A law firm's first AI agent should be small, narrow, and finished in under an hour. An agent is not a robot associate. It is a loop: it takes a goal, reads context, performs an action, checks its own work, and either continues or stops. The fastest way for a firm to build one is the stack I run my entire business on: Claude, Anthropic's AI model, doing the thinking, and Cloudflare Workers doing the running. No servers, no IT department, no code you have to write yourself, and the free tier covers your first agent entirely. This guide gives you the exact prompt, the way to test the result, and the mistakes that stall most firms before they ever ship anything.
The Robot Associate Fantasy Is Why Most Firms Never Start
When I talk to law firm owners about AI, almost every one of them is picturing the same thing: a digital associate that answers the phones, drafts the demand letters, manages the docket, and runs the tort campaigns while the partners sleep. That picture is exciting, and it is exactly why most firms are still stuck at zero. The gap between that fantasy and a blank screen is so wide that nobody takes the first step.
I build and run AI agents every single day, and every one of them lives on the same two pieces: Claude and Cloudflare. The Stable, the autonomous SEO agent that maintains this very website, is a Cloudflare Worker that wakes up on a schedule each morning, evaluates the site, asks Claude what should change, and ships approved actions to production without me touching anything. AI Jacob, the chat widget on this site, is the same pattern. None of those systems started as anything impressive. Every one of them started as a small loop that did one boring job correctly. That is the whole secret, whether you are an agency owner in Winter Garden or a managing partner running a hundred-lawyer shop.
What an Agent Actually Is (One Sentence, No Buzzwords)
An AI agent is a system that takes a goal, looks at its context, takes an action, verifies the result, and then decides whether to keep going or stop.
That is it. Goal, context, action, verification, decision. A chatbot answers once and disappears. An agent works through a task in steps, on its own schedule, and checks itself along the way. Every impressive agent demo you have seen is that same five-part loop with richer context and stricter verification. Once you have built the loop once, at any size, the intimidating vocabulary around agents shrinks back down to normal.
The Stack: Why Claude Plus Cloudflare
You need two things to run an agent: a brain and a body. Claude is the brain, the model that reads, reasons, and writes. Cloudflare Workers is the body, a place where a small program runs on the internet twenty-four hours a day without you owning a server, hiring IT, or installing anything. A Worker can be triggered on a schedule, called a cron trigger, which is what turns a one-time script into an agent that shows up for work every morning.
Why this stack and not the dozen platforms being pitched to law firms right now? Three reasons. It is nearly free at this scale; Cloudflare's free tier runs your first agent without a credit card conversation. It is yours; no vendor sits between you and your own system, and nothing you build gets held hostage by a subscription. And it scales; the exact same pattern that runs your first one-job agent is the pattern running $250M worth of advertising infrastructure at my agency. You will not outgrow it.
And here is the part that removes the last excuse: you do not write the code. Claude writes the code. Your job is to describe the job.
Your First Build: A Daily Briefing Agent
Do not build a research assistant for everything. Do not build an autonomous intake department. Build an agent that does one narrow job on a schedule and reports to you. My recommendation for a first build: a daily briefing agent. Every morning it fetches a handful of pages you care about, your firm's practice area news sources, a court's press page, a regulatory feed, asks Claude to summarize what changed and flag anything that needs a human, and emails you one clean briefing.
That sounds almost insultingly small, and that is the point. Small means you can understand it, test it, and trust it, and trust is the currency that lets you build the second agent.
The build takes three steps. First, create a free Cloudflare account and open the Workers dashboard. Second, go to Claude at claude.ai and use the prompt below to have it write your complete Worker file. Third, paste that file into the Cloudflare dashboard editor, add your Claude API key as a secret, set the cron trigger to run each morning, and deploy. That is the entire build. I deploy everything I run through the dashboard exactly this way.
The Exact Prompt
This is what you give Claude to have it build your agent:
"Write a complete, single-file Cloudflare Worker in JavaScript that acts as a daily briefing agent. On a scheduled cron trigger, it should: fetch the following URLs [list your three to five sources]; send the fetched content to the Claude API with instructions to summarize what is new, flag anything that requires human attention, and note any source that could not be read; then email the briefing to [your email] using the Resend API. Requirements: the file must be complete and ready to paste into the Cloudflare dashboard with no placeholders left for me to code; API keys must be read from environment secrets, never hardcoded; if any step fails, the agent must fail loudly by emailing me an error report instead of failing silently; and the briefing must end with a verification line confirming every source was checked. Then give me step-by-step dashboard instructions for adding the secrets and the cron trigger."
Look at what that one paragraph contains, because this is the anatomy of every agent you will ever build. It has a defined scope (these specific sources), a clear objective (summarize and flag), a concrete deliverable (one emailed briefing), and mandatory verification (confirm every source was checked, and fail loudly when something breaks). The verification and fail-loud requirements are the parts beginners always leave out, and they are the parts that separate an agent from a very fast guesser. A system that acts without checking is not saving you time. It is manufacturing risk, and in a law firm risk has a billing rate.
How to Grade the Result Like a Partner, Not a Fan
Here is where most beginners fail: the briefing arrives, the formatting looks clean, and they declare victory. You already know better than that, because you have reviewed first-year associate work. A memo can be beautifully formatted and completely wrong.
Grade your agent the way you would grade that associate. Did it actually check every source, or did it skip two and never mention it? Does the summary reflect what those pages really said, or does it sound plausible while missing the substance? Open the sources yourself for the first week and compare. Break it on purpose: feed it a dead URL and confirm you get the error email instead of silence. Would you act on this briefing? AI output is polished by default, so polish tells you nothing. Accuracy is the only test, and you can only verify accuracy because you kept the job small enough to check by hand. That is another reason your first agent should be boring.
When the output drifts, tighten the instructions and redeploy. Small loop, test, adjust, rerun. That cycle, repeated, is what building agents actually looks like. It is not magic. It is management.
A Note on Confidentiality, Because You Are a Lawyer
Do not wire your first agent to client files, your case management system, or anything privileged. Your first agent is a learning exercise, not a production system, and you have confidentiality obligations that a marketing consultant does not. Public web pages, your own marketing content, and internal non-privileged material are the right training ground. Build competence on safe material first, and understand exactly where data flows before anything sensitive ever touches the pipeline.
The Four Mistakes That Stall Law Firms
First, over-ambition. If your first agent needs ten integrations, memory, and a chain of sub-agents, you have not built a beginner project, you have built confusion wearing a nice suit. Second, vague goals. "Keep me informed" gives an agent no definition of success. "Check these five sources every morning and flag anything new" gives it a target it can hit or miss. Third, accepting silent failure. An agent that breaks quietly and keeps emailing you nothing is worse than no agent, because you will trust it. Demand the error email. Fourth, drowning in vocabulary. You do not need to master orchestration frameworks before you build. Build first. The vocabulary catches up on its own.
What to Build Second
Once the briefing agent works and you trust it, level up along the same pattern rather than jumping to something exotic. A docket-watch agent that checks specific court pages for filings in matters you track. A reputation agent that reads your firm's new reviews each week and drafts responses for approval. A marketing-recap agent that turns a month of campaign notes into a partner-ready briefing. A content agent that turns your rough notes into structured first drafts for client alerts. Notice that every one of these is still the same loop, still one Worker, still Claude in the middle. The context gets richer and the verification gets stricter, but the shape never changes.
Your first agent should feel almost disappointingly small. Mine did. The Stable started life as an agent that could read a list of pages and suggest one fix. Today it runs this site's SEO on a daily schedule. Small loops compound.
Go Deeper
This article is the one-hour version. If you want the full playbook for how a plaintiff firm should think about building AI, from agents and Workers to the infrastructure decisions behind them, I wrote an entire book on it: A Lawyer's Guide to Building AI. And if you would rather talk about what AI and smart advertising can do for your firm's tort campaigns, you know where to find me.
