AI Prompting for Lawyers: Complete Guide + 24 Examples

Last Updated: August 8, 2026

Lawyers are experimenting with LLMs (Large Language Models) like Claude and ChatGPT. These tools sit alongside the purpose-built legal platforms now in wide use at firms. Improper use can lead to fabricated citations, privilege exposure, and compliance failures.

AI prompting for lawyers is the practice of writing structured instructions that guide large language models to produce accurate, professional legal outputs. Effective legal prompts specify a role, context, task, and output format. They exclude confidential client data from public AI tools to protect privilege and meet professional-responsibility duties.

This guide covers legal AI ethics, confidentiality safeguards, prompting techniques, common failure modes, the LEGAL framework, 24 reusable prompts, and compliance considerations for selecting AI tools.

What Is AI Prompting for Lawyers?

Two related concepts sit underneath that definition. Prompt engineering is the practice of writing and refining instructions that steer a model toward a specific output, where the level of detail supplied directly shapes what comes back. Priming comes first, in the initial instructions that set context, structure, and style before any substantive question is asked. A primed model that knows the case type, jurisdiction, and expected format produces tighter, more usable output than one given a bare request.

Prompt engineering for lawyers adds requirements that general prompting advice omits. These include jurisdiction and governing law alongside source-bounded drafting that prohibits invented authority. Citation verification against primary sources and confidentiality screening before anything is entered add further requirements.

Unlike casual use of tools like ChatGPT or Gemini, legal prompting requires deliberate structure. LLMs are accessible and powerful. General LLMs lack the purpose-built controls of legal AI platforms, and compliance obligations fall squarely on the lawyer.

AI Ethics and Compliance Requirements for Lawyers

The American Bar Association applied the existing Model Rules to generative AI in Formal Opinion 512. Formal Opinion 512 emphasizes competence, communication, confidentiality, oversight, and candor. These duties must be upheld in every AI-assisted task.

Whether a tool can safely touch client data depends on the product and tier rather than the brand. OpenAI trains on consumer ChatGPT conversations by default, with an opt-out for new conversations, and Anthropic's consumer Claude plans train on inputs by default and are excluded from zero-data-retention arrangements. No-training commitments, retention controls, and BAAs exist only on specific enterprise products, and even then with caveats.

Solo and small firms carry these obligations with the least infrastructure behind them. Written AI policy, vendor review, and a verification workflow all take time a lean practice has already committed elsewhere. Smaller firms face particular pressure to balance these duties against the operational push toward expanding case capacity.

The sanctions record shows what is at stake. The California Court of Appeal imposed a $10,000 sanction and State Bar referral in Noland v. Land of the Free, L.P. (September 2025) on counsel whose AI-drafted appellate briefs cited fabricated authorities. Six months later, in Couvrette v. Wisnovsky, an Oregon federal judge ordered plaintiffs' counsel to pay $94,704.38 in defense fees and costs, split 85/15 between pro hac vice and local counsel, on top of the sanction imposed in the court's December 2025 opinion, the largest AI penalty from an Oregon federal court.

Fines are no longer the ceiling. In June 2026 the Ninth Circuit paired $2,500 fines with six-month court suspensions and a two-year firm-wide AI disclosure requirement.

Key AI Rules & Obligations for Lawyers

Essential Model Rules support compliant AI prompting in legal practice and define the lawyer's continuing responsibility for AI-assisted work.

• Rule 1.1 (Competence): Understand the AI tool's capabilities and risks. Verify outputs before relying on them, and maintain ongoing education about evolving AI technology.
• Rule 1.4 (Communication / Informed Consent): Disclose to clients when AI is used in ways that affect cost, confidentiality, or decision-making. Formal Opinion 512 states that "merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use GAI is not sufficient."
• Rule 1.5 (Fees): Fees must remain reasonable. Lawyers cannot bill clients for hours not performed, may only bill for oversight and review of AI outputs, and cannot charge for time spent learning AI systems.
• Rule 1.6 (Confidentiality): Never paste client-identifiable data into public LLMs. Review terms for retention, training, or third-party access risks.
• Rule 1.7 (Conflicts of Interest): Check whether AI tools might retain or reuse data in ways that create conflicts among clients or across matters.
• Rule 3.3 (Candor Toward the Tribunal): Lawyers are responsible for ensuring the accuracy of AI outputs submitted to courts. False statements, incorrect citations, or misleading arguments violate this duty.
• Rules 5.1–5.3 (Supervision & Vendor Due Diligence): Firms must set internal AI policies, train staff, supervise associates and non-lawyers, and review vendor practices.

Model Rule interpretations and state bar guidance are advisory. The State Bar of California's updated Practical Guidance, approved May 14, 2026, states that as a general matter a lawyer must not input any confidential client information into a generative AI tool that may present material risks to confidentiality or security, absent informed client consent. Proposed amendments to California's Rules of Professional Conduct would carry disciplinary force, but they remain in the comment stage and require Supreme Court adoption.

Court rules are already binding. The Florida Supreme Court amended Rule 2.515(d)(2), effective June 15, 2026, so signers of court filings represent that "the legal authorities identified exist and are accurately cited."

New York's 22 NYCRR Part 161, effective June 1, 2026, applies across the Unified Court System. It permits AI use without general disclosure, while individual courts may adopt a model rule requiring certification that filings contain no fabricated cases, statutes, or material.

AI Practice: Always vs. Never

The Model Rules translate into a short list of operational habits. The highest-stakes obligations provide a quick reference for daily AI use.

Always:

• Verify every AI-generated citation against an authoritative reporter or database
• Confirm the LLM's data retention and training policies before any professional use
• Document AI use in matters where it affects scope, cost, or work product
• Review vendor BAAs and security certifications before handling regulated data

Never:

• Enter client-identifiable information, privileged material, or PHI into public consumer-tier AI tools
• Submit AI-generated work product to a court without independent verification
• Bill clients for time spent learning AI tools or for unperformed work
• Treat AI output as final analysis rather than a draft requiring professional review

The "never" rules target public consumer tools. Approved enterprise products with a signed BAA, contractual no-training commitments, and retention controls change the confidentiality analysis. Verification duties remain identical for every tier.

Legal AI Prompting Checklist

This checklist draws on Formal Opinion 512 and state bar guidance. It addresses the main safeguards for lawyers using LLMs. These safeguards help maintain alignment with professional duties:

• Check the LLM's privacy and data policies.
• Evaluate whether to inform the client or obtain consent.
• Verify all AI outputs for accuracy and citations.
• Ensure billing reflects oversight and actual attorney time.
• Establish firm-wide policies on when and how LLMs may be used.
• Never upload client PHI or confidential files into public tools.
• Assess conflicts from potential data reuse across matters.

How to Prompt AI Without Compromising Client Confidentiality

Confidentiality failures happen at the moment of input, before the model returns a word. Formal Opinion 512 requires lawyers to "evaluate the risks that the information will be disclosed to or accessed by others outside the firm" before entering anything relating to a representation. The workflow runs from tool selection through file documentation.

1. Confirm the tool is on the firm's approved list, with terms of service, privacy policy, and data-retention policy reviewed, as Florida Bar Opinion 24-1 directs.
2. Confirm the vendor does not train models on user inputs at the tier in use.
3. For PHI or medical records, confirm a signed BAA covers the specific product and features involved, not just the vendor.
4. Remove all client names, identifiers, and confidential facts; substitute consistent fictitious placeholders such as "Plaintiff" or "Driver A," a technique recommended in the Kentucky Bar's AI materials.
5. Change nonmaterial facts that do not affect the legal analysis.
6. Strip metadata (for example, by saving to PDF) before uploading any document.
7. Check whether the remaining combination of facts, dates, locations, and roles could still identify the matter to a third party.
8. Specify jurisdiction, time period, and controlling authority in the prompt itself.
9. Independently verify every output before use.
10. Save inputs, outputs, and decisions to the client file, per the Arizona Bar checklist. Document which tool was used, how it was used, and what verification was performed.

A firm representing an individual claimant falls outside the covered entity definition. HHS identifies attorneys whose services to a health plan involve PHI access as business associates, and its cloud computing guidance treats any vendor that processes or stores PHI as a business associate, even when the vendor cannot view the encrypted data.

BAAs are product- and tier-specific. OpenAI offers BAAs for ChatGPT Enterprise with Regulated Workspace and its API with Modified Retention, but does not list ChatGPT Business, while Anthropic's BAA covers its HIPAA-ready API and Enterprise services. Configuration and use determine compliance even when a signed BAA is in place, and the same HHS cloud guidance makes clear that a vendor is not answerable for failures attributable solely to the customer.

De-identification under the HHS Safe Harbor method removes 18 identifier categories. Names and full-face photographs are among them. It takes a vendor outside business-associate status for that data, though re-identification risk never fully disappears.

State authorization rules add separate constraints on releasing medical records into any third-party system.

Courts are split on whether AI prompts stay privileged. In United States v. Heppner, a federal court denied both privilege and work-product protection for AI queries because the platform's privacy policy let the provider collect, review, and disclose user inputs. Warner v. Gilbarco found no waiver on similar facts a week earlier.

Morgan v. V2X extended work-product protection to a pro se litigant's AI-assisted preparation, then held the tool's identity was not itself protected and ordered it disclosed. Until the doctrine settles, a public tool's privacy policy is the practical measure of privilege exposure.

Essential AI Prompting Techniques for Lawyers

Six prompting techniques define the level of guidance given to an LLM in legal tasks and directly influence outcome clarity and reliability. Role, objective, context, format, constraints, and iteration are the variables each technique sets.

They escalate from setting constraints on scope and tone to producing reasoning that can be audited line by line. Each maps to a different legal task, such as one-off statute analysis or repeatable contract review.

Priming Prompts

Priming prompts set the case type, jurisdiction, and expected format before the substantive question arrives. They establish the context, structure, and style that govern everything the model returns afterward.

Example: "Matter parameters: premises liability under Illinois law as of [DATE]. Answers should use numbered, issue-by-issue formatting with a citation for every legal proposition. Confirm these parameters, then await the substantive question."

Role-Based Prompting

Role-based prompting assigns the LLM a defined professional identity to constrain its perspective and tone. Specifying role, practice area, and jurisdiction reduces generic output and keeps responses within scope.

Example: "Act as an experienced intellectual property attorney reviewing a software licensing agreement. Identify provisions that deviate from market standards and flag enforceability concerns under New York law."

Zero-Shot Prompting

Zero-shot prompting works for one-off analysis, statute interpretation, or fact review without examples, and requires explicit constraints and context. It is the weakest fit for open-ended legal research, where every uncited assertion demands verification from scratch.

Example: "Summarize the following statute in plain English, identifying its purpose and main requirements."

Few-Shot Prompting

Few-shot prompting provides examples of the desired response to help the AI produce consistent formatting and content. No fixed number of examples applies, though two or three carrying the firm's house style are usually enough to establish the pattern.

Example: "Two model statute summaries follow: [Example 1: summary of Consumer Protection Act] / [Example 2: summary of Labor Code § XYZ]. Summarize the new statute in a similar style, highlighting purpose, obligations, and penalties."

Chain-of-Thought Prompting

Chain-of-thought prompting breaks analysis into a stated sequence of issue, rule, application, and conclusion so each step can be audited against source material. A stated chain of steps provides a checkable audit trail. The prompt should require a citation or a labeled assumption at every step because the narrative does not prove how the answer was produced.

Example: "Explain step by step whether contract [X] would be enforceable under state law. First, identify issues. Then identify relevant rules. Then apply those rules to the contract facts. Then conclude. For each step, cite the authority relied on or label the point as an assumption."

Iterative Refinement

Iterative refinement treats the first AI output as a working draft, not a final product. Follow-up prompts can narrow the tone, correct errors, or expand specific sections. The process mirrors how a supervising attorney would revise work from a junior associate.

Example: "Revise the prior summary to focus only on procedural defenses available in the first 90 days. Remove the substantive defenses and add citations to the controlling rule."

AI Prompting Pitfalls and Failure Modes

Most AI failures in legal work trace to a handful of repeatable patterns. A Stanford study of more than 200 legal queries found purpose-built legal research tools hallucinated between 17% and 33% of the time, while general-purpose LLMs hallucinated on 58% to 82% of legal queries. Grounding a model in retrieved authority narrows that gap, but the same study found the hallucination problem persists at significant levels.

• Hallucinated citations and fabricated quotations. Invented authority survives casual review because it looks plausible; AI-related sanctions have begun with this failure. Every cited case should be opened and read, every quotation checked against the opinion at the cited page, and the authority confirmed as still good law. The prompt-level control requires a source for each proposition.
• Vague, under-specified, or overloaded prompts. More specific prompts leave less room for the model to fill gaps incorrectly. Complex requests work better split into smaller sequential tasks with the output format specified explicitly, whether a checklist, a table, or a client email under 150 words.
• Missing jurisdiction and stale law. A model's knowledge stops at its training data, so it may miss recent repeals and amendments when the prompt is silent on timing. Accuracy also falls off for less prominent jurisdictions and lower courts, where localized law is thinner in the training corpus. Every research prompt needs the jurisdiction and a controlling-law date, plus an instruction that the model list every assumption it made about facts, jurisdiction, and procedural posture, labeled "given" or "inferred."
• Sycophancy. The Oklahoma Bar Association warns that generative AI is "optimized to maximize user satisfaction," and this design rewards responses that echo the lawyer's assumptions rather than testing them. Adversarial prompts counter it: "Give the strongest argument against the prior conclusion," or ask how opposing counsel would attack the reasoning and how each attack would be answered.
• Long-document and scanned-record limits. ChatGPT's non-Enterprise tiers extract digital text from uploaded files and discard images, so a scanned medical record loses its content entirely unless OCR runs first. Claude rejects files over 1,000 pages and reads pages 101–1,000 as text only, and Gemini caps uploads at 100 MB per file and ten files per prompt. Records-heavy PI work calls for OCR on everything, long records chunked into provider-level sections, chained prompts so one focused output feeds the next, and page or Bates citations so every extraction can be checked.
• Confidential data in the prompt itself. An entered input cannot be un-entered, and Heppner measured privilege by the platform's privacy policy. The confidentiality workflow is the control for this failure mode.

The LEGAL Prompt Engineering Framework

Whether applying zero-shot, few-shot, or chain-of-thought prompting, the LEGAL framework provides a systematic approach to structuring prompts that meet professional standards when working with LLMs. Each letter maps to a decision the lawyer makes before typing: who the model is speaking as, what artifact is due, what sources it may use, how its output gets checked, and which professional rules constrain it. The five elements collapse into the fill-in-the-blank template.

L – Legal Role Assignment

Define the AI's role, such as experienced paralegal or appellate attorney. This keeps analysis within scope. Role assignments should specify:

• Experience level
• Practice area expertise
• Relevant certifications or specializations
• Analytical responsibilities

E – Explicit Goal Definition

State the deliverable, format, length, and audience. This helps outputs align with workflow needs. Effective goals specify:

• Desired output format
• Required length or scope
• Intended audience
• Integration requirements with existing legal technology systems

G – Grounding

Anchor the model to the sources and law it is allowed to use. This prevents the answer from drifting into invented or inapplicable authority. Effective grounding includes:

• Jurisdiction and governing law, stated in every prompt
• A controlling-law date ("state the law as of [DATE]")
• Attached source documents, with an instruction to rely only on the supplied materials
• Relevant statutes, regulations, and applicable procedural rules
• An explicit prohibition on relying on law or facts not stated in the prompt

A – Accuracy Controls

Embed verification protocols for citation requirements, cross-checking, fact-checking, and confidence assessments. These protocols make areas requiring attorney review visible. Built-in accuracy measures include:

• Source document verification requirements
• Cross-referencing protocols
• Fact-checking instructions
• Confidence assessment requests
• Identification of areas requiring human review

L – Legal Standards

Include compliance cues in prompts. These cues define how the model should handle authority and unsupported points. For example:

• Instruct the AI not to fabricate citations
• Format in Bluebook style
• Require unsupported points to be flagged rather than filled in

One caution applies to confidentiality instructions. Telling a model to "assume confidentiality applies" provides only a drafting cue. Confidentiality depends on excluding protected information from the prompt.

A Reusable Master Prompt Template

Each bracket in this reusable prompt maps to a LEGAL element.

"Act as [ROLE, e.g., a senior workers' compensation paralegal] practicing in [JURISDICTION]. Task: [DRAFT / SUMMARIZE / COMPARE / ANALYZE] + [DELIVERABLE]. Sources: rely only on the attached [DOCUMENTS] and the authorities cited in them; do not rely on law or facts not provided. Key facts (anonymized): [FACTS]. Constraints: [LENGTH, ISSUES TO PRIORITIZE, EXCLUSIONS]. Output format: [TABLE / MEMO / NUMBERED LIST], with a citation to the source page or Bates number for each proposition. Mark anything uncertain as 'UNCLEAR' rather than guessing, and list every assumption made about facts, jurisdiction, or procedural posture. End with a checklist of items requiring attorney verification."

24 Essential AI Prompts for Lawyers

Matter-specific facts require the confidentiality workflow before any paste. Each of these AI legal prompts remains scaffolding that still requires attorney verification.

Legal Research & Analysis Prompt Examples

Research prompts carry the highest hallucination risk, because their output is authority the lawyer may go on to cite. Each prompt below fixes a jurisdiction and a controlling-law date, and requires unsettled points to be flagged rather than resolved.

1. Elements & Defenses Map (Zero-Shot)

Use when scoping a new matter or training junior staff. Returns a structured reference map; verify all case cites and pattern jury instructions against primary sources before relying on them. "For [cause of action] in [jurisdiction], list elements, common defenses, leading cases, and any pattern jury instructions with citations. State the law as of [DATE] and flag any element where authority is unsettled. Provide a short practitioner note on proof pitfalls."

2. Statute Quick Sheet (Chain-of-Thought)

Use for unfamiliar statutes or quick refreshers. The step-by-step structure makes the AI's reasoning auditable; flag any deadlines or controlling cases for independent confirmation. "Interpret and summarize [statute/rule]. Step 1: explain its purpose. Step 2: define key terms. Step 3: identify deadlines/limitations. Step 4: outline defenses/exceptions. Step 5: cite controlling cases. Step 6: cross-reference related rules. List every assumption made about jurisdiction or effective date, labeled 'given' or 'inferred.'"

3. Comparative Negligence Table (Zero-Shot)

Use for multi-state intake, conflict-of-laws analysis, or jurisdiction-shopping decisions. Tables are scannable but error-prone; treat caps and citations as starting points for verification, not conclusions. "Build a table that compares negligence and fault allocation rules in [State A], [State B], [State C] and notes caps, joint-and-several rules, and citations to controlling authority. Flag any figure you cannot cite to a specific statute."

4. Affidavit/Certificate of Merit (Zero-Shot)

Use when intaking medical malpractice matters or evaluating filings in unfamiliar jurisdictions. Confirm statutory triggers and expert qualification rules against the current state code, since these requirements change frequently. "Outline certificate of merit requirements for medical malpractice in [state]: triggering statute, timing, content, expert qualifications, and dismissal consequences, with citations."

Document Drafting & Discovery Prompt Examples

Drafting and discovery prompts produce structure rather than substance. The return is a complete section list, an objections framework, or an exhibit map that the attorney then fills with case-specific facts and local rule citations.

5. Contract Review & Risk Flag (Zero-Shot)

Use for first-pass review of standard commercial agreements. Avoid uploading executed contracts containing client-identifiable terms to consumer-tier tools; redline suggestions still require attorney judgment on enforceability and negotiation posture. "Review the attached [contract type] governed by [jurisdiction] law. Identify missing standard clauses, ambiguous language, and provisions that deviate from market norms. For each issue, note the risk to [client role: licensee/buyer/employer] and suggest a redline. Do not assume facts not present in the document."

6. Interrogatories/RFP Bank (Few-Shot)

Use to accelerate discovery drafting for repeatable case types. Few-shot examples train the model on firm style and topic depth; output should be reviewed for jurisdictional rule compliance and case-specific tailoring. "Using these examples [insert 2–3 sample interrogatories], draft a set of standard interrogatories and requests for production for [case type] under [jurisdiction rule], organized by topic (liability, damages, defenses). Include an objections checklist."

7. Meet-and-Confer Letter (Zero-Shot)

Use for routine discovery disputes where format and tone are largely standardized. Tailor the deficiencies table to actual responses and confirm local rule citations before sending. "Draft a meet-and-confer letter citing [Rule] addressing deficient discovery responses: deficiencies table, requested cure, and notice of potential motion to compel."

8. Deposition Outline & Digest (Chain-of-Thought)

Use to scaffold deposition preparation for technical or regulated witnesses. Step-by-step reasoning surfaces topical gaps; the outline is a starting framework, not a substitute for case-specific strategy. "Create a deposition outline for a [witness type] in [case type]. Step 1: identify governing regs (e.g., FMCSA/OSHA). Step 2: map questions to elements. Step 3: align exhibits and impeachment anchors. Step 4: for any transcript excerpts provided, digest key testimony with page:line citations and flag internal contradictions."

9. Protective Order Skeleton (Zero-Shot)

Use for first drafts in matters involving sensitive documents or trade secrets. Local rules and judges' preferences vary widely; treat the template as a structure to refine, not a final form. "Draft a protective order template tailored to [jurisdiction/local rule] with definitions, categories of confidential material, challenge procedure, clawback under FRE 502(d), and sealing process."

10. Privilege Log Package (Few-Shot)

Use when standardizing privilege log practice across a matter or training a review team. Use document-specific factual support for each live-log entry instead of copying sample descriptions verbatim. "Given these sample privilege log entries [insert 2–3], provide a template and drafting guidance compliant with [jurisdiction]. Include recurring drafting errors and sample descriptions for attorney–client and work-product."

Plaintiff Personal Injury Prompt Examples

Medical records are where structured prompting pays off most, and where sloppy prompting costs most. The ABA's 2026 guide to AI prompts for personal injury lawyers puts it plainly: medical records are usually complex, and a vague prompt produces either a vague or a completely incorrect answer.

Disciplined chronology development moves a case from raw records to verifiable analysis. The prompts below form a sequence in which each output feeds the next, so an incomplete inventory propagates into every step that follows.

11. Medical Records Inventory (Zero-Shot)

Use at intake before any chronology or demand work. The inventory catches missing providers early, when a follow-up request costs days rather than weeks. "Attached are medical records and bills for a claimant injured in a [TYPE OF INCIDENT] on [DATE]. Inventory what is provided: (1) every distinct provider or facility, with the date range of treatment for each; (2) the document types present for each provider (ER records, office notes, imaging reports, PT notes, itemized bills); (3) a gaps-and-flags list covering illegible pages, apparent duplicates, bills without matching treatment notes, and references to providers whose records are not in this set. Do not draft a chronology yet. Where a date, name, or figure is unclear, write 'UNCLEAR' rather than guessing."

12. Bates-Cited Medical Chronology (Zero-Shot)

Use once the inventory confirms the record set is complete. The Source column makes every entry checkable against the underlying page, which separates a usable chronology from an unverifiable summary. "Build a medical chronology from these records as a table sorted by date of service: Date | Provider/Facility | Type of Visit | Key Findings/Diagnoses | Treatment/Plan | Source (page or Bates number). Use one row per encounter and the date of service rather than the dictation date. Quote diagnoses closely; do not paraphrase a diagnosis into something more severe. If two records conflict on a date or finding, add a row and flag the conflict. Mark unreadable content 'UNCLEAR' with the page reference. After the table, list any treatment gaps longer than 30 days."

13. Treatment Gap & Inconsistency Analysis (Chain-of-Thought)

Use before demand drafting to anticipate the defense's causation arguments. Every finding must cite specific entry dates so it can be verified. "Using the chronology above, identify: (a) all treatment gaps exceeding 30 days between consecutive entries, with exact date ranges and durations; (b) any referral mentioned in a provider note that does not appear as a later entry; (c) any inconsistency between symptoms the patient reported and findings the provider documented. Present findings as a numbered list citing the specific entry dates for each item."

14. Medical Billing Audit (Zero-Shot)

Use for specials calculation and duplicate detection. Every dollar figure requires verification against the bills before it enters a demand. "Extract all medical billing entries and their CPT and ICD codes from the records. Organize them by date, provider, and type of service. Calculate total medical expenses and flag any duplicate bills or other discrepancies. Where a figure is unreadable, list the provider with 'AMOUNT UNCONFIRMED' rather than estimating."

15. Demand Letter Draft (Zero-Shot)

Use only after the chronology, gap analysis, and billing audit are verified; the letter is only as strong as its inputs. "Draft a demand letter for a [INCIDENT TYPE] claim. Liability facts: [DESCRIBE]. Injuries: [LIST]. Treatment summary: [SUMMARIZE]. Medical expenses: [TOTAL AND BREAKDOWN]. Lost wages: [AMOUNT AND CALCULATION BASIS]. Structure: (1) introduction identifying parties and the incident; (2) liability analysis with specific negligent acts; (3) injury narrative connecting the incident to each injury; (4) itemized damages with supporting calculations; (5) pain-and-suffering analysis using both multiplier and per diem methods; (6) settlement demand of $[AMOUNT] with a 30-day response deadline; (7) reservation of rights to file suit."

Workers' Compensation Prompt Examples

Workers' compensation prompting turns on impairment ratings and wage-loss calculations. It must also account for authorization deadlines that vary by state. The recurring documents in a comp practice reflect these structures.

16. Impairment Rating Review (Zero-Shot)

Use when treating-physician and IME ratings diverge. Confirm the AMA Guides edition against the state's controlling standard before relying on the output. "Organize the impairment rating information for this claim. Date of injury: [DATE]. Body parts: [LIST]. Treating physician rating: [PERCENTAGE]. IME rating: [PERCENTAGE]. Rating basis: [AMA GUIDES EDITION]. Organize: (1) how each rating was determined; (2) what the rating means for benefits; (3) any disputes about the rating; (4) medical evidence supporting or contradicting each rating."

17. Settlement Demand With Wage-Loss Narrative (Zero-Shot)

Use to assemble the economic-damages skeleton before attorney valuation. "Organize a settlement demand for a workers' compensation claim. Injury: [DESCRIBE]. Date of injury: [DATE]. Treatment: [SUMMARY]. Impairment rating: [PERCENTAGE]. Wage loss: $[AMOUNT]. Future medical: $[ESTIMATE]. Settlement demand: $[AMOUNT]. Sections: (1) liability summary; (2) medical treatment summary; (3) impairment and disability; (4) economic damages including wage loss and medical expenses; (5) future medical needs; (6) settlement demand with justification."

18. Treatment Authorization Letter (Zero-Shot)

Use for adjuster correspondence on contested care. Authorization deadlines vary by state; verify the statutory citation before sending. "Draft a letter to an insurance adjuster requesting medical treatment authorization. Injury: [DESCRIBE]. Treatment requested: [DESCRIBE]. Medical necessity: [EXPLAIN WHY IT IS NEEDED]. Include: date of injury and claim number, treatment description, medical-necessity documentation, [STATE] law requirements for authorization, a response deadline, and the consequences of denial."

Client Communication & Practice Management Prompt Examples

Client-facing AI use spans communication, intake, and engagement workflows. Firms often automate earlier touchpoints first. These workflows commonly begin where matters originate before expanding into broader practice management.

19. Plain-English Explainer (Zero-Shot)

Use for client newsletters, intake materials, or website updates explaining public legal developments. Avoid using for matter-specific advice and review for accuracy before publication. "Draft a client-friendly explainer of [public legal development]: what changed, who is affected, likely timelines, and 'what happens next', plus a short FAQ."

20. AI Disclosure Clause (Few-Shot)

Use when updating engagement letters to reflect AI-assisted work. Run final language past firm ethics counsel and confirm alignment with applicable state bar guidance, which evolves quickly in this area. "Using these example clauses [insert 2–3], generate engagement-letter language covering scope, supervision, confidentiality, and billing consistency with Model Rules. Include optional client-consent language."

21. Demand Letter Structure Guardrails (Chain-of-Thought)

Use for outlining demand structure before assembling case-specific facts and damages. The output is scaffolding only; substantive content requires medical records, damages calculations, and attorney judgment on settlement posture. "Lay out the structure of a demand letter for [case type/state]. Step 1: list required sections. Step 2: categorize damages. Step 3: identify supporting docs. Step 4: add statutory references."

Quality, Accuracy & Compliance Prompt Examples

Verification prompts govern the work the other prompts produce. Each one builds a repeatable check rather than a single answer, which is what makes an AI-assisted filing defensible after the fact.

22. Citation Audit Protocol (Chain-of-Thought)

Use to build a repeatable verification workflow for any AI-assisted brief or memo. The protocol can be AI-drafted, but execution runs against authoritative reporters and databases; asking an AI tool to confirm its own citations is not verification. "Provide a citation verification workflow for an AI-assisted brief. Step 1: pull each cited case from an official reporter or research database and confirm it exists. Step 2: confirm each quotation appears in the opinion at the cited page. Step 3: confirm the holding supports the proposition and check subsequent history. Step 4: replace or remove any authority that fails, and document the check for the file."

23. Bluebook Formatter Prompts (Few-Shot)

Use for high-volume citation cleanup in briefs and memos. Format conversion still requires verification of case names, reporter pinpoints, and parallel cites against the original source. "Using these rough citations [insert 2–3], convert them into Bluebook format (cases, statutes, regs). Remind user to verify against official reporters."

24. Damages Category Checklist (Zero-Shot)

Use for damages scoping during case evaluation or pre-demand preparation. Statutory caps and special damages rules change frequently; confirm current figures against state code before relying on the output. "List economic and non-economic damages available in [state] PI cases, note any caps or special statutes, and identify documentary support typically required (types only)."

Choosing the Right AI Tool for Legal Work

AI tools for legal work fall into two broad categories. Each is suited to different tasks. Each also carries a different risk profile.

General-purpose LLMs include ChatGPT (OpenAI), Claude (Anthropic), and Gemini (Google). These tools handle open-ended drafting and brainstorming well. They can also summarize statutes.

Their consumer tiers carry training defaults and retention gaps, as well as file-handling limits. They are appropriate for non-confidential work product and learning the techniques in this guide.

Legal-specific platforms are built for legal workflows, and some support confidential or regulated data. Categories include legal research, contract drafting and citation checking, case management with AI features, and medical records and demand letter automation built for plaintiff personal injury and workers' compensation work, a category where capabilities vary by vendor. Security capabilities vary by product and tier, so retention, training, and BAA terms belong in due diligence for each specific product rather than being assumed category-wide.

Purpose-built platforms can also shrink the prompting burden by encoding prompt structure into repeatable legal workflows.

Tool selection depends on the data sensitivity and repeatability of the task. The firm's integration needs also influence whether a general or legal-specific platform fits the work.

Beyond General LLMs: Platforms With Additional Compliance Controls

Compliant AI prompting combines structure, grounded sources, accuracy controls, and confidentiality screening before the first input. General LLMs suit non-confidential drafting and research scaffolding. Medical records and privileged material call for purpose-built platforms, especially in case preparation workflows that depend on complete, verified records.

Tavrn applies that model for plaintiff personal injury and workers' compensation firms, pairing same-day provider outreach with specialist-reviewed chronologies that hyperlink to their source pages. Bigos Law reported roughly 10x faster case closing after consolidating retrieval, chronologies, and demand letters into that single workflow.

To learn more, book a demo.