

• Matter type, such as securities litigation, M&A, governance, financing, disclosure counseling, diligence, or public-company contract review.
• Known registrant names, filing types, dates, periods, exhibit references, or accession numbers.
• Source inventory showing whether official filings, internal summaries, or third-party summaries are available.
• Lawyer-approved rule for when EDGAR or another official source should control over summaries.

Expected output

The output should be a source-preference plan that tells the drafting team which propositions need official filing support and which details remain unverified. It should not quote or summarize SEC filings unless the filing text is actually supplied or accessed through an approved, authorized research process.

Verification checkpoint

Before relying on any public-company proposition, verify the exact filing, exhibit, amendment history, period covered, and quoted text against the official source or an approved legal research system. A lawyer should decide whether subsequent filings, incorporated exhibits, or market-context materials alter the drafting position.

Prompts 10–18: Code, Verify, Screen, and Explain the Draft

OpenAI’s Harvey customer story describes a legal workflow pattern built from matter context, lawyer preferences, and structured outputs, including examples such as numbered lists, prioritizing EDGAR as a source, and color-coding issues by priority. The prompts in this section turn that pattern into operational drafting controls without claiming to reproduce Harvey, guarantee legal correctness, or replace a lawyer’s professional judgment.

These nine prompts are designed for the middle of a legal-drafting workflow, after the team has assembled an authorized source packet and before any final advice, filing, client communication, negotiation position, or court submission. They focus on issue-priority coding, quotation verification, citation verification, tables of authorities, missing evidence, privilege and confidentiality screening, conflict questions, defined-term consistency, and redline explanation.

This article explains Chain-of-Verification prompting with templates, examples, and real-world applications designed to reduce AI hallucinations in ChatGPT and Claude outputs. The Chain-of-Verification Prompting: The Advanced Technique That Eliminates AI Hallucinations in 2026 article is a focused companion for Citation Verification Workflow because citation verification in legal drafting depends on checking claims and reducing hallucinated support, making a verification-prompting workflow semantically appropriate.

Prompt 10: Issue-priority coding

Purpose

Use this prompt to convert an issue map into a priority-coded review plan. The goal is not to decide the client’s strategy automatically; it is to help the legal team see which issues appear dispositive, time-sensitive, underdeveloped, fact-dependent, or suitable for later treatment. OpenAI reports that Harvey’s workflow can incorporate preference examples such as color-coding issues by priority; this prompt adapts that idea as a transparent issue-control method rather than a legal conclusion engine.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting under this legal safety contract: Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task: Review the issue map and source packet excerpt provided below. Assign each issue a priority code for drafting control only:
RED = potentially dispositive, jurisdiction-sensitive, deadline-sensitive, or requiring immediate lawyer review.
AMBER = important but dependent on additional facts, unresolved authority, or strategic judgment.
GREEN = background, framing, or lower-risk issue that can be drafted after RED and AMBER items.
GRAY = insufficient information to code; requires missing-evidence entry.

For each issue, provide:
1. Issue label.
2. Priority code.
3. Reason for code.
4. Source-location references supporting the code.
5. Missing evidence or uncertainty.
6. Required lawyer review question.
7. Whether privilege, confidentiality, conflict, or professional-responsibility review is implicated.

Do not create new facts, authorities, deadlines, or client instructions. If the record does not support a priority code, use GRAY.

Required inputs

• Authorized issue map from the matter record.
• Jurisdiction and date scope already approved by the responsible lawyer.
• Source packet index with document names, dates, source types, and page, paragraph, section, exhibit, or line references.
• Any lawyer-approved drafting preferences for priority colors, labels, or section ordering.

Expected output

The expected output is a table with one row per issue and columns for priority code, rationale, source references, uncertainty, missing evidence, review question, and risk flag. A strong answer will explain, for example, that an issue is RED because it may control a motion deadline or statutory element, while another is GRAY because the record does not show the governing contract version or effective date.

Verification checkpoint

A qualified lawyer should confirm that each priority code reflects litigation, transactional, regulatory, or advisory strategy rather than model confidence. The reviewer should downgrade any unsupported RED issue, elevate any overlooked deadline or jurisdictional question, and confirm that the coding scheme does not disclose confidential strategy to unauthorized recipients.

Prompt 11: Quotation verification

Purpose

Use this prompt when a draft relies on exact language from cases, statutes, regulations, contracts, policies, testimony, correspondence, or corporate filings. The prompt requires the model to create a quotation ledger and mark unverified text, because legal drafts can be materially harmed by an altered word, omitted qualifier, missing ellipsis, or inaccurate attribution.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting under this legal safety contract: Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task: Create a quotation verification ledger for the draft excerpt and source excerpts provided below. For every quoted phrase or block quotation:
1. Identify the exact quoted text in the draft.
2. Identify the source document and precise source location.
3. Compare the draft quote against the source text character by character where possible.
4. Mark status as VERIFIED, VARIATION, INCOMPLETE SOURCE, or NOT FOUND.
5. Identify any omissions, bracketed changes, ellipses, capitalization changes, punctuation changes, or emphasis.
6. State whether the surrounding sentence accurately describes the quote without overstating it.
7. Flag any quote that requires lawyer review before use.

Do not supply replacement quotations from memory. If the source text is not provided or the location is unclear, mark INCOMPLETE SOURCE or NOT FOUND.

Required inputs

• Draft excerpt containing every quotation to be checked.
• Underlying source excerpts or approved access to the relevant source packet locations.
• Required citation style or court rule for alterations, ellipses, and emphasis if already known.
• Instruction on whether paraphrases should also be checked for faithful characterization.

Expected output

The output should be a quotation ledger, not a polished brief section. Each row should show the draft quote, source location, status, detected differences, characterization risk, and recommended correction for lawyer review. If a statute excerpt is missing a limiting phrase such as “unless otherwise provided,” the output should flag the variation instead of silently fixing the draft.

Verification checkpoint

The responsible lawyer or trained reviewer must inspect the original source, not only the model’s ledger. The final verification should check the official reporter, statute, regulation, docket entry, contract execution copy, transcript, or filing source as applicable, and should confirm that quotation formatting complies with local rules and professional obligations.

Prompt 12: Citation verification

Purpose

Use this prompt to audit citations before a draft is treated as reliable. The prompt does not ask the model to invent missing authorities or “clean up” unsupported citations; it forces every citation into a status category so the team can distinguish verified authority from placeholders, stale law, incorrect pin cites, and propositions that lack support.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting under this legal safety contract: Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task: Audit the citations in the draft excerpt using only the provided source packet and approved citation materials. For each cited authority:
1. Extract the citation exactly as written.
2. Identify the proposition the citation is offered to support.
3. Verify whether the cited authority exists in the provided materials.
4. Verify jurisdiction, court or issuing body, date, and pin cite if the source is available.
5. Mark status as VERIFIED, PIN CITE NEEDS CHECK, WRONG PROPOSITION, NOT IN SOURCE PACKET, POSSIBLY STALE, or FORMAT ONLY.
6. Flag any need for currentness checks, adverse-authority checks, treatment history, or jurisdictional review.
7. Provide a lawyer-review note.

Do not fabricate citations, treatment history, docket numbers, reporter information, statutory text, or currentness. If verification cannot be completed from the supplied sources, say so.

Required inputs

• Draft excerpt with citations intact.
• Source packet containing the cited materials or approved excerpts.
• Jurisdiction, court, agency, transaction type, or governing-law scope.
• Any applicable citation manual, local rule, style guide, or client preference if available.

Expected output

The expected output is a citation audit table. It should separate form issues from substance issues; for example, a citation may be formatted imperfectly but still support the proposition, while another may be perfectly formatted but attached to a proposition the authority does not support.

Verification checkpoint

No citation should move into a final draft solely because the model marked it VERIFIED. A lawyer or qualified legal professional should check the authority in an approved research system, confirm currentness and treatment, verify pin cites, and decide whether the citation is appropriate for the forum, client objective, and professional responsibility context.

Prompt 13: Table of authorities preparation

Purpose

Use this prompt to prepare a working table of authorities from a draft. The output is a drafting aid, not a filing-ready table, because courts and jurisdictions may impose technical formatting, category, pagination, and citation requirements that must be checked by a lawyer or trained filing professional.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting under this legal safety contract: Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task: Prepare a working table of authorities from the draft text below. Extract authorities only from the draft; do not add authorities. Organize them into categories such as Cases, Statutes, Regulations, Rules, Constitutional Provisions, Administrative Materials, Secondary Sources, Contract Documents, Public Filings, or Other, as appropriate to the draft.

For each authority, provide:
1. Authority name as written.
2. Full citation as written.
3. Category.
4. Draft locations where it appears, using section, paragraph, heading, or page markers available in the draft.
5. Apparent proposition or topic.
6. Verification status if already available from the citation audit.
7. Formatting or duplication issues.
8. Required lawyer-review note.

Do not create page numbers unless the draft provides reliable pagination. Do not imply court-rule compliance.

Required inputs

• Draft text with stable section, paragraph, page, or line markers if available.
• Any prior citation audit table.
• Forum-specific table-of-authorities requirements if known.
• Instruction on whether non-legal materials, exhibits, or public filings should be included.

Expected output

The expected output is a categorized working table that helps reviewers find duplicate authorities, inconsistent short forms, missing pin cites, and category problems. It should make clear when page references are unavailable because the draft has not been paginated in the final filing format.

Verification checkpoint

Before filing or sending, a qualified reviewer should regenerate or verify the table in the actual word-processing or filing environment, confirm pagination after final edits, check local rules, and ensure that confidential exhibits or sealed materials are not exposed through authority labels or document descriptions.

Prompt 14: Missing-evidence log

Purpose

Use this prompt when the draft contains propositions, assumptions, dates, party descriptions, damages figures, procedural facts, or legal elements that are not yet tied to a source. The missing-evidence log prevents the model from smoothing over gaps and gives the legal team a practical checklist for research, client follow-up, discovery review, or partner decision-making.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting under this legal safety contract: Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task: Create a missing-evidence log for the draft excerpt below. Identify every material assertion, element, date, procedural statement, damages number, defined term, client instruction, quotation, citation, or strategic assumption that lacks a clear source-location reference in the supplied materials.

For each gap, provide:
1. Draft location.
2. Unsupported or under-supported statement.
3. Type of evidence needed.
4. Likely source category to check, limited to authorized systems and approved sources.
5. Risk level: HIGH, MEDIUM, LOW, or UNKNOWN.
6. Consequence if unresolved.
7. Suggested question for the lawyer or matter team.
8. Whether privilege, confidentiality, client-consent, or conflict review is implicated.

Do not fill gaps from memory. Do not assume missing facts are true.

Required inputs

• Current draft excerpt or full draft.
• Available source packet and source-location references.
• List of already-known missing materials, if any.
• Instructions on which sources are authorized for follow-up research or fact development.

Expected output

The output should be a prioritized log, not a rewritten draft. A useful entry might identify that the draft asserts notice was sent on a specific date, but the record contains only a later email discussing notice; the suggested follow-up would be to locate the actual notice letter, service record, or client confirmation before relying on the assertion.

Verification checkpoint

The matter team should resolve high-risk gaps before the draft is circulated outside the approved review group. If a gap cannot be resolved, the lawyer should decide whether to remove the statement, qualify it, seek additional evidence, or disclose uncertainty in a manner consistent with legal obligations and client strategy.

Prompt 15: Privilege and confidentiality screen

Purpose

Use this prompt to identify possible privilege, confidentiality, sealing, personal-data, regulated-data, or protective-order issues before content is shared, quoted, summarized, or incorporated into a draft. The prompt is intentionally conservative because the model cannot determine privilege waiver, consent, protective-order compliance, or legal ethics obligations on its own.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting under this legal safety contract: Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task: Screen the draft excerpt and source references for potential privilege, confidentiality, sealing, protective-order, personal-data, regulated-data, trade-secret, or client-consent issues. Do not make final privilege determinations. For each flagged item:
1. Identify the draft location and source reference.
2. Describe the apparent sensitivity category.
3. Explain why the item should be reviewed.
4. Identify the possible risk of disclosure or misuse.
5. Suggest a conservative handling option, such as remove, anonymize, summarize at a higher level, replace with public source, seek client consent, or escalate to supervising lawyer.
6. Identify whether professional-responsibility, conflict, or court-order review is needed.

Do not ask the user to paste unnecessary sensitive details. If the information appears unauthorized or too sensitive for this system, instruct the user to stop and use approved secure review channels.

Required inputs

• Draft excerpt with sensitive sections minimized to what is necessary for screening.
• Source labels that show whether materials are public, client-provided, discovery, internal work product, sealed, regulated, or otherwise restricted.
• Known protective orders, confidentiality designations, or access limitations summarized without exposing unnecessary confidential text.
• Approved handling rules from the matter team or organization, if available.

Expected output

The expected output is a risk-screening table that flags items for human review. It should avoid definitive conclusions such as “privilege is waived” or “safe to disclose” and instead state that a supervising lawyer must decide whether the material can be used, redacted, anonymized, or replaced.

Verification checkpoint

A qualified lawyer should review every flagged item before any external circulation, filing, publication, client communication, opposing-party communication, or vendor upload. If the draft includes regulated or personal data, the team should use approved systems and organization-specific privacy, security, and records controls rather than ad hoc copying into unapproved tools.

Prompt 16: Conflict and professional-responsibility questions

Purpose

Use this prompt to generate questions that a lawyer or designated conflicts team should consider. It does not perform a conflicts check, interpret professional-conduct rules, or decide whether representation is permitted; it produces a structured list of issues for approved human review.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting under this legal safety contract: Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task: Based on the authorized matter summary and draft excerpt below, generate a conflict and professional-responsibility question list for lawyer review. Do not determine whether a conflict exists. Do not provide legal ethics advice.

Include questions about:
1. Parties, affiliates, witnesses, funders, insurers, vendors, and counterparties.
2. Former, current, and prospective client relationships if disclosed in the provided summary.
3. Confidential information from other matters.
4. Positional conflicts or inconsistent obligations if relevant.
5. Joint representation, waivers, consent, engagement scope, and client identity if relevant.
6. Candor, fairness, confidentiality, supervision, competence, unauthorized practice, and tribunal obligations where implicated by the draft.
7. Any human approvals required before filing, sending, publishing, or advising.

For every question, cite the draft or matter-summary location that triggered it. If the record lacks enough information, say what is missing.

Required inputs

• Authorized matter summary with parties and roles, limited to information approved for this review.
• Draft excerpt or outline showing the work being prepared.
• Jurisdictional and forum context at a high level, without asking the model to decide ethics rules.
• Any firm-approved conflicts intake fields or professional-responsibility checklist headings.

Expected output

The output should be a question list grouped by category. For example, if a draft refers to an insurer directing defense strategy, the prompt should ask whether the client identity, reporting obligations, confidentiality rules, and any tripartite relationship issues have been reviewed by the appropriate lawyer.

Verification checkpoint

The conflicts or professional-responsibility review must occur in the organization’s approved systems and under its policies. The model output should not be stored as a substitute for a formal conflicts clearance, waiver analysis, engagement-letter review, supervisory approval, or ethics opinion.

Prompt 17: Defined-term consistency

Purpose

Use this prompt for contracts, policies, settlement papers, pleadings, briefs, internal memoranda, and transaction documents that rely on defined terms. Defined-term errors can change meaning, create ambiguity, or make a cross-reference unusable; this prompt helps locate inconsistent capitalization, undefined terms, duplicate definitions, circular references, and mismatches between defined terms and source documents.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting under this legal safety contract: Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task: Audit defined-term consistency in the draft below. Use only the provided draft and authorized source materials. Create a defined-term register with:
1. Defined term.
2. Definition text or reference.
3. First use and definition location.
4. Later uses and any inconsistent capitalization, pluralization, abbreviation, or quotation marks.
5. Terms used but not defined.
6. Terms defined but not used.
7. Duplicate or conflicting definitions.
8. Cross-references affected by the term.
9. Source-document consistency issues.
10. Lawyer-review recommendation.

Do not rewrite legal definitions unless asked by a lawyer. Do not infer that two similar terms are interchangeable unless the draft or source expressly says so.

Required inputs

• Draft document or selected provisions with definitions and operative language.
• Source documents containing original defined terms, such as agreements, policies, orders, pleadings, exhibits, or deal documents.
• Instruction on whether the review should be case-sensitive and whether short-form party names are included.
• Any drafting style rules approved for the matter.

Expected output

The expected output is a defined-term register and issue list. A useful response might flag that “Services Agreement” is defined once, later appears as “Service Agreement,” and is cross-referenced in an indemnity provision where the singular form could create ambiguity.

Verification checkpoint

The responsible lawyer should decide whether inconsistent terms are harmless drafting style, negotiated language, intentional distinctions, or material defects. In transaction documents and court filings, defined-term changes should be reviewed against the full document set before acceptance because a small term edit can affect obligations, admissions, releases, remedies, or party identities.

Prompt 18: Redline explanation

Purpose

Use this prompt to explain changes between two draft versions for a supervising lawyer, client team, negotiation team, or internal reviewer. The prompt should summarize what changed and why it may matter, but it must not approve the changes, treat them as legal advice, or send a redline to another party without human authorization.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting under this legal safety contract: Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task: Compare Version A and Version B provided below and prepare a redline explanation for internal lawyer review. Do not accept, reject, negotiate, transmit, or implement any change.

For each material change, provide:
1. Location.
2. Version A text summary.
3. Version B text summary.
4. Change type: clarification, expansion, narrowing, deletion, new obligation, new condition, risk allocation, procedural change, citation change, factual change, defined-term change, confidentiality change, privilege concern, or other.
5. Potential legal or practical significance stated cautiously.
6. Source or instruction that appears to support the change, if provided.
7. Missing evidence or uncertainty.
8. Questions for the responsible lawyer.
9. Whether client approval, opposing-party communication, court permission, privilege review, confidentiality review, conflict review, or professional-responsibility review may be required.

Do not infer intent where the record does not show it. Do not characterize a change as acceptable, enforceable, strategic, or compliant without qualified lawyer review.

Required inputs

• Version A and Version B, with stable numbering, tracked changes, comparison text, or clause references if available.
• Known instruction history, negotiation notes, client comments, or court comments approved for use.
• Source packet references supporting factual, citation, or authority changes.
• Audience for the explanation, such as supervising lawyer, client internal team, transaction team, litigation team, or compliance reviewer.

Expected output

The expected output is an internal redline explanation organized by materiality and document location. It should distinguish cosmetic edits from changes that affect obligations, admissions, standards, deadlines, remedies, confidentiality, or defined terms. Where the reason for a change is unknown, it should say that the record does not show the drafter’s intent.

Verification checkpoint

Before any redline is circulated or relied upon, a qualified lawyer should review the comparison against the full document set, confirm that no privileged comments or internal strategy notes are exposed, and decide whether client approval, counterparty communication, court permission, or additional review is required.

Prompts 19–25: Preserve Uncertainty, Escalate Risk, and Improve the System

The final seven prompts move the workflow from draft production into controlled review. OpenAI reports that Harvey uses GPT-6 Astra with court information, law-firm documents, case-law research, and other matter context, and the prompt pattern below follows that documented idea of combining source context with explicit drafting preferences. These prompts do not reproduce Harvey, do not establish legal correctness, and do not authorize a model to make professional judgments; they are templates for organizing work so a qualified lawyer can inspect, correct, and approve it.

OpenAI’s prompt-engineering guidance emphasizes explicit instructions, structured context, and testing, while OpenAI’s safety guidance emphasizes constrained inputs and human-in-the-loop review for high-stakes domains. In legal drafting, that means the final stage should not ask the model to “make it final.” It should ask for uncertainty flags, adverse-authority review, client questions, supervisory sign-off materials, filing checks, lessons learned, and a quality-improvement plan owned by humans.

This playbook covers frontier AI monitoring for agents, including incident reporting, alignment gates, human control, and stop criteria. The Frontier AI Monitoring Playbook: Incident Reporting, Alignment Gates, Human Control, and Stop Criteria for AI Agents article is a focused companion for Human Review for High Stakes AI because high-stakes legal AI use requires human control and escalation boundaries, and this article directly addresses governance mechanisms for keeping AI systems under review.

Prompt 19: Uncertainty register

Purpose

Use this prompt to convert hidden weaknesses into a structured uncertainty register before a draft is relied on. The register should identify unsupported propositions, ambiguous record facts, currentness concerns, conflicting authorities, procedural assumptions, undefined client instructions, and places where the model may have inferred more than the sources justify.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting support under the legal safety contract below.

Legal safety contract:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
Review the draft and source packet described below. Create an uncertainty register that lists every proposition, fact, citation, procedural statement, client instruction, or drafting assumption that requires lawyer confirmation or source verification.

For each uncertainty, provide:
1. Uncertainty ID.
2. Draft location.
3. Affected issue code or section number.
4. Type of uncertainty: fact, law, quotation, citation, procedure, currentness, authority weight, client instruction, privilege/confidentiality, conflict/professional responsibility, or drafting preference.
5. Source currently supporting the point, with page, paragraph, exhibit, docket, section, or other location if provided.
6. Why the support is insufficient or requires confirmation.
7. Risk level: low, medium, or high.
8. Recommended verification step.
9. Person who should resolve it: drafter, research lawyer, supervising lawyer, client contact, records custodian, or other named role.
10. Do not resolve any uncertainty by guessing. If a source is missing, say “source not provided.”

Matter context:
[Insert authorized, redacted matter context.]

Draft excerpt or full draft:
[Insert authorized draft text.]

Source packet:
[Insert source list with locations.]

Jurisdiction and date boundary:
[Insert jurisdiction, court/forum if applicable, and currentness date.]

Required inputs

• Authorized draft text or an excerpt that can be reviewed under applicable confidentiality and privilege controls.
• Source packet manifest with source names and source locations, not just generalized descriptions.
• Jurisdiction, forum, procedural posture, and currentness date.
• Issue codes or section numbers if the draft uses a structured issue map.
• Any known professional-responsibility, privilege, conflict, or client-instruction boundaries.

Expected output

The output should be a table that makes uncertainty visible instead of smoothing it over. A useful register separates “missing source,” “source exists but location was not provided,” “source may be outdated,” “authority appears adverse or distinguishable,” and “client instruction not confirmed.” The model should not fill in missing docket numbers, holdings, quotations, or dates from memory.

Uncertainty ID	Draft location	Type	Why it matters	Risk	Verification step	Owner
U-001	Section II.A	Currentness	The draft relies on a statute section, but no currentness date was supplied.	High	Research lawyer checks current version and any recent amendments.	Research lawyer
U-002	Facts paragraph 4	Fact	The draft states that notice was received on a specific date, but the source packet contains only an email thread summary.	Medium	Confirm against original email metadata or client-approved chronology.	Drafter

Verification checkpoint

A qualified lawyer should compare each uncertainty item against the underlying source, decide whether the risk level is accurate, and remove or rewrite any draft language that lacks support. The register should become part of the review file, not a substitute for legal research or supervisory judgment.

Prompt 20: Adverse-authority review

Purpose

Use this prompt to test whether the draft acknowledges and distinguishes adverse or limiting authority. It is designed for a source packet that already contains candidate authorities; it should not be treated as a complete legal research search across all possible sources.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting support under the legal safety contract below.

Legal safety contract:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
Review the draft and the authority list. Identify any authority in the provided materials that is adverse, limiting, distinguishable, procedurally harmful, jurisdictionally weaker than presented, superseded, questioned, or incomplete.

Important constraints:
- Do not claim to have searched outside the provided source packet unless a lawyer separately confirms that an approved research system was used.
- Do not invent adverse authority.
- Do not omit an authority merely because it is inconvenient.
- If the source packet is insufficient for an adverse-authority review, state that clearly.

For each authority, provide:
1. Authority name or source identifier.
2. Citation exactly as provided, or “citation not provided.”
3. Source location.
4. Relationship to our position: adverse, partially adverse, limiting, distinguishable, supportive but weak, unknown, or not assessable.
5. Relevant holding or proposition only if it is supported by the provided source location.
6. How the current draft treats the authority.
7. Recommended lawyer-review action: cite, distinguish, omit with reason, research further, update currentness, or escalate.
8. Professional-responsibility or candor concern to consider, if any.
9. Confidence level based only on provided materials.

Matter position:
[Insert position or argument being drafted.]

Draft excerpt:
[Insert authorized draft excerpt.]

Authority list and source packet:
[Insert cases, statutes, regulations, rules, secondary sources, filings, contracts, or other authorities with locations.]

Jurisdiction and currentness date:
[Insert jurisdiction and date.]

Required inputs

• The legal position or argument being tested.
• Draft excerpt showing how authorities are currently used.
• Authority list with source locations and currentness information where available.
• Jurisdiction and forum rules relevant to candor, citation, and disclosure obligations.
• Any supervising-lawyer instructions about authority hierarchy and treatment of nonbinding sources.

Expected output

The model should produce an adverse-authority matrix that helps a lawyer decide what requires further research, disclosure analysis, or drafting revision. It should not say that no adverse authority exists unless the assigned research scope and source packet justify that narrow conclusion.

Authority	Relationship to position	Draft treatment	Recommended action	Candor issue to consider
[Case or source identifier]	Partially adverse	Not discussed in current draft	Supervising lawyer to decide whether to distinguish or conduct further research	Assess forum-specific duty to disclose controlling adverse authority

Verification checkpoint

A lawyer must independently confirm the existence, currentness, procedural posture, precedential weight, and disclosure implications of each authority. This checkpoint is especially important because adverse-authority duties depend on jurisdiction, forum, context, and professional-responsibility rules.

Prompt 21: Client-question list

Purpose

Use this prompt to convert drafting gaps into a disciplined question list for the legal team to review before any client communication. The model should not contact the client or frame questions as legal advice; it should prepare a proposed internal list that a lawyer can approve, edit, or discard.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting support under the legal safety contract below.

Legal safety contract:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
Create an internal client-question list for lawyer review. Use the draft, uncertainty register, and source packet to identify factual, procedural, business, timing, document, and instruction questions that may need client input.

Do not write to the client. Do not ask for unnecessary confidential, personal, regulated, or privileged information. Do not ask for secrets, credentials, account numbers, private keys, or irrelevant personal details. If sensitive information might be needed, propose a safer category-level request and note that the lawyer should choose the approved channel.

For each proposed question, provide:
1. Question ID.
2. Draft section or issue code.
3. Proposed question.
4. Why the answer matters to the draft.
5. Source or gap that triggered the question.
6. Sensitivity level: routine, confidential, privileged, personal, regulated, sealed, or unknown.
7. Recommended safer framing.
8. Whether the question should be asked of the client, internal records custodian, supervising lawyer, or not asked.
9. Deadline relevance, if any.
10. Lawyer approval status: pending.

Draft and issue map:
[Insert authorized draft and issue map.]

Uncertainty register:
[Insert register from Prompt 19 if available.]

Source packet:
[Insert authorized source list with locations.]

Known communication restrictions:
[Insert client-contact rules, protective orders, litigation holds, confidentiality limits, or workspace policies.]

Required inputs

• Draft sections or issue codes that reveal unresolved factual or instruction gaps.
• Uncertainty register from the prior stage, if available.
• Known client-contact rules, engagement boundaries, protective orders, or confidentiality restrictions.
• Source packet locations showing which documents have already been reviewed.
• Deadline information when the question affects filing, negotiation, or transaction timing.

Expected output

The output should be a lawyer-facing triage list, not a client email. Good questions are narrow, source-linked, and necessary for the draft; weak questions ask for broad narratives, duplicate information already in the record, or invite unnecessary disclosure of sensitive information.

Question ID	Issue code	Proposed question	Why it matters	Sensitivity	Recommended recipient
CQ-001	FACT-Notice	Can the records custodian confirm the date the notice was received and identify the source document?	The draft currently relies on an unverified summary date.	Confidential	Internal records custodian before client escalation

Verification checkpoint

A lawyer should approve every question before it is sent, decide the proper recipient, and choose the secure communication channel. No model-generated question should be transmitted automatically, and no client answer should be treated as verified until it is reconciled with the matter record.

Prompt 22: Supervisory review package

Purpose

Use this prompt to prepare a concise supervisory package that allows a partner, supervising lawyer, or designated reviewer to inspect the draft’s provenance, key choices, unresolved risks, and required approvals. The output should reduce review friction without implying that the model has approved the legal work.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting support under the legal safety contract below.

Legal safety contract:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
Prepare a supervisory review package for a qualified lawyer. Do not approve the draft. Do not state that the draft is ready to file, send, publish, sign, or provide to a client. Instead, summarize what must be reviewed.

Create the package with these sections:
1. Matter and draft identifier.
2. Authorized source packet summary.
3. Jurisdiction, forum, procedural posture, and currentness date.
4. Draft purpose and intended audience.
5. Authority hierarchy used.
6. Numbered structure and issue codes.
7. Key propositions with source locations.
8. Adverse or limiting authority requiring attention.
9. Uncertainty register summary.
10. Privilege/confidentiality issues requiring review.
11. Conflict and professional-responsibility questions.
12. Open client questions.
13. Citation and quotation verification status.
14. Redline or version-control notes.
15. Decisions only the supervising lawyer can make.
16. Recommended review order.

Inputs:
Matter summary:
[Insert authorized matter summary.]

Draft:
[Insert draft or section list.]

Source packet:
[Insert source manifest.]

Prior review artifacts:
[Insert uncertainty register, adverse-authority matrix, citation ledger, quotation ledger, client-question list, and redline explanation if available.]

Reviewer preferences:
[Insert preferences for numbered lists, source priority, color coding, issue labels, or other drafting conventions.]

Required inputs

• Draft identifier, version, date, and intended use.
• Source packet summary and authority hierarchy.
• Prior review artifacts, including uncertainty and citation-verification materials.
• Reviewer preferences, such as numbered lists, preferred source hierarchy, or issue-priority coding.
• Explicit statement of what the model must not approve or transmit.

Expected output

The package should resemble a review cover memo with tables for open issues, verification status, and lawyer-only decisions. It should help a supervisor focus attention on high-risk sections first, particularly unsupported facts, adverse authority, currentness, confidentiality, and client-instruction gaps.

Verification checkpoint

The supervising lawyer should compare the package against the draft and source packet, not just rely on the summary. If the review package omits a major issue, misstates the source hierarchy, or overstates verification status, the team should correct the prompt inputs and preserve the error as a quality-improvement item.

Prompt 23: Filing-readiness checklist

Purpose

Use this prompt to create a pre-filing or pre-submission checklist that is explicitly human-owned. The model can organize checks, but it must not determine that a document is actually ready to file, send, serve, publish, sign, or submit.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting support under the legal safety contract below.

Legal safety contract:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
Create a filing-readiness or submission-readiness checklist for lawyer review. Do not state that the document is ready. Do not file, send, serve, publish, sign, or submit anything.

Organize the checklist into:
1. Authorization and client-instruction confirmation.
2. Jurisdiction, forum, rule, and deadline checks.
3. Source and record support checks.
4. Citation and quotation verification.
5. Adverse-authority and currentness review.
6. Factual accuracy and exhibit checks.
7. Privilege, confidentiality, sealing, and redaction checks.
8. Conflict and professional-responsibility checks.
9. Formatting, numbering, table, signature, certificate, and service checks.
10. Version control and attachment checks.
11. Human approvals required before action.
12. Final “do not proceed unless” conditions.

For each checklist item, provide:
- Item ID.
- Question to be answered.
- Required evidence or source.
- Responsible human role.
- Status: not started, in progress, verified by human, blocked, or not applicable.
- Consequence if skipped.
- Notes for lawyer review.

Matter and document:
[Insert authorized matter/document description.]

Forum or recipient:
[Insert court, agency, counterparty, internal committee, or other recipient if applicable.]

Known rules, deadlines, and formatting instructions:
[Insert only if authorized and verified.]

Review artifacts:
[Insert citation ledger, uncertainty register, supervisory review package, and redaction notes if available.]

Required inputs

• Document type, intended recipient, and procedural posture.
• Known rules, deadlines, formatting requirements, and service requirements, if already verified by a human or authorized source.
• Citation, quotation, adverse-authority, uncertainty, and redaction review artifacts.
• Named human owners for final approval, filing, service, publication, or signature decisions.
• Workspace policy on external submissions and consequential actions.

Expected output

The checklist should make “blocked” a valid and visible status. For example, if the certificate of service, redaction review, or deadline calculation has not been verified by a human, the output should mark the item as blocked rather than infer completion from surrounding text.

Item ID	Readiness question	Required evidence	Owner	Status	Consequence if skipped
FR-001	Has a qualified lawyer approved the final version?	Documented approval in the matter file	Supervising lawyer	Not started	Unauthorized filing or communication risk
FR-002	Have citations and quotations been checked against source text?	Citation and quotation ledger with human initials	Research lawyer or assigned reviewer	In progress	Misstatement of authority or record risk

Verification checkpoint

Before any consequential action, an authorized human must confirm every applicable checklist item. The model-generated checklist is only an organizing aid; it is not a docketing system, court-rule engine, service platform, legal opinion, or filing authorization.

Prompt 24: Post-matter lessons learned

Purpose

Use this prompt after a matter milestone to identify reusable drafting lessons without retaining or exposing unnecessary confidential content. The goal is to separate durable process improvements from matter-specific facts, privileged strategy, protected client information, or nontransferable outcomes.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting support under the legal safety contract below.

Legal safety contract:
Use only authorized sources and approved systems. Do not request or expose privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never fabricate, silently complete, or imply verification of facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Require source-location references, uncertainty flags, jurisdiction and currentness checks, privilege/confidentiality review, conflict and professional-responsibility checks, and qualified lawyer review. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. Educational content only; not legal advice. Do not claim Harvey’s outcomes are caused by or transferable through these prompts.

Task:
Create a post-matter lessons-learned summary for internal process improvement. Do not include unnecessary client-identifying information, privileged strategy, sealed material, personal data, regulated data, proprietary information, or confidential details unless the user confirms authorization and controls. Prefer abstracted, de-identified, and process-level lessons.

Organize the summary into:
1. Matter milestone reviewed.
2. Drafting task type.
3. What worked well in source organization, structure, citation checking, issue coding, and review.
4. What failed or required rework.
5. Prompt inputs that caused ambiguity or error.
6. Source-packet gaps.
7. Verification failures or near misses.
8. Human review interventions that improved the draft.
9. Lessons that are reusable across matters.
10. Lessons that are matter-specific and should not be generalized.
11. Confidentiality, privilege, retention, and access-control concerns.
12. Proposed updates to templates, checklists, or training materials.
13. Items requiring approval before reuse.

Matter milestone:
[Insert abstracted milestone, such as motion draft completed, contract markup reviewed, memo finalized, or closing checklist updated.]

Permitted information level:
[State whether content must be de-identified, aggregated, internal-only, or matter-file only.]

Review artifacts:
[Insert redacted artifacts, error logs, reviewer notes, citation check results, and approved process observations.]

Do not include:
[Insert exclusions, such as client names, personal data, sealed excerpts, privileged strategy, settlement positions, account details, or confidential business terms.]

Required inputs

• Clear permission level for what may be summarized and retained.
• Redacted or abstracted reviewer notes, error logs, and process observations.
• List of information categories that must be excluded.
• Human-approved description of the matter milestone and drafting task type.
• Retention and access-control rules for lessons learned.

Expected output

The summary should produce process lessons such as “source packets need page-level locations before citation review” rather than matter-specific lessons such as a client’s negotiation strategy. It should identify which recommendations need approval before being added to templates, training materials, or future prompt defaults.

Verification checkpoint

A supervising lawyer or authorized knowledge-management owner should review the summary before it is stored, shared, or reused. If the summary contains client-identifying details, privileged strategy, sealed material, or unnecessary personal information, it should be redacted or discarded according to the organization’s policies.

Prompt 25: Human-owned quality improvement plan

Purpose

Use this prompt to turn reviewer findings and evaluation evidence into proposed improvements while keeping policy, adoption, and production decisions with authorized lawyers and governance owners.

Copy-paste prompt

MANDATORY SAFETY AND AUTHORITY CONTRACT: Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions; verify every citation, quotation, and authority against a source location. Define the jurisdiction and currentness date. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

You are assisting with legal drafting support under this mandatory safety and authority contract. Use only authorized sources and approved systems. Do not request, expose, or process privileged, confidential, personal, regulated, sealed, or proprietary material unless the user confirms authorization and required controls. Never invent or silently complete facts, quotations, citations, authorities, dates, docket numbers, statutes, regulations, holdings, jurisdictions, or client instructions. Define the jurisdiction and currentness date; require source-location references and verification of every citation, quotation, and authority. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action. A qualified lawyer must review and approve the output and make every final decision.

Task:
Create a human-owned quality-improvement plan for the legal drafting prompt or workflow described below. Do not change policy, approve a prompt for production, relax a review gate, or authorize external use. Label all causal explanations as hypotheses unless independently verified.

Return these sections:
1. Workflow scope and intended legal task.
2. Authorized source-packet boundary.
3. Known error categories with evidence references.
4. Root-cause hypotheses and competing explanations.
5. Proposed prompt changes.
6. Proposed source-packet and authority-hierarchy changes.
7. Proposed quotation, citation, adverse-authority, and currentness checks.
8. Proposed privilege, confidentiality, conflict, and professional-responsibility checks.
9. Reviewer-training changes.
10. Evaluation fixtures and tests to run before adoption.
11. Human-defined passing criteria.
12. Rollback and stop conditions.
13. Approval owners and segregation of duties.
14. Access-control, retention, and deletion impacts.
15. Versioned change-log entry and next review date.

Inputs:
- Authorized, redacted lessons learned: [insert]
- Reviewer findings and source locations: [insert]
- Evaluation observations and sample failures: [insert]
- Jurisdiction and currentness date: [insert]
- Governance and client constraints: [insert]
- Current prompt or workflow version and owner: [insert]

Required inputs

• An authorized and redacted lessons-learned summary.
• Reviewer findings tied to draft sections, source locations, and error categories.
• Evaluation observations, including failed or borderline examples rather than invented benchmark results.
• The governing jurisdiction, currentness date, client instructions, and professional-responsibility constraints.
• The current prompt or workflow version, owner, approval path, rollback method, and retention policy.

Expected output

A proposed improvement plan that separates evidence from hypotheses, assigns human owners, defines tests and rollback, and identifies every legal, confidentiality, privilege, access, and verification control that must remain in force. The plan must not claim production approval or legal correctness.

Verification checkpoint

A qualified supervising lawyer and the relevant privacy, security, and governance owners should compare each recommendation against the underlying examples, current law and authority, client instructions, professional obligations, and existing controls. No prompt or workflow change proceeds unless humans approve the evidence, test criteria, access boundary, rollback plan, and permitted use.

Reviewer Workflow for the 25-Prompt Masterclass

Use the prompts as controlled drafting aids, not autonomous legal decision makers. Before each use, confirm source authority, privilege and confidentiality handling, jurisdiction, currentness, and the human reviewer. After each use, compare every factual proposition, quotation, citation, and authority with the authorized source packet; record uncertainty and missing authority; check adverse material; and preserve a versioned review record. Do not file, send, publish, sign, advise a client, contact a court, or take any consequential action through the model.

The Harvey customer story is useful because it documents a pattern: legal source context can be combined with explicit preferences such as numbered lists, EDGAR source priority, and issue color coding. It does not establish that any prompt here reproduces Harvey, its product controls, its security posture, or its legal quality. These prompts are educational content, not legal advice, and they require qualified lawyer review in every matter.

Useful Links

• OpenAI: Harvey turns legal context into stronger drafts with GPT-6 Astra
• OpenAI Developers: Prompt engineering
• OpenAI Developers: Safety best practices
• OpenAI Developers: Evaluation best practices
• OpenAI Developers: GPT-6 Astra model documentation