Claude is the AI tool I most often recommend to Canadian lawyers and the one I run my own business on. This post covers what it does well for legal work, the settings to check before client information goes anywhere near it, and how to get output worth your time.
The mental model
Claude is extremely capable predictive text. Take that seriously rather than as a dismissal, because it means there are only two levers that improve your results: a better model, or better context. When the output is bad on a top model, the context was bad, and in my experience that's usually a me problem. The lawyers who get good at this stop blaming the tool and start asking what they failed to provide.
That's also why "prompt engineering" is mostly dead vocabulary. There is no magic prompt. What matters is everything feeding the answer: your message, the documents you attach, the whole conversation history, the memory you've let it keep. Manage that, and the prompts write themselves.
What it does well for legal work
Big-document work. Claude holds an enormous amount of material in one conversation (full agreements, complete transcripts, a file's worth of correspondence) and works across it. One practical caution from the workshop: quality degrades well before the context window is technically full, and the tool won't warn you. Start a new conversation for each distinct task, and when a long thread is carrying something valuable, have Claude write a summary memo and take that into a fresh conversation instead of pushing on.
Repeatable workflows, through skills. A skill is a written procedure Claude follows: store the instructions for a task you do repeatedly, get consistent output, share it across the team so everyone's version runs the same way. This is the single highest-value feature I teach. My own CPD-accreditation skill takes a workshop topic, pulls the regulator's criteria and a past successful application, and drafts the new application — the last time I ran it I didn't edit a single word. The prerequisite: a skill captures a task you already know how to do. If you've never reviewed an NDA, the gap is mentorship, not software.
Working in your other tools, carefully. Out of the box, Claude can't touch anything: no files, no email, no calendar. Connectors grant access one tool at a time, and the guardrails are well designed. The email connector can draft but can't send or delete, you see every action it takes, and risky permissions default to needs-your-approval. Two rules from the workshop. Start restrictive and loosen as you build trust. And install connectors only from the official marketplace inside Claude: document management systems like NetDocuments and iManage are already there, and the marketplace is the difference between a vetted integration and some stranger's YouTube tutorial.
The confidentiality question
Lawyers store every client secret in cloud practice management and cloud document storage without losing sleep, then treat AI as uniquely radioactive. The duty is the same one you applied when you adopted those systems: reasonable steps to protect client information, which means reading the terms and controlling the settings. Here's what that looks like for Claude, as of June 2026:
On consumer plans (Free, Pro, Max): whether your chats can be used for model training is a setting you choose. Turn "Help Improve Claude" off (Settings, then Privacy) before anything client-related goes in. Off, conversations are retained about 30 days. On, training data can be retained up to five years. While you're in there: the thumbs-up/down feedback buttons share that conversation regardless of your setting, so don't use them.
On commercial plans (Team, Enterprise) and the API: no training on your data by default, with admin-controlled retention. The commercial terms also incorporate Anthropic's Data Processing Addendum (DPA), a contractual data-protection commitment the consumer plans don't carry. That's why I recommend the commercial terms for every lawyer using Claude for work. Team runs US$20–125 a seat depending on billing and seat type (the premium seat carries 5x the usage), with a five-seat minimum. Anthropic publishes its compliance posture (SOC 2, ISO 27001) at trust.anthropic.com, and firms with data-residency requirements can run Claude through AWS Bedrock or Google Vertex for location control.
Two more habits round it out. Document the steps you took (the terms you read, the settings you verified), because reasonable steps you can't evidence might as well not have happened. And address AI in your retainer the way you address any other technology: commit to taking reasonable steps. The common mistake is promising more than the regulator asks, like "we only use tools that guarantee the safety of your data." Nothing guarantees anything, and you don't need it to. Your law society's expectations are the floor either way: I've summarized the LSO's guidance, and a firm AI policy makes the whole posture explicit.
Getting output worth your time
The workshop spends most of its hours here, because the tool is rarely the problem. The habits that move the needle:
Frame neutrally. You'd never run the Google search "do candles cause cancer?" because you know what it returns. Claude reads your framing the same way, like a junior bracing for a leading question. Ask "find support for the argument that remote workers are more productive" and you'll get factum fodder. Ask "is there any relationship between remote work and productivity, based strictly on published studies?" and you'll get an answer you can use.
Say why you're asking. "Pull the key facts from this affidavit" gets a list. Add "I'm asking because I'm preparing cross-examination and want inconsistencies with prior testimony" and the same request comes back organized for the job. The phrase I teach: "I'm asking because…"
Plan before you execute. The old habit is a quick prompt followed by twenty minutes of fixing weak output. Move that time to the front: have Claude interview you about the task ("ask me follow-up questions, multiple choice") before it produces anything. The editing time doesn't disappear, it shifts to planning, and the plan is reusable where the edits never were.
Know when to start over. A bad answer stays in the conversation and quietly feeds every later one. When a thread goes sideways, don't argue with it. Take what's useful and open a fresh conversation.
Verification is non-negotiable throughout: Claude is not a citation database, it will sometimes be confidently wrong, and Canadian courts have already sanctioned lawyers who filed AI-invented cases. Nothing leaves your desk unchecked.
What about ChatGPT and Gemini?
I get asked at every session. Claude's models are strong, and the desktop app currently has the features that make AI feel most usable for lawyers. That's a June 2026 statement, not a permanent one. A lawyer who gets good at this can switch between tools without issue, and should be ready to, because models change and different companies make different advances. Everything in this post about context, framing, and verification moves with you. Consumer pricing is comparable across the major tools (Claude Pro is US$20 a month), so the decision was never about price.
Where to start
The first hour, in order: set the privacy toggle deliberately instead of leaving it wherever the signup flow put it. Turn memory on so the tool accumulates context about how you work. Run one real task end to end with the habits above, and verify the output against the source. Then take the best conversation you've had and tell Claude to turn it into a reusable skill, so next time starts where this time finished.
That loop is the whole skill, and the hands-on version is the live training I run for Canadian lawyers and their staff, including the four-hour Master the Claude Suite workshop where you build your own setup on your own files. Inn Laws members attend free. Everything above is accurate as of June 2026 and gets re-tested every time I teach, because this field doesn't sit still.