Harvey AI is $14,400 per attorney per year. Most lawyers don't need that.
Harvey AI, the most-funded legal AI product on the market, charges roughly $1,200 per attorney per month. CoCounsel from Thomson Reuters runs $150-400 per seat. Spellbook for contracts charges $249/month. Paxton hits $499/month. The entry-level Legal Prompts tool charges $49/month, recurring.
If you are a solo attorney or a partner at a 5-15 lawyer firm, you have probably looked at those prices, looked at your matter mix, and decided that the math does not work.
Here is what those tools are actually selling. Not magic. Not unique training data (most run on the same large language models from Anthropic, OpenAI, or Google that anyone else can call). What they are selling is prompt patterns — the carefully-structured instructions that take a generic LLM and turn it into something useful for a specific legal task.
This post breaks down what those prompt patterns look like, why they are worth paying for, and why $49 lifetime is a fair price for the prompts themselves once you remove the proprietary UI markup.
What does Harvey AI actually do?
Harvey AI markets itself as a domain-specific assistant for elite law firms. Strip the marketing and the actual feature set across legal AI tools is consistent:
- Contract review. Take an inbound contract, identify load-bearing clauses, flag deviations from market, suggest redlines.
- Drafting assistance. Take a fact pattern and produce a first draft of a complaint, motion, brief, memo, or contract.
- Legal research synthesis. Take a research question and produce a structured analysis with case authority (though hallucination risk remains real and every citation must be verified).
- Discovery and document review. Process large document sets to surface relevant material.
- Client communication drafting. Convert raw notes into client-readable correspondence.
The SaaS tools wrap these capabilities in a UI, add some collaboration features, and bill monthly. The underlying engine is the same as the public Claude or GPT models you can already access.
The value-add of the SaaS, when there is one, is the prompts — the structured way they ask the model to do legal work. That's what makes Harvey's contract review feel different from asking Claude.ai "review this contract."
What makes a legal AI prompt actually work?
After writing and testing 50 prompts for the Legal Practice + Claude Code Prompt Pack, the patterns that consistently produce work-quality output share seven structural elements. These are not specific to legal work — they apply to any high-stakes AI prompting — but they matter most when the output will be reviewed by another attorney or filed in court.
1. Role anchoring. The prompt opens with what Claude IS, not what Claude should DO. "You are a senior backend engineer reviewing a junior dev's pull request" produces dramatically different output from "review this code." In legal context: "You are a senior litigation partner drafting a motion to dismiss under Federal Rule 12(b)(6)" anchors voice, depth, and scope before any instruction lands.
2. Context fences. Explicit declaration of what Claude has and what Claude does not have. The single most common reason Claude hallucinates is that it assumes it has information it does not. "You have: the complaint. You do not have: the test suite, prior pleadings, or the docket. If a finding requires any of those, name the missing context and stop."
3. Output contract. Name the exact format Claude must return. Not "give me your analysis" but "return three sections in this order: (1) one-paragraph summary, (2) bulleted list of findings tagged [CRITICAL] [WARNING] or [INFO], (3) one-sentence verdict."
4. Failure mode declaration. Tell Claude what NOT to do with one specific example. "Do not invent case citations. Example of bad output: 'See Smith v. Jones, 442 F.3d 511 (3d Cir. 2022)' when no such case exists. Example of good output: 'I would cite a case for this point but my training data does not let me verify the citation; you should verify against Westlaw.'"
5. Reasoning scaffolding. Decide whether you want plan-first or answer-first reasoning. For legal drafting, plan-first almost always wins: "Before drafting, briefly outline the elements you'll address. Then execute."
6. Verification handles. Build the prompt so it produces output the prompt can self-check against. For a brief: "After the argument section, in a separate block, name one assumption that could change the outcome if it turned out to be wrong."
7. Iteration anchors. Leave named hooks the user can later edit. [VOICE: senior backend engineer], [DEPTH: line-by-line], [JURISDICTION: California]. When the prompt isn't doing what you want, you edit one bracketed handle instead of rewriting.
These seven patterns are detailed in the seven-patterns post with worked before/after examples. Every prompt in the Legal Pack is built on this framework.
Three example prompts from the Legal Pack
These are three of the 50 prompts in the pack. Each shows the structure that separates a working prompt from a generic ask.
Example 1: Initial client intake interview
The goal is a structured intake that surfaces the issues a paralegal would catch, without drifting into pre-engagement advice that could create an implied attorney-client relationship.
You are a senior attorney conducting an initial client intake. Be professional but warm.
You have: a prospective client wanting to discuss [matter type].
You do not have: any documents yet, jurisdiction confirmation, conflict-check results, or fee discussion authority.
Conduct a 12-question intake covering: (1) the precipitating events and dates, (2) the parties involved (and party affiliations that could create conflicts), (3) the relief the client is seeking, (4) any prior counsel involved, (5) statute-of-limitations exposure, (6) documents the client should preserve immediately, (7) communication preferences and confidentiality concerns.
Output format: numbered question list followed by a 3-line "what I am hearing" summary the attorney can read back to confirm understanding.
Do NOT give legal advice in this intake. Do NOT commit to representation. Flag any answer that requires immediate counsel attention with [URGENT].
Why this works: the [URGENT] flag pattern surfaces SOL and document-preservation issues without the lawyer having to remember to ask. The "do not give advice" instruction prevents the prompt from drifting into territory that creates an implied attorney-client relationship.
Example 2: Contract indemnification clause analysis
Indemnification provisions are often the load-bearing risk allocation in a contract. They are also where most reviewers either skim or over-lawyer.
The prompt forces explicit treatment of the "arising from" vs "related to" vs "in connection with" distinction — three phrases that look interchangeable and are construed very differently across jurisdictions. Different states have produced wildly different case law on these phrases, and the difference can change the financial exposure by an order of magnitude.
The prompt also forces a "silent points" review — what the clause does NOT address that it should. Settlement consent rights, notice timing, mitigation duty, defense control. These are the things that turn an indemnification dispute into a separate lawsuit.
Example 3: Motion to dismiss drafting
The full prompt is in the pack. The structural insight is the "if I were the judge" memo. After drafting the brief, the prompt asks for a separate one-page memo identifying the strongest objection the court might raise to the motion, even if plaintiff does not.
Most motion-to-dismiss drafts get DENIED on grounds the lawyer did not consider. Forcing the prompt to surface the judge's likely objection — and forcing the lawyer to address it before filing — is what separates a motion that wins from one that gets briefed and lost.
What the pack contains
The Legal Practice + Claude Code Prompt Pack ships 50 prompts across four roles:
Solo / General Practice (12 prompts): initial client intake, conflict-check email, fee agreement drafting, demand letter, withdrawal letter, "I am not your lawyer yet" prospective-client follow-up, SOL checklist by claim type, scope-creep redirect, client status update, missed-deadline apology, initial case-strategy outline, final client closeout letter.
Litigation (13 prompts): complaint drafting against specific causes of action, interrogatories, requests for production, requests for admission, deposition outline, motion to dismiss, summary judgment brief, trial brief, settlement / mediation brief, jury instruction drafting, witness preparation script, end-of-day trial brief, post-trial motion (JNOV / new trial).
Transactional (13 prompts): MSA review from party A vs party B perspective, NDA generation (mutual vs one-way), employment agreement review for problematic clauses, indemnification clause analysis, term sheet markup, SaaS / software license review, M&A diligence checklist, partnership / JV structure analysis, AI / data compliance audit, vendor / supplier contract review, contract drafting from a deal memo, termination / wind-down letter, LOI / non-binding term sheet drafting.
Practice Management (12 prompts): time-entry narrative from terse notes, billing dispute response, calendar / deadline audit, paralegal task delegation, client onboarding kit, SOL reminder for closed matters, privilege log creation, file-closure / records-retention review, conflict-of-interest disclosure letter, bar grievance / disciplinary response, fee-increase notification, year-end matter review.
The full pack is at clskillshub.com/pack/legal.
Why $49 lifetime vs $1,200/month
The $49 lifetime price is not a discount on Harvey. It is a different business model.
Harvey, Spellbook, CoCounsel, and Paxton charge monthly because they pay for: (a) ongoing access to large language models via API, (b) proprietary UI and integrations, (c) sales, marketing, and customer success teams, (d) investor returns on hundreds of millions of dollars of venture capital.
Claude Code is the official Anthropic CLI. You install it once. You bring your own Claude or Anthropic plan. The prompts in the pack ride on top — Claude Code skills are just structured Markdown files. You install them, they show up in your Claude Code sessions as @legal, you invoke them by name.
The pack ships the prompts. You bring the model access. There is no recurring subscription on the pack itself.
This only works if you are willing to use Claude Code (the CLI), or to copy-paste the prompts into Claude.ai, the API, or another tool. If you require a managed SaaS UI with logo, login, and a customer success manager, Harvey is still your option. The pack is for attorneys who would rather own the patterns than rent the wrapper.
Disclaimer (real, not boilerplate)
This pack is a drafting assistance tool, not legal advice. Every output produced by these prompts must be reviewed and edited by a licensed attorney before being sent to a client, filed with a court, or used in any matter of substance.
Claude will sometimes hallucinate case citations. Verify every citation against Westlaw, Lexis, or Bloomberg Law before relying on it. For privileged communications, run the prompts only on systems where the conversation is not being mined for training data — Anthropic offers HIPAA / enterprise tiers with proper agreements; consumer Claude.ai does not have those agreements.
The pack is for your assistance. You remain the attorney.
Who should buy this
Solo attorneys and small-firm partners (1-15 lawyers) who:
- Are already using Claude or another LLM ad hoc for drafting assistance
- Want structured prompts so the output is consistent across matters
- Do not want to commit to $250-1,200 per month per seat for a SaaS wrapper
- Are comfortable using Claude Code (CLI) or copying prompts into Claude.ai
- Will review every output before it leaves their desk
The pack is NOT for: large firms that need single-sign-on, matter-management integration, and audit trails — those firms should stay on Harvey or CoCounsel. Or attorneys who want a managed-SaaS experience and do not want to manage prompts as files.
How to actually use the pack
Three options:
1. Copy-paste into Claude.ai or the API. The prompts work in any Claude conversation. Pick the prompt, paste it, fill in the bracketed inputs.
2. Install as Claude Code skills. Drop the skills-install/legal/ folder into ~/.claude/skills/, restart Claude Code, type @legal in any session.
3. Treat as a reference. The Markdown is searchable. Ctrl+F for "discovery", "NDA", "motion to dismiss" and the relevant prompt is one search away.
Related reading
- The 7 Patterns Behind High-Performance Claude Prompts — the prompt-engineering framework every prompt in the pack uses
- Claude Code This Week — recent Claude Code feature updates
- Claude for Small Business Review — Anthropic's broader SMB AI play and how it compares
- Pricing page — all 7 industry packs
- Legal Practice Pack on clskillshub.com
FAQ
How does this compare to Harvey AI specifically? Harvey is a managed SaaS at roughly $1,200/seat/month. The Legal Pack is a one-time $49 purchase of the underlying prompt patterns. Harvey's UI, support, and enterprise features are not in the pack — but the prompt structure that does most of the actual work is. If you are paying for the UI and support, stay on Harvey. If you are paying for the prompts, switch.
What about hallucinated citations? Every legal AI tool, including Harvey and CoCounsel, can hallucinate. Always verify case citations against Westlaw or Lexis. The pack includes disclaimers and structural prompts that flag uncertainty rather than hide it.
Can I use this for billable work? That is your judgment as the attorney. The pack produces drafting assistance that you review and edit before it leaves your desk. Most attorneys can ethically bill for time spent reviewing and editing AI-assisted work product at the same rate as time spent reviewing and editing a junior associate's work.
Is this HIPAA-compliant? The pack itself is just text. Whether your use of it is HIPAA-compliant depends on whether you are running it on a system with a BAA. Anthropic offers enterprise tiers with BAAs. Consumer Claude.ai does not.
Lifetime updates? Yes. When v1.1 ships with additional prompts and refined patterns, the same unlock link serves the latest version.
Refund policy? Digital product, all sales final. If something genuinely does not land for you, reply to your purchase email and I will add you to the full Skills Library (lifetime access) as a goodwill gesture.