Claude is the AI assistant I recommend lawyers start with: its drafts sound less like AI, it reads about 500 pages in one conversation, and Pro costs about $20 a month.
I own 10 law firms. My flagship, a national lemon law firm with 62 staff, runs on systems I check for about 30 hours a week. I touch zero daily operations, and the single tool I open most is Claude.
Not because I’m loyal to a brand. Because I’ve run the same drafting tasks through Claude and ChatGPT side by side for two years, and Claude wins the ones that matter for legal work. Here’s the full picture: what it does, what it costs, and exactly how my firms use it.
Why Claude over ChatGPT for legal work
Three reasons, in order of how much they matter.
First, the writing. Claude’s drafts sound less like AI. ChatGPT loves the tics that make clients and opposing counsel roll their eyes: everything is a “journey,” every list has exactly three items, every email closes with false warmth. Claude still needs editing, but I edit less. When a client email goes out under my firm’s name, that gap is real money.
Second, long documents. Claude reads about 500 pages in a single conversation (200,000 tokens, in technical terms). A lemon law file with two years of repair orders, a purchase contract, and a warranty booklet fits in one chat. You ask questions against the whole file instead of feeding it chunks and hoping the AI remembers chunk one by chunk nine. When a matter outgrows even that, NotebookLM handles bigger source sets.
Third, instructions. Legal work is full of nuance: “cite the statute but don’t quote it,” “firm but not threatening,” “plain English, eighth grade reading level, no legalese.” Claude follows layered instructions like that more faithfully. ChatGPT tends to drift back to its defaults after a few exchanges.
I still keep a ChatGPT subscription. It’s a decent second opinion and its web search is good. But when the output is going in front of a client or a court, I start in Claude.
Claude Projects: the feature that changes everything
A Project is a workspace with two things a normal chat lacks: permanent instructions and uploaded files. Everything you put there is available in every conversation inside that Project, forever.
This fixes the biggest waste in how lawyers use AI: re-explaining the firm every single chat. Who you are, what you practice, how you write, what your fee structure is. In a Project, you say it once.
Here’s a real one from my lemon law firm. The Project holds our intake criteria: qualifying repair-attempt thresholds by state, mileage cutoffs, which manufacturers we take, what a dead-on-arrival case looks like. The instructions tell Claude to read an intake summary and return a strength assessment with the specific criteria the case passes or fails. Intake staff paste in raw notes and get a structured evaluation in seconds. No prompt engineering, no re-explaining lemon law. The Project already knows.
Your first three Projects should be: one loaded with your fee agreement and engagement letters, one with your five most-used letter templates, and one with your intake criteria. That took me under an hour to set up and it pays out daily. The step past Projects is connecting Claude to your live systems through MCPs, so it reads your email and case data instead of waiting for uploads.
Artifacts: documents you can actually use
When Claude generates something substantial, a letter, a table, a checklist, even a small working web tool, it opens in a side panel called an Artifact instead of scrolling by in the chat. You can edit it, ask for revisions, and export it.
For lawyers this matters because drafting is iterative. “Make paragraph two firmer. Add the statute cite. Cut the last sentence.” The document updates in place while the conversation continues next to it. I’ve also had Claude build small interactive tools as Artifacts: a settlement calculator, a deadline checker. No code knowledge required. You describe it, it appears, it works.
The three models, in plain English
Claude comes in three models and the picker confuses everyone for about a day.
Opus is the deepest reasoner. Use it for complex analysis: multi-issue memos, dissecting a dense contract, anything where you’d assign your smartest associate.
Sonnet is the daily workhorse. Fast, strong, and what I use for 90% of tasks: emails, summaries, letters, intake evaluations.
Haiku is the fast, cheap one. It matters mostly if you build automations that make thousands of calls, where speed and cost compound. For chat, you’ll rarely pick it.
Default to Sonnet. Switch to Opus when the question is genuinely hard.
What it costs
The free tier is real but you’ll hit usage limits within a serious workday. Fine for kicking tires, not for practice.
Pro runs about $20/month and covers most lawyers completely. That’s the plan I’d start on.
Max plans ($100 to $200/month) exist for heavy users: people running long documents through Claude all day or using Claude Code. A few of my staff are on Max. Most are on Pro.
| Plan | Price | Who it fits |
|---|---|---|
| Free | $0 | Testing only; limits hit fast |
| Pro | ~$20/month | Most lawyers’ daily use |
| Max | $100 to $200/month | Heavy users and Claude Code |
One line that matters more than price: paid plans do not train on your data. Your client conversations are not teaching the model.
What my firms actually use it for
Four uses carry most of the weight at CPLG:
Client emails in firm voice. A Project holds tone examples, staff paste in the situation, and the draft comes back sounding like us. Editing time dropped from writing time to review time.
Repair history summaries. Two hundred pages of dealer service records become a chronological table of visits, complaints, and days out of service. That table used to be hours of paralegal time per file.
KPI brief drafting. My daily numbers brief across the firms starts as a Claude draft against raw data. I read it before 10:30 AM every weekday.
Plain-English explanations. Clients don’t want a statute recited. “Explain the civil penalty under the Song-Beverly Act to a frustrated client at an eighth grade level” produces something a human being can actually read.
The confidentiality section, because you were going to ask
Three rules cover it.
Paid plans don’t train on your data. Free tier policies differ, which is one more reason not to practice law on a free account. Firms with stricter needs can get commercial terms from Anthropic with stronger guarantees.
Verify everything. The real malpractice risk isn’t a data leak, it’s the lawyer who filed an AI-drafted brief with invented citations. For case law, use AI legal research tools built on real databases, not a chatbot’s memory. Nothing Claude writes touches a client file until a lawyer has read it. That rule has no exceptions at my firms.
Check your state bar. Most bars now permit AI use under normal supervision and confidentiality duties, and several have published specific guidance. Read yours once. It takes 20 minutes.
Getting started this week
Skip the research phase. Do this instead:
- Sign up at claude.ai and pay the $20 for Pro.
- Create one Project. Name it after your firm.
- Upload three documents: your fee agreement, one letter template, your intake criteria (or whatever three documents define how your firm works).
- Draft one real client email in it today. Not a test. A real one you were going to write anyway.
That last step is where most lawyers stall, and it’s the only one that builds the habit. If you haven’t read the full roadmap, this is stage one of five. Once the habit sticks, steal my prompt library, and when you’re ready to build instead of chat, Claude Code is waiting.
Open Claude. Write the email. That’s the whole assignment for today.