How to Use Claude Like the Top 1% of Lawyers

Here is how most lawyers use Claude.

Open it. Paste a clause. Ask a question. Copy the answer. Close it.

Every session starts from zero. Claude knows nothing about your practice, your jurisdiction, your client, or how you draft. So it gives you something generic. And generic is exactly what a good lawyer can't use.

The lawyers getting genuinely different results do something else. They treat Claude the way they treat a new associate. They onboard it once. They set the rules once. Then the returns compound on every matter after.

This is that system, built for legal work. Files, prompts, agent workflows, and the habits that separate real leverage from party tricks.

One thing first, because it is the thing that makes this a lawyer's guide and not a generic one.

The three rules you never break

AI changes how fast you work. It does not change what you owe your client.

• Confidentiality and privilege come first. Know where your data goes before you send it. Privileged or client-identifying material belongs only in a governed, firm-approved setup — never pasted into a personal account or an ungoverned tool.
• Verify everything the model asserts. Citations, quotes, section numbers, figures. Claude can be confidently wrong. Courts have already sanctioned lawyers for fabricated case citations. You sign the work. You own every word.
• You are the partner. Claude is the associate. It can draft, compare, summarise, and pressure-test. It cannot exercise legal judgment. That never delegates.

Hold those three, and everything below makes you faster without making you reckless.

Part 1: Onboard Claude like you'd onboard an associate

Set this up once. Benefit on every matter after.

The idea is simple. Claude reads a small set of files before it starts work. Those files tell it who you are, how you draft, and what good looks like. You stop re-explaining yourself. It stops starting from scratch.

Four files are worth building first.

• practice-profile.md — who you are before every task. Your practice area and seniority. Your jurisdiction, or jurisdictions. The kinds of clients you act for. The positions your firm has already settled, so Claude builds on them instead of relitigating them. What you are focused on this quarter.
• drafting-voice.md — how you write. British or American spelling. Whether you use "shall." Your defined-terms convention. Plain-English client notes versus formal deeds. The fastest way to build this: ask Claude to interview you. Tell it to play a sharp editor and ask hard questions about how you draft and why. You will surface conventions you follow without noticing.
• never-do-this.md — taste is what you reject. List the tells you never want to see when Claude writes as you. The empty throat-clearing ("It is important to note that"). The robotic hedging. The over-bolding. And the legal-specific ones: never invent a citation, never use US contract terms in an Indian document, never soften clear advice into mush.
• Your matter folders — the structure Claude reads when it works on your machine. Keep it boringly consistent: a MATTERS folder with one subfolder per matter (the brief, key documents, prior drafts), a PRECEDENTS folder with your firm's templates (NDA, MSA, SHA, term sheet, SAFE), and an OUTPUTS folder as the only place Claude delivers finished work.

Then set your standing rules once, so they run on every task: read the practice profile before starting, open the matching matter folder, never fabricate authority, flag anything uncertain rather than papering over it, and save output to the OUTPUTS folder.

A morning to build properly. The compounding starts the same day.

Part 2: The prompting shifts that change legal outputs

Prompting is not about a magic phrase. It is about giving Claude the structure to reason like a lawyer would.

A useful test: would a junior who has never seen the matter understand your instruction? If they would be confused, so will Claude.

Make Claude ask before it drafts. This is the single highest-leverage habit. Instead of ordering an output, ask Claude what it needs to know first.

I need to review this agreement so that we protect [CLIENT]
as the [buyer / vendor]. Read the matter folder, then ask me
the questions you'd need answered before you start. Don't draft
anything until we've agreed the approach.

Claude comes back with the right questions. Are we buyer or vendor? What is the client's risk appetite on liability? Is there a data-processing addendum? Governing law? You answer. It proposes an approach. You approve or redirect. Then it works. This one loop prevents more bad output than anything else.

Structure complex prompts with tags. For a real review or memo, separate the parts so Claude processes each cleanly.

<facts>       the deal, the parties, what's already agreed
<task>        what you want produced
<constraints> jurisdiction, length, who reads it, what to avoid
<precedent>   the playbook or template to follow

Give it a role, an audience, and the stakes. "You are a senior M&A associate. This memo goes to the general counsel of a listed company. Prioritise accuracy and flag every assumption" produces a completely different result from "summarise this." The model responds to framing. Tell it why the work matters and who will read it.

Build Skills for the work you repeat. A Skill is a saved workflow Claude triggers automatically when the task matches. Teach it once, refine it to your standard, and it stops asking you to re-explain. The obvious candidates: first-pass NDA markup, clause-by-clause review against your playbook, a standard due-diligence summary format, your house style for a board note. Claude's Word integration can run exactly this kind of skill — drafting, redlining, and comparing clauses against an approved playbook while tracking every change and its reasoning.

Part 3: Agent workflows most lawyers miss

Claude Cowork is Anthropic's desktop agent for non-coding work. It reads and writes files in folders you permit, connects to your apps, and runs multi-step tasks end to end. (It is in research preview on the paid plans, so check what your account includes — and only point it at firm-governed data.)

A few moves separate the power users.

• Describe the outcome, not the steps. Don't say "open this file, copy that clause." Say: "Review this vendor MSA against our playbook and produce a redline plus a one-page risk memo flagging liability, IP ownership, and data protection." Claude plans, you approve, it executes.
• Run due diligence in parallel. Reviewing thirty NDAs one by one is an evening. Framed as a batch — "produce a one-line summary of the term, governing law, and any off-market clause for each of these" — Claude spins up sub-agents and does them at once. Minutes, not hours. You review the exceptions, not every page.
• Schedule the recurring work. A weekly sweep of regulatory updates that matter to your clients — DPDP, SEBI, MCA, RBI. A Monday matter-status note. Tomorrow's hearing or call prep, drafted overnight. You describe the job once and it runs on a cadence. One honest limit: scheduled tasks only run while your computer is awake and the desktop app is open, so they fit a machine that stays on, not a closed laptop.
• Stack your tools. Connected to the firm's systems, Claude can cross-check a deal's closing checklist against the executed documents and the email trail, and surface the commitment that never made it onto the list. The value compounds with each system you connect — which is also why each connection deserves a real privacy check first.
• Use Projects to keep matters clean. Give every matter or client its own Project — its own files, instructions, and memory. This is not just tidiness. It stops Client A's confidential assumptions from bleeding into Client B's work. For a lawyer, separation of context is a confidentiality control, not a nice-to-have.
• Hand work off from your phone. You can assign a task from the Claude mobile app and let your desktop do the work, then walk into a finished first draft. Queue the review on your commute. Read the markup at your desk.

Part 4: Context tricks that change how Claude reasons

A long, cluttered session does not just run slower. It reasons worse. Your instructions get buried under everything that came before, and the model starts making mistakes it would not make with a clean window. When accuracy is the whole job, that matters.

Start fresh for each new matter or topic. Counterintuitive, but true. A clean session outperforms a bloated one. Open a new conversation when you switch matters or when quality starts to slip.

Write a handoff before you start fresh. Mid-draft and need a new session? Ask for a clean baton:

Summarise what we've drafted, the decisions made, the open
points, and the next step — written so a fresh Claude with no
prior context can pick up exactly where we left off.

Save it to the matter folder. Load it next time. Clean context, full continuity.

Tell Claude the context it's working in. "You're in a Cowork session with access to this matter folder; save the redline to OUTPUTS." One line removes a whole class of avoidable errors.

The shift underneath all of this: the leverage is no longer in clever wording. It is in what you load. The right precedent, the right facts, the playbook. Structure beats phrasing.

Part 5: Four legal use cases worth going deep on

• Contract review and redlining. The highest-volume win. Claude marks up against your playbook, explains each change, and produces the risk memo alongside the redline. You move from drafter to reviewer.
• Due diligence at scale. Batch review of NDAs, DPAs, and vendor agreements; issue lists; data-room synthesis. The work that used to swallow a team of juniors becomes a structured pass plus your judgment on the exceptions.
• Research and memos — with the brakes on. Use Claude to structure the analysis, draft the memo, and pressure-test your reasoning. Do not use it as a source of authority. Every case, statute, and section it cites gets confirmed independently before it leaves your desk. Treat it as a brilliant associate who occasionally misremembers — because that is exactly what it is.
• Founder-facing work. Plain-English explainers of a term sheet. A SAFE-versus-priced-round breakdown. A cap-table scenario a non-lawyer founder can actually follow. This is where good lawyers win trust — and where Claude, given your voice file, drafts something that sounds like you and lands for them.

The one shift that ties it together

Most lawyers still use Claude the way they used a search engine in 2015. One question, one answer, start over tomorrow.

The lawyers pulling ahead build a system instead. Context files that load automatically. Skills that trigger on the right task. Reviews that run in parallel. Sweeps that run overnight. The repeatable work handled, so their hours go to the part no machine can do — the judgment, the strategy, the client in front of them.

None of it replaces the lawyer. It removes everything that was never lawyering in the first place.

That system is not a download. It is something you design — the files, the permissions, the guardrails, the verification step that never gets skipped. Build it deliberately and the gap between you and the lawyer still pasting clauses into a blank box widens every single week.

Adapted for legal practice from Ruben Dominguez's "How to use Claude like the top 1% of users" (The AI Corner). Product features described here were checked against Anthropic's documentation and are accurate as of writing; capabilities and plan availability change often, so confirm the current state before relying on any specific feature.