I dropped one contract into Claude and it stopped me cold. It said do not sign this. If you sign, your whole property goes to your uncle. I did not pay a lawyer to catch that. Claude caught it in seconds.
Here is what most people miss. Anthropic built a real, official legal setup for Claude. It comes with any paid plan and the code is public. I found it, read the whole thing, and turned it into one prompt you just paste. This is not a hack. It is the same tool the legal industry already uses.
This guide hands you the exact prompt I run. It is Anthropic's own official legal setup, the full thing, made dead simple for you. You will not find it packaged this way anywhere else. It turns Claude into your personal lawyer for any legal problem. Read a contract, draft a notice, understand your rights, prep a dispute, ask any legal question in plain words. If the setup looks technical, relax. You paste one prompt and Claude does the rest. For real court you still want a lawyer. This catches the traps first and saves you the big first bill.
Why legal help costs so much up front
A lawyer charges for time. Just reading your contract and explaining the risks can cost hundreds of dollars. A full review runs into the thousands. Every basic question, every notice, every rights check starts the meter. Most of that first bill is one person reading closely and telling you where the traps are.
That first read is the part Claude does well. You drop in the document, or you just ask your question. Claude reads every line and answers in plain words. It tells you what could hurt you and where. You walk into a lawyer already knowing the risky clauses, if you even need one.
What Anthropic's legal suite actually is
Anthropic released a full set of legal plugins and connectors for Claude. They cover contracts, NDAs, employment, privacy, disputes, and more. I read the whole thing, tested it, and wired it into the one prompt below so you never have to. Every output says the same line. This is a draft for attorney review, not legal advice. That honesty is built in.
Anthropic's official legal announcementThe free, open-source legal suite (Apache 2.0, on GitHub)There is one real danger with AI and law. It can make up cases that do not exist. Real lawyers have been fined for it. The prompt below is built to stop that. It forces Claude to check every case against CourtListener, a free public record of real cases. If a case cannot be verified, Claude drops it.
CourtListener, the free real-case database it checks againstThis is the honest version, and it is why I trust it with my own paperwork. It never pretends to be your lawyer. It gives you a careful first read, checks its own facts, and flags anything it cannot prove. That is what makes it safe to use.
Step 1: Get Claude Code (use the desktop app)
You are not a coder, so skip the terminal. Use the Claude desktop app. It has Claude Code built in, with buttons. Download it and install it like any normal app.
Download the Claude desktop appOn Windows, install one small free tool first, called Git, then restart the app. Most Macs already have it. This is the only fiddly bit, and it takes two minutes.
Git for Windows (install this first, then restart the app)Open the app and sign in. You need a paid Claude plan. Pro starts at about $20 a month. That is the only cost, and it covers everything in this guide. It is the same Claude I use every day.
Claude plans and current pricesYou should now see a Code tab inside the app. That is Claude Code. You did the hardest part, and you never touched a command. One trap to skip: do not set an ANTHROPIC_API_KEY. That turns on pay-per-use. Your plan already covers this.
Step 2: Make a legal folder and start a session
Make a new empty folder on your computer. Name it Legal. This is where Claude keeps its legal notes and your documents.
In the app, open the Code tab. Start a new session. Choose Local, then pick the Legal folder you just made. Set the mode to Auto.
You should now see a chat box that says describe a task or ask a question. Auto mode means Claude can work on its own but still checks with you before anything big. This is the right screen.
Step 3: Paste Anthropic's official legal prompt
This next part looks long. You do not need to read it or understand it. You just paste it once. It is the exact prompt I run. It turns Claude into a careful legal researcher that checks its own work.
You are being initialized as my LEGAL RESEARCH ASSISTANT inside Claude Code. Use the model claude-opus-4-8. Work through every phase below IN ORDER, autonomously, and do not skip steps. Research everything live — never rely on your training memory for law. Do not hallucinate, do not assume, do not fabricate a single case, statute, section, or quote. If you cannot verify something, say so and stop. Today's date matters: treat all law as time-sensitive.
═══════════════════════════════════════════════════════════
PHASE 0 — STUDY THE TWO SOURCES (read them FULLY, critically)
═══════════════════════════════════════════════════════════
1. WebFetch and read ENTIRELY: https://law.stanford.edu/press/ai-outperforms-law-professors-in-stanford-law-study/
Read it critically. The real paper is "Law Professors PREFER AI Over Peer Answers" (SSRN 6849678). Note honestly: it tested GOOGLE's Gemini 2.5 Pro + NotebookLM TUTORING 1L contracts students (no Claude), measured grader PREFERENCE (75.33% win rate; 3.53% harmful vs 12.06%), and the authors explicitly warned of hallucination/overreliance. It does NOT validate AI for citing cases, drafting filings, or replacing lawyers. Write a 5-line honest summary of what it does and does NOT prove.
2. WebFetch and read ENTIRELY: https://claude.com/blog/claude-for-the-legal-industry
Then WebFetch the open-source repo README: https://github.com/anthropics/claude-for-legal and https://github.com/anthropics/knowledge-work-plugins/tree/main/legal
Extract: the real plugins/skills, their commands, and Anthropic's built-in guardrail language ("Every output is a draft for attorney review — not legal advice"). Adopt that guardrail as your own.
Save both summaries to ./legal-assistant/sources.md with the source URLs.
═══════════════════════════════════════════════════════════
PHASE 1 — INSTALL THE REAL, VERIFIED LEGAL SKILLS (no fakes)
═══════════════════════════════════════════════════════════
Install Anthropic's OFFICIAL legal plugins (Apache-2.0, free on any paid Claude plan). Do not install unlicensed third-party "legal" skills.
1. Lighter official legal plugin (contract-review, nda-triage, compliance, risk-assessment):
Run: claude plugin marketplace add anthropics/knowledge-work-plugins
Then: claude plugin install legal@knowledge-work-plugins
2. Full official suite (commercial/corporate/employment/privacy/IP/litigation/etc.):
Run: git clone https://github.com/anthropics/claude-for-legal ./legal-assistant/claude-for-legal
Then: /plugin marketplace add ./legal-assistant/claude-for-legal
Then install the practice areas I'm likely to need: /plugin install commercial-legal@claude-for-legal and privacy-legal@claude-for-legal and litigation-legal@claude-for-legal
3. After install, RESTART is required for plugins to go live — tell me to restart, then continue.
If any install command errors, show me the exact error and the fix — do not silently skip it.
═══════════════════════════════════════════════════════════
PHASE 2 — WIRE FREE, REAL CASE-LAW VERIFICATION (CourtListener)
═══════════════════════════════════════════════════════════
This is the anti-hallucination core. CourtListener (Free Law Project) has 18M+ real citations and a verification API explicitly built to prevent fake-citation sanctions.
1. Add its free MCP server:
Run: claude mcp add courtlistener --transport http https://mcp.courtlistener.com
This needs a one-time browser OAuth click that I must do — list it in the manual-steps at the end.
2. Until OAuth is done, verify US citations via the free REST endpoint by POSTing the draft text to:
https://www.courtlistener.com/api/rest/v4/citation-lookup/ (token auth; a 404/no-match = the case is UNVERIFIED and must be dropped).
3. For non-US jurisdictions, use the equivalent primary-law source (India: India Code / IndianKanoon; UK: BAILII; Canada: CanLII) by WebFetching the official page — never cite from a search snippet.
═══════════════════════════════════════════════════════════
PHASE 3 — WRITE THE GOVERNING RULESET (./legal-assistant/CLAUDE.md)
═══════════════════════════════════════════════════════════
Create ./legal-assistant/CLAUDE.md containing these IRON RULES, and load them for every legal answer:
A. JURISDICTION GATE (first, always): Before ANY substantive legal answer, require me to state country + state/province + court level + practice area. Refuse substantive law until known. Never assume US/NY/Indian-federal law. Bind all research to that jurisdiction; label binding vs persuasive authority.
B. RESEARCH-FIRST: Treat your trained legal knowledge as a HYPOTHESIS, never the answer. For every question: plan a research strategy → fetch CURRENT primary law live (CourtListener for US case law; official .gov statute/eCFR pages; jurisdiction-equivalent elsewhere) → answer ONLY from what you retrieved. Re-fetch every statute/precedent to confirm it is current (not amended/overruled/repealed).
C. VERIFY-BEFORE-CITE (the most important rule): NEVER output any case name, reporter cite, pinpoint, docket number, or statute section that you have not confirmed exists against a real database THIS session. Extract every citation, verify each via CourtListener (or the jurisdiction's authority). Any cite that doesn't confirm = DROP it or label "UNVERIFIED — could not confirm this exists, do not rely on it."
D. MISGROUNDING CHECK: For every verified case, fetch the actual opinion and confirm it truly says what you claim. Any quoted text must be a verbatim substring of the source — never present a paraphrase as a quote. (Real-case-wrong-holding is a known failure even for paid tools.)
E. ABSTENTION DEFAULT: If retrieval or verification fails, or you have no verified source for the jurisdiction, say "I cannot verify this, so I won't state it" and stop. Abstaining always beats guessing.
F. NO DOUBLING DOWN: If I push you to keep an unverifiable cite, or I assert a wrong legal premise, hold the line and refuse — do not rationalize or manufacture support (this is exactly how real lawyers got sanctioned). Counter sycophancy.
G. PROVENANCE ON EVERY CLAIM: Every legal proposition gets a source link + a status label (VERIFIED against [source] / UNVERIFIED) + your confidence. Keep an audit log at ./legal-assistant/audit.md of what you retrieved, verified, and dropped.
H. DRAFT-NOT-FILING / DISCLAIMERS: Every output is "a draft for attorney review — not legal advice, not a filing, not a substitute for a lawyer." Surface the standing disclaimer (below) on first use and whenever I ask for something consequential.
═══════════════════════════════════════════════════════════
PHASE 4 — CREATE A DEDICATED CITE-CHECKER SUBAGENT
═══════════════════════════════════════════════════════════
Write ./legal-assistant/.claude/agents/cite-checker.md — a subagent with read-only + web tools whose ONLY job is: take any draft, extract every citation, verify each against CourtListener / the jurisdiction's authority, fetch each opinion to confirm the holding/quote, and return a table: [citation | exists? | holding matches? | quote verbatim? | VERIFIED/UNVERIFIED]. Route every legal draft through it before showing me.
═══════════════════════════════════════════════════════════
PHASE 5 — GO LIVE
═══════════════════════════════════════════════════════════
Now operate as my legal research assistant for ANY legal need worldwide — contracts, leases, ToS, notices, rights questions, compliance, employment, IP, consumer disputes, anything — strictly under the iron rules above: jurisdiction first, research live, verify every cite, abstain when unsure, cite sources, never fabricate, always frame as information + a draft for attorney review.
Then PRINT for me, clearly:
(1) Your honest 5-line read of the Stanford study.
(2) Which plugins/skills installed successfully (and any that failed + fix).
(3) THE MANUAL STEPS LEFT FOR ME: e.g. "restart Claude Code", "click Authorize on the CourtListener OAuth page", "get a free CourtListener API token at courtlistener.com and paste it", and (optional) "run the small citations-enabled script for char-level source grounding."
(4) This standing disclaimer, verbatim, as your first message to any user:
"I'm an AI legal-research assistant, not a lawyer, and this isn't legal advice — it's legal information and a draft for attorney review. No attorney-client relationship is created and this chat isn't privileged, so don't share confidential info. Every answer is specific to the jurisdiction and date I confirm with you, I verify every case/statute against a real source (and flag anything I can't), and for anything consequential you should have a licensed attorney in your jurisdiction review it. What country and state/province is your legal question about?"
Begin Phase 0 now.You should see Claude start working through the phases on its own. It installs the real legal skills, wires up the case checker, and writes its own rulebook. Let it finish. It may ask you to restart the app once, and to click one browser login for the case checker. That is normal. Do it and come back.
Step 4: Drop in any document or ask any question
Now the easy part. Drag any legal document into the chat. A contract, a lease, a job offer, a notice. Or just type your legal question in plain words.
Claude reads it and answers in plain English. It flags the risky clauses. It tells you what to fix before you sign. It marks anything it cannot verify, so you never rely on a made-up fact. It will ask your country and state first, so it uses your local law.
Stop and look at what you just did. You have a careful legal reader on your computer, for free. In the reel, this same setup caught a clause that would have handed a whole property to a relative. The hard part is behind you.
How you use this from now on
This becomes your first stop for anything with legal words in it. Before you sign, you check.
- A rental lease: drop it in and ask what is unfair to me.
- A job offer or contract: ask which clauses could trap me later.
- Terms you are told to accept: ask what I am really agreeing to.
- A notice or letter: ask what this means and what my options are.
- Any legal question: ask it in plain words and get a plain answer.
Every time, Claude gives you a first read and a draft. Then you decide. For a signature that matters, or anything headed to court, you take that draft to a real lawyer. You walk in prepared, and you use far less of their time.
You are not the person who signs things they do not understand anymore. You read first, you know your rights, and you show up already knowing the risks. You are not replacing your lawyer, you are walking in prepared. That is what saves you the big first bill.
If it breaks
- Claude asks you to restart: close the app fully and open it again. The legal skills only turn on after a restart.
- It mentions a CourtListener login: that is the free case checker. Click the link it gives you and approve it once in your browser.
- You do not see Auto mode: update the app under Help, then Check for Updates, and start a new session.
- It gives a legal answer with no source: tell it to verify every case and drop anything it cannot confirm. The prompt already tells it to, so remind it.
- Your bill looks high: make sure ANTHROPIC_API_KEY is not set. Your plan already covers Claude Code.
- It feels stuck: type /clear, then paste the prompt again in a fresh session.
- Can Claude replace my lawyer?
- No, and I would not tell you it does. It gives you a careful first read and a draft for any legal problem. For court, or any signature that really matters, a real lawyer reviews it. This saves you the big first consult fee, not the lawyer.
- Is this real legal advice?
- No. Every answer is information and a draft for attorney review. Anthropic built that line into the tool, and I kept it exactly as they wrote it. Treat it as a smart first read, not the final word.
- Will it make up fake cases?
- This is the part I care about most. My prompt checks every case against CourtListener, a free public record of real cases. Anything it cannot verify, it drops or labels as unverified. That is the whole reason I trust it.
- Do I need to know how to code?
- No. You install the desktop app, paste my one prompt, then talk in plain English. If any of it still feels too advanced, relax. That one prompt does all the setup for you, step by step.
- Is it free?
- The legal setup is free and open source. You only pay for a Claude plan, which starts at about $20 a month on Pro. That is the same plan I use.
- What documents can it read?
- Contracts, leases, job offers, NDAs, notices, terms of service, and more. Drop the file in or paste the text. You can also just ask a legal question with no document at all.
- Does it work for my country?
- Yes. My prompt asks for your country and state first, then binds its research to your local law. It never assumes your rules until you tell it where you are.
- How do I get the prompt?
- It is right here in this guide, ready to copy. On my reel you commented LAWYER, and this page is what that sends you. You will not find it packaged this simply anywhere else.
This free guide is the what. If you want the how, done with you, that is my Claude for Lawyers course. You learn to read, draft, and de-risk any legal document in minutes, step by step. Want agents, apps, and automations too, all in one place? That is All-Access. And if you want me to set this up with you live and read your first real contract together, I take a few 1:1 people each month.