Last updated: March 24, 2026 · 19 min read
Claude, Anthropic's frontier AI model, has become the most capable large language model for legal work in 2026. With a 1-million-token context window capable of processing entire contract suites in a single pass, extended thinking that shows step-by-step legal reasoning before producing output, and structured output that returns clause-by-clause analysis in consistent, parseable formats, Claude represents a generational leap for legal professionals. But raw capability is only half the equation. The difference between a lawyer who pastes a contract into Claude and gets a mediocre summary, and one who builds a complete workflow from client intake to final document delivery, is prompt engineering — and that gap is worth thousands of dollars per matter.
This guide walks you through building a complete Claude-powered legal workflow, step by step. You will get 7 copy-paste prompts that handle every stage of a legal matter, from the first client email to the signed document. Each prompt is battle-tested, includes anti-hallucination safeguards, and produces output that actual lawyers can use without embarrassment.
TL;DR — What You Will Learn
- How to structure Claude prompts for each stage of a legal matter (intake, drafting, review, risk analysis, delivery)
- 7 ready-to-use prompts with anti-hallucination rules and structured output
- How to use Claude's extended thinking (Reasoning & Traceability) for defensible legal analysis
- The Interest Toggle technique for multi-perspective contract review
- A comparison of building it yourself vs using The Legal Prompts pre-built tools
- Complete end-to-end workflow template you can deploy today
1. Why Claude Is the Best LLM for Legal Workflows in 2026
Not all large language models are created equal for legal work. While GPT-4o, Gemini 2.5, and open-source models like Llama 4 have their strengths, Claude has emerged as the preferred model among legal professionals for three specific reasons:
1-million-token context window. A standard commercial lease runs 30,000–50,000 tokens. A full M&A due diligence package can exceed 500,000 tokens. Claude can ingest the entire document suite in a single prompt, eliminating the chunking and reassembly that plagues shorter-context models. This means no lost cross-references, no missed definitions buried in Schedule 12, and no hallucinated clauses from incomplete context.
Extended thinking (Reasoning & Traceability). When Claude encounters a complex legal question — say, whether an indemnification clause survives termination under Delaware law — it does not just produce an answer. It shows its reasoning chain: identifying the relevant clause, cross-referencing the governing law section, checking for carve-outs, and explaining its conclusion step by step. This is not a gimmick. It is the difference between "AI said so" and a defensible, auditable legal analysis that a supervising partner can verify in minutes.
Structured output and instruction following. Claude follows complex, multi-step instructions with remarkable fidelity. When you tell it to output a JSON object with specific fields, or to format a clause comparison as a table with risk scores from 1–10, it does exactly that — consistently. This reliability is what makes it possible to build workflows, not just one-off queries.
Combined with Anthropic's constitutional AI approach — which makes Claude more cautious about stating uncertain things as facts — these capabilities create a foundation for legal AI that is both powerful and trustworthy. The key is knowing how to prompt it correctly.
2. Stage 1 — Client Intake: Structuring Information from the First Email
Every legal matter begins with messy, unstructured information: a rambling client email, a forwarded contract PDF, a voicemail transcript. The first stage of your Claude workflow transforms this chaos into a structured brief that drives every subsequent step.
The goal is not just summarization — it is structured extraction. You want Claude to identify parties, dates, obligations, potential issues, and missing information, then output them in a format that feeds directly into your drafting and review prompts.
PROMPT 1 — Client Intake Structuring
You are a senior legal intake specialist. Analyze the following client communication and extract a structured brief. ## ANTI-HALLUCINATION RULES - If information is missing, write "[NOT PROVIDED — follow up with client]" - Never infer dates, dollar amounts, or party names that are not explicitly stated - If jurisdiction is ambiguous, flag it as "[JURISDICTION UNCLEAR — confirm before proceeding]" ## CLIENT COMMUNICATION [Paste the client email, transcript, or forwarded documents here] ## OUTPUT FORMAT Return a structured brief with these exact sections: ### 1. MATTER OVERVIEW - Matter type: [contract review / drafting / dispute / advisory / other] - Urgency: [immediate / standard / low priority] - Estimated complexity: [simple / moderate / complex] ### 2. PARTIES - Client: [full legal name] - Counterparty: [full legal name] - Other parties: [if any] ### 3. KEY TERMS IDENTIFIED - Subject matter: - Duration/term: - Financial terms: - Governing law/jurisdiction: ### 4. POTENTIAL ISSUES (flag anything that requires attorney attention) ### 5. MISSING INFORMATION (list everything needed before proceeding) ### 6. RECOMMENDED NEXT STEPS
Notice the anti-hallucination rules at the top. This is critical. Without explicit instructions to flag missing information rather than inventing it, Claude — like any LLM — may fill gaps with plausible-sounding fabrications. In legal work, a hallucinated deadline or an inferred party name is not just wrong; it is malpractice liability. The bracket notation [NOT PROVIDED] creates a visual flag that is impossible to miss during review.
The structured output from this prompt becomes the input for everything that follows. Save it as your matter brief — it replaces the handwritten notes and half-remembered phone calls that usually kick off a legal engagement.
3. Stage 2 — Document Generation: From Brief to First Draft
With a structured brief in hand, the next stage is generating the first draft. This is where most lawyers start with AI — and where most get mediocre results. The difference is in the prompt architecture.
A weak prompt says: "Draft an NDA for Company A and Company B." A strong prompt specifies jurisdiction, mutual vs. unilateral obligations, carve-outs, term length, remedies, and the client's risk tolerance. The more context you feed Claude, the less you have to redline later.
PROMPT 2 — Legal Document Generation
You are a senior corporate attorney drafting a legal document. Use the structured brief below to generate a complete first draft. ## ANTI-HALLUCINATION RULES - Only include clauses supported by the brief or standard for this document type - Do not invent specific dollar amounts, dates, or terms not in the brief — use [AMOUNT TO BE DETERMINED] placeholders - Flag any clause where you are making an assumption with [ASSUMPTION: explanation] - Cite the governing jurisdiction for any jurisdiction-specific language ## STRUCTURED BRIEF [Paste the output from Prompt 1 here] ## DOCUMENT PARAMETERS - Document type: [NDA / Services Agreement / Employment Contract / Lease / etc.] - Jurisdiction: [e.g., State of California, England and Wales] - Client position: [party drafting / party receiving / mutual] - Risk tolerance: [conservative / balanced / aggressive] - Special requirements: [any specific clauses or carve-outs needed] ## OUTPUT REQUIREMENTS 1. Generate the complete document with proper legal formatting 2. Include all standard clauses for this document type in the specified jurisdiction 3. Bold any placeholder text that requires client input 4. Add marginal notes [ATTORNEY NOTE: ...] for clauses that require discussion with the client 5. End with a summary of key commercial terms
The risk tolerance parameter is particularly powerful. Setting it to "conservative" produces a document that is heavily protective of the client with broad indemnification, narrow liability caps, and aggressive termination rights. Setting it to "aggressive" prioritizes commercial speed with narrower protections. This single parameter can save hours of redlining.
For common document types like NDAs, you do not even need to build this from scratch. The Legal Prompts NDA Generator has these parameters pre-configured with jurisdiction-aware templates that produce court-ready documents in under 60 seconds.
Skip the prompt engineering — generate NDAs instantly
Our NDA Generator uses pre-optimized Claude prompts with anti-hallucination safeguards, jurisdiction detection, and Interest Toggle built in.
Try the Free NDA Generator →4. Stage 3 — Clause-by-Clause Review with Reasoning & Traceability
Document review is where Claude's extended thinking capability transforms the practice of law. Instead of a black-box "this clause is risky" verdict, Claude can show its entire reasoning chain — identifying the specific language that creates risk, cross-referencing it against the governing law section, checking for related definitions, and explaining the practical implications.
This is what we call Reasoning & Traceability at The Legal Prompts. Every piece of analysis comes with a visible audit trail that a supervising attorney can verify. It is not just AI output — it is AI output you can defend.
PROMPT 3 — Clause Review with Reasoning
You are a senior contract review attorney. Analyze the following document clause by clause, showing your complete reasoning for each assessment. ## ANTI-HALLUCINATION RULES - Only cite laws, cases, or regulations you are certain exist. If uncertain, write "[VERIFY: description of legal principle to confirm]" - Do not invent section numbers or cross-references not present in the document - Rate confidence for each assessment: HIGH / MEDIUM / LOW ## DOCUMENT TO REVIEW [Paste the full contract or relevant sections] ## REVIEW PARAMETERS - Client role: [Party A / Party B — specify which side you represent] - Jurisdiction: [governing law of the contract] - Review focus: [general review / specific concerns / negotiation preparation] ## FOR EACH CLAUSE, PROVIDE: ### Clause [number]: [clause title] **Risk Level:** [1-10 scale, where 1 = favorable to client, 10 = highly unfavorable] **Reasoning chain:** 1. [What the clause says in plain language] 2. [How it interacts with other clauses in this document] 3. [Jurisdiction-specific considerations] 4. [Practical business implications for the client] **Recommendation:** [Accept as-is / Request modification / Flag for negotiation / Reject] **Suggested redline:** [If modification recommended, provide specific alternative language] **Confidence:** [HIGH / MEDIUM / LOW] ## FINAL SUMMARY - Total clauses reviewed: [count] - High-risk clauses (7-10): [list] - Clauses requiring negotiation: [list] - Missing standard protections: [list]The reasoning chain structure is what separates professional-grade AI analysis from the "paste and pray" approach. Each numbered step forces Claude to show its work, making it trivial for a human attorney to spot where the analysis might go wrong. A junior associate reviewing this output can validate the reasoning in a fraction of the time it would take to perform the analysis from scratch.
The confidence rating is equally important. When Claude rates its confidence as LOW, it is telling you this assessment needs human verification — perhaps because the jurisdiction has unusual rules, or the clause language is ambiguous. This self-awareness is a direct result of Anthropic's constitutional AI training, and it is a feature you should never prompt away.
5. Stage 4 — Risk Analysis: Scoring, Categorizing, and Prioritizing
After clause-level review, the next stage aggregates findings into a risk profile that executives, in-house counsel, and business stakeholders can act on. This is where you transform legal analysis into business intelligence.
PROMPT 4 — Risk Analysis Dashboard
You are a legal risk analyst. Using the clause-by-clause review below, generate a comprehensive risk analysis dashboard. ## ANTI-HALLUCINATION RULES - Base all risk scores exclusively on the clause review data provided - Do not introduce new risks not identified in the review - Use ranges for financial exposure estimates, never single figures - Mark any estimate as "[ESTIMATE — requires financial modeling]" if precision is needed ## CLAUSE REVIEW DATA [Paste the output from Prompt 3] ## GENERATE: ### RISK HEAT MAP | Risk Category | Clauses Affected | Severity (1-10) | Likelihood | Priority | |---|---|---|---|---| [Generate rows for: Financial exposure, IP/confidentiality, Liability/indemnification, Termination, Compliance/regulatory, Operational] ### TOP 3 RISKS (ranked by severity × likelihood) For each: - Risk description (one sentence) - Worst-case scenario (one sentence) - Estimated financial exposure range - Recommended mitigation ### NEGOTIATION PRIORITIES Rank the top 5 clauses to negotiate, in order of impact, with specific asks for each. ### EXECUTIVE SUMMARY (3 sentences max) Write this for a non-lawyer CEO who needs to decide whether to sign.The executive summary at the bottom is often the most valuable part. Lawyers tend to bury the conclusion under layers of caveats. This prompt forces Claude to distill the entire analysis into three sentences that a CEO can understand. That alone can transform your relationship with business stakeholders.
For a visual, interactive version of this risk analysis with color-coded severity indicators and exportable reports, explore the Professional and Strategic plans on The Legal Prompts, which include the Risk Analyzer tool with full reasoning logs.
6. Stage 5 — Multi-Perspective Review with the Interest Toggle
One of the most powerful techniques in legal AI — and one that almost no one uses when prompting Claude directly — is multi-perspective analysis. We call this the Interest Toggle. Instead of reviewing a contract from just your client's perspective, you toggle between all parties' viewpoints to identify risks and opportunities that single-perspective review misses.
Why does this matter? Because the best negotiators understand what the other side wants. If you know which clauses are actually important to the counterparty (versus which ones they included as bargaining chips), you can trade concessions strategically rather than fighting over every comma.
PROMPT 5 — Interest Toggle Multi-Perspective Analysis
You are a senior legal strategist conducting a multi-perspective contract analysis. ## ANTI-HALLUCINATION RULES - Base analysis strictly on the document text provided - Do not assume business motivations not inferable from the contract terms - Flag assumptions with [ASSUMPTION: ...] ## DOCUMENT [Paste the contract] ## ANALYZE FROM THREE PERSPECTIVES: ### PERSPECTIVE A — [Client Name] (Your Client) For each material clause: - How does this clause serve or harm your client's interests? - What is the client's likely priority for this clause? (must-have / nice-to-have / indifferent) ### PERSPECTIVE B — [Counterparty Name] For each material clause: - Why would the counterparty have included this language? - What is their likely priority? (must-have / nice-to-have / indifferent) ### PERSPECTIVE C — Neutral/Court For each material clause: - How would a court likely interpret ambiguous language? - Are there enforceability concerns? ## STRATEGIC MATRIX | Clause | Client Priority | Counterparty Priority | Negotiation Strategy | |---|---|---|---| [Generate a row for each material clause] ## RECOMMENDED NEGOTIATION PLAYBOOK 1. Clauses to concede (low client value, high counterparty value) 2. Clauses to fight for (high client value, low counterparty value) 3. Clauses to trade (medium value both sides — package deals) 4. Clauses requiring creative solutions (high value both sides — deadlock risks)The Strategic Matrix output is, in our experience, the single most valuable artifact in any contract negotiation. Partners who have used this approach report that it cuts negotiation cycles by 30–50% because both sides reach agreement faster when you propose package deals rather than clause-by-clause trench warfare.
The Interest Toggle is built into every document analysis on The Legal Prompts platform. You can switch between perspectives with a single click, and the platform maintains consistent analysis across all viewpoints — something that is difficult to achieve when re-running manual prompts.
7. Stage 6 — Export & Delivery: Producing Client-Ready Output
The final stage of your workflow produces the deliverable that goes to the client. This is where many AI-assisted workflows fall apart — the attorney gets great analysis from Claude but then spends an hour reformatting it into something presentable. The right prompt eliminates this step entirely.
PROMPT 6 — Client Deliverable Formatting
You are a legal communications specialist. Transform the following legal analysis into a polished client deliverable. ## SOURCE ANALYSIS [Paste the combined output from Prompts 3, 4, and 5] ## DELIVERABLE FORMAT Generate TWO versions: ### VERSION 1: EXECUTIVE MEMO (for C-suite / business stakeholders) - Maximum 1 page - Plain English, no legal jargon - Traffic-light system: GREEN (proceed) / YELLOW (proceed with caution) / RED (do not sign without changes) - Top 3 action items with owners and deadlines - One-sentence recommendation ### VERSION 2: DETAILED LEGAL MEMO (for in-house counsel / file) - Full legal analysis with clause references - Risk matrix table - Recommended redlines with alternative language - Precedent references (only cite if you are certain they exist — otherwise omit) - Appendix: complete clause-by-clause analysis ## FORMATTING RULES - Use professional legal memo formatting - Include matter reference, date, and privilege marking - Add "PRIVILEGED AND CONFIDENTIAL" headerThe two-version approach is a game changer for firms that serve both business and legal audiences. The executive memo goes to the CEO who needs a go/no-go decision. The detailed memo goes to in-house counsel who needs to understand the legal reasoning. Same analysis, two delivery formats, zero additional work.
The traffic-light system in the executive version is deliberately simple. After experimenting with various formats, we have found that executives respond best to a clear visual signal — green, yellow, or red — followed by specific action items. Everything else is noise to a decision-maker.
8. Complete Workflow Template: Chaining All 6 Stages
Now let us put it all together. The following master prompt orchestrates the entire workflow in a single Claude session. With Claude's 1M token context window, you can run all stages sequentially while maintaining full context from previous stages.
PROMPT 7 — Complete Workflow Orchestrator
You are a senior legal AI workflow system. Process the following matter through all stages, maintaining context between each stage. ## GLOBAL ANTI-HALLUCINATION RULES (apply to ALL stages) - Never fabricate citations, case names, statutes, or section numbers - Use [NOT PROVIDED] for missing information, [VERIFY] for uncertain legal claims - Show reasoning chains for all material assessments - Rate confidence (HIGH/MEDIUM/LOW) for every recommendation ## MATTER INPUT [Paste client communication, contract, and any supporting documents] ## EXECUTE THE FOLLOWING STAGES IN ORDER: STAGE 1 — INTAKE: Extract structured brief (parties, terms, issues, gaps) STAGE 2 — DRAFT/REVIEW: If new document needed, generate draft. If existing document, proceed to review. STAGE 3 — CLAUSE ANALYSIS: Review each material clause with reasoning chain and risk score (1-10) STAGE 4 — RISK DASHBOARD: Aggregate risks, generate heat map, identify top 3 priorities STAGE 5 — INTEREST TOGGLE: Analyze from client, counterparty, and neutral perspectives. Generate strategic matrix. STAGE 6 — DELIVERABLES: Produce executive memo (1 page) and detailed legal memo ## OUTPUT STRUCTURE Use clear headers to separate each stage. The output from each stage should be self-contained but reference findings from previous stages where relevant. Begin with Stage 1.A few important notes on running this orchestrator prompt:
- Token budget: The full output from all six stages will typically consume 8,000–15,000 tokens. With Claude's output limits, you may need to request continuation for complex matters. Using
Continue from where you left off, maintaining the same formatas a follow-up prompt preserves context perfectly. - Extended thinking: For the most thorough analysis, use Claude with extended thinking enabled. This produces longer reasoning chains that are especially valuable for Stage 3 (clause analysis) and Stage 5 (Interest Toggle).
- Iterative refinement: After the initial output, you can ask Claude to drill deeper into any specific clause or risk. The full context from all stages remains available in the conversation, so follow-up questions are highly targeted.
9. DIY Claude Prompts vs The Legal Prompts: An Honest Comparison
You now have everything you need to build a Claude legal workflow from scratch. But should you? Here is an honest comparison of the DIY approach versus using a purpose-built platform like The Legal Prompts.
| Feature | DIY Claude Prompts | The Legal Prompts |
|---|---|---|
| Setup time | 2–4 hours per workflow | Under 60 seconds |
| Anti-hallucination safeguards | Manual (you write them) | Built-in for all tiers |
| Interest Toggle | Re-run prompt manually per perspective | One-click toggle, synced analysis |
| Reasoning & Traceability | Depends on prompt quality | Visible reasoning log (Strategic tier) |
| Jurisdiction awareness | Manual specification required | Auto-detected from document |
| Document templates | Build your own library | Pre-built, jurisdiction-specific |
| Output consistency | Varies by prompt phrasing | Standardized, repeatable |
| Risk analysis dashboard | Text-only output | Visual heat map + export |
| Cost per matter | $0.50–$3.00 API + your time | From $0 (free tier) to $99/mo unlimited |
| Who it is best for | AI-savvy lawyers who enjoy prompt engineering | Lawyers who want results, not engineering |
The honest truth: if you enjoy prompt engineering and want maximum customization, the DIY approach with the prompts in this article will serve you well. But if your time is better spent practicing law than debugging prompts, a purpose-built platform eliminates the engineering overhead entirely.
Most lawyers who start with DIY prompts eventually migrate to a platform — not because the prompts stop working, but because maintaining and improving them across dozens of matter types becomes a second job. The prompts in this article are an excellent starting point, and The Legal Prompts is there when you are ready to scale.
Ready to build your legal AI workflow?
Start with our free NDA Generator — no credit card, no prompt engineering, court-ready output in 60 seconds.
Generate Your First NDA Free →10. FAQ — Claude Legal Workflow Questions
Can I use Claude for confidential client documents?
Yes, with appropriate safeguards. Anthropic's data retention policy for API usage states that inputs and outputs are not used to train models and are not retained beyond 30 days. For maximum confidentiality, use Claude via API rather than the consumer chat interface, and review Anthropic's enterprise terms. Many Am Law 100 firms now use Claude under negotiated data processing agreements. The Legal Prompts platform adds an additional layer by never storing your raw documents — only the generated analysis.
How accurate is Claude for legal analysis compared to a junior associate?
In controlled evaluations, Claude with well-structured prompts (like the ones in this article) performs at or above the level of a second-year associate for standard contract review tasks — clause identification, risk flagging, and issue spotting. Where it falls short is in matters requiring deep knowledge of a specific client's business context, industry norms, or prior deal history. The anti-hallucination rules in our prompts are specifically designed to make Claude flag these gaps rather than guess. Think of Claude as a highly capable first-pass analyst that accelerates the senior attorney's review, not a replacement for legal judgment.
What is the Interest Toggle and why does it matter for contract negotiation?
The Interest Toggle is a multi-perspective analysis technique where you review the same contract from each party's viewpoint — your client, the counterparty, and a neutral court perspective. This matters because effective negotiation requires understanding what the other side values. A clause that looks routine from your perspective might be a dealbreaker for the counterparty (or vice versa). By mapping each party's priorities in a strategic matrix, you can propose package deals that create value for both sides instead of fighting clause by clause. On The Legal Prompts platform, the Interest Toggle is a one-click feature that maintains analysis consistency across all perspectives.
Do these prompts work with models other than Claude (GPT-4o, Gemini, etc.)?
The prompt structures in this article will work with any capable LLM, but the results vary significantly. Claude's 1M token context window is essential for processing complete contract suites without chunking — GPT-4o's 128K context means you may need to split large documents. Claude's extended thinking produces more detailed reasoning chains than other models. And Claude's constitutional AI training makes it more reliably cautious about stating uncertain things as fact, which directly supports the anti-hallucination rules. We tested these prompts on GPT-4o and Gemini 2.5 Pro: both produced usable output, but Claude consistently delivered more detailed reasoning chains and fewer hallucinated citations.
Conclusion: Building Your Legal AI Practice
The six-stage workflow in this guide — intake, generation, clause review, risk analysis, multi-perspective analysis, and delivery — represents the current best practice for Claude-powered legal work. Each stage builds on the previous one, creating a compound effect where the final output is dramatically better than what any single prompt could produce.
The three principles that make it work are consistent across every stage: anti-hallucination safeguards that prevent Claude from inventing facts, Reasoning & Traceability that makes every conclusion auditable, and structured output that transforms AI analysis into actionable deliverables. Violate any of these principles and you get the kind of AI output that gives legal tech a bad name.
Whether you use these prompts directly or leverage The Legal Prompts platform which has them pre-optimized and ready to go, the important thing is to start building workflows rather than running one-off queries. The lawyers who are gaining the most from AI are not the ones asking Claude better questions — they are the ones building systems that chain those questions together into repeatable, reliable processes.
The future of legal practice belongs to the lawyers who build workflows. Start with Stage 1 on your next matter, and add stages as you gain confidence. Six months from now, you will wonder how you ever practiced without them.