Claude Cowork for Real Estate Agents: A Complete Beginner's Guide (2026)

I get some version of this every week: “I signed up for Claude, I typed a question into the box, it gave me a nice paragraph back, and I still don’t see how this runs my business.” That’s a fair complaint, and it’s almost always the same problem. You’re in the chat window. Chat gives you an answer. Claude Cowork for real estate agents is the part that actually does the work, and it’s a completely different experience once you set it up right.

I’m William Zhang, a real estate agent in Austin, Texas, and the founder of Real Estate AI Society. I’ve been running my practice on Claude for about a year now, and Cowork is the piece I’d hand a brand-new agent first. Not because it’s the flashiest thing Claude does: it’s the one that gets you from “AI is interesting” to “something happened while I was at a showing.”

So this is the beginner’s walkthrough, step by step, the same way I ran it on camera. We’re going to download the desktop app and pick a plan, understand the difference between chat and Cowork, build one project from scratch (an executive assistant I named Christine), give her the four things every project needs, connect Gmail and Google Calendar, set the permissions so she can’t do anything you didn’t approve, and then run a real task on a real transaction.

That last part is the reason I made the video. I asked Christine to look at my deal at 1117 Terrace View, find every contract sitting in my inbox, read them, and put the deadlines on my calendar. She went and did it: pulled the contracts, researched how TREC contracts actually work so she’d understand what an option period is, derived the third-party financing deadline and the closing date, and wrote them onto my Google Calendar while I stepped away from the desk. That’s transaction coordinator work, running unattended, on a $20 a month plan.

If you’re brand new to Claude entirely and haven’t installed anything yet, start with my complete beginner’s guide to Claude for real estate agents first, then come back here. Everything below assumes you’re ready to build something.

Claude Cowork for Real Estate Agents: Chat Answers, Cowork Does the Work

Inside the Claude desktop app there’s a home tab and a Cowork tab, and the difference between them is the whole ballgame.

Chat is a back-and-forth conversation, the same shape as ChatGPT. You ask, it answers, maybe it does some research for you. Useful, but you’re still the one doing every step of the actual job.

Cowork is where you train an AI employee to do the task on your behalf. If you want it to read through your email, put things on your calendar, and tell you what needs a response, you need to be in Cowork. That’s the framing I keep coming back to: this is hiring an AI employee. You train them once on what to do, and they repeat that work for you going forward.

My own Cowork setup goes further than email. I’ve got mine connected to my CRM so it manages Follow Up Boss for me, and connected to the MLS so it can pull comps and run a CMA. But none of that is where you should start. Start with one assistant, two connectors, and one job.

Step 1: Download the Desktop App and Get on the Pro Plan

Go to claude.ai and download the desktop app. There’s a Mac version and a Windows version. You can absolutely use Claude in a browser instead, and it works, but I use the desktop app because it makes everything easier and it’s what I’m demoing here.

Then register for an account. There are three tiers: free, Pro, and Max. I’m on Max because I’m a power user and I’ve been on this platform for a year, but that’s not where a beginner starts.

Get on Pro. It’s $17 a month if you pay annually, or $20 a month if you pay month to month. The reason Pro is the answer and free isn’t: Pro gives you access to Claude Code, Claude Cowork, Claude Design, and Claude Science. Cowork is the one this entire guide is about, and it’s the one that turns Claude from a chatbot into an employee.

Step 2: Build Your First Project (I Named Mine Christine)

Once you’re in the desktop app, you create a project. A Claude project is a repeatable workflow that follows your files, your rules, and has its own memory.

That sounds abstract, so here’s what it looks like in my actual account. I have an executive assistant project, whose whole job is my inbox and my calendar. I have a seller assistant project that takes a seller lead, pulls comps out of the MLS, sends the seller follow-up email, and walks through the seller process I designed. I have a buyer assistant project that runs my buyer consultation, sends the buyer rep agreement, sends listings, and sets up the search alert. I also have a YouTube content assistant, an Instagram assistant, and a Meta ads assistant.

Every one of those is a repeatable workflow that already existed in my business before AI touched it. So the question isn’t what can AI do: it’s what part of my business do I do the same way every time? A probate-focused agent wants a probate assistant that runs a client from first call to close. If you work move-up buyers or downsizers, you already have a process, and a process is exactly what a project is for.

For this build I created a project called Demo Executive Assistant and named her Christine. That’s my third: Ashley and Betty are already running elsewhere in my business. The human name isn’t a gimmick. It changes how you talk to her, and you write better instructions when it feels like you’re briefing a person.

The Four Parts of a Claude Cowork Project: Instructions, Rules, Memory, and Context

This is the part most people skip, and it’s the part that decides whether your assistant is useful or useless.

Instructions are the brain. Think of the instruction file as the soul of the project. Every single time you start a conversation inside this project, Claude reads the instructions before it answers you. Mine opens plainly: she is my AI executive assistant, she helps me stay organized, manage my inbox, review my calendar, prep for appointments, draft replies, and identify follow-up tasks. So no matter what I ask her, she already knows what she’s for.

Rules are what she’s not allowed to do. Mine are short and specific. Never send anything without my approval: I don’t want AI sending client-facing communication on its own, so she drafts and I review. Don’t invent transaction facts: AI is confident even when it’s wrong, and in this business a made-up number is a real problem. Keep it simple and communicate straightforward, because I want the most information in the fewest words. And “nothing here” is a valid answer for any section, which matters more than it sounds like: without that line, an assistant will manufacture something to fill an empty slot.

Memory is per project. This is the actual reason I keep the executive assistant, the seller assistant, and the buyer assistant separate instead of running one giant assistant. Each project remembers its own conversations and carries its own context, so Christine isn’t confused by what my seller workflow was doing last Tuesday.

Context files are how it sounds like you. I feed the project a set of my previous emails so it understands my tone of voice: how I write to clients, how I write to vendors, how I write to a new lead. That’s what lets a draft come back in my voice instead of in generic AI voice. Context is also where property specifics live. My seller assistant has a context file on the listing itself, so when I say the address, it already knows the bed count, the bath count, and the list price instead of guessing.

If you want my actual executive assistant instruction file rather than writing one from a blank page, it’s in the free Claude Starter Guide you can grab through the newsletter.

Step 3: Connect Gmail and Google Calendar

Christine can’t do anything until she has access to something. In the desktop app, click Customize, then Connectors.

The way I explain connectors to agents: they’re apps on your iPhone. Each one gives Claude the ability to operate a piece of software you already use. Search for Gmail, click it, click connect, log in, and grant access. Do the same for Google Calendar. Google Drive is right there too, along with a growing list of others worth browsing on my tools page.

For this build that’s all you need: email in, calendar out. Two connectors, one assistant.

Permissions: Read-Only Can Be Always Allow, Write and Delete Need Approval

As soon as you connect an app, Claude shows you a permissions screen, and this is the single most important screen in the entire setup. It’s worth understanding what you’re letting your AI assistant do and what you’re not, because that’s how you stay protected.

Split it in two.

Read-only tools can be set to always allow. Reading an email doesn’t change anything in the world. Making Claude ask before every single message it opens defeats the purpose of automating anything.

Write and delete tools should be set to needs approval. Don’t set them to blocked. Blocked would stop her from drafting an email at all, which is most of the value. Needs approval means she can compose it, but she has to come ask you before anything gets written, sent, or deleted. You can go tool by tool and set always allow, needs approval, or blocked individually, but read-only always, write and delete on approval is the setting I’d start every agent on.

Picking a Model: Haiku, Sonnet, Opus, and Fable 5

Before you run anything, you pick a model. Here’s how I explain the lineup to my bootcamp students, because staff analogies land better than benchmarks.

Haiku is the cheapest, and it’s your college graduate: genuinely good at simple, well-defined tasks. Sonnet is the master’s student: more capable, a little pricier, and it handles most real work efficiently. Opus is the PhD: incredibly intelligent, the one you bring in for complex problems, and one of the more expensive options. Fable 5 is the newest one, and I haven’t landed on a perfect analogy yet. Postdoc is close, but I think it’s really more like a VP or director-level employee who’s excellent at a lot of things at once. It’s also the most expensive model available right now.

Next to the model there’s an effort setting, which is how much time and computing resource the model spends on your task. Max effort gives you the best result, but it eats more of your usage limit and takes longer to come back. Most of the time I run high or medium and get what I need.

Don’t default to the most expensive option. Reading your email doesn’t require a PhD.

Run “Catch Me Up” First

The first thing I do on any new assistant is a warm-up: I select Haiku, because this is easy work, and I type “catch me up.”

A progress bar runs while she moves through the connectors, and then she came back with what was actually sitting in my business that morning: showing feedback owed on a home we toured that day, a buyer agreement I’d sent out that was still waiting on the buyer’s signature, and a pile of marketing email from builders and communities that didn’t need me at all.

That’s the test. If “catch me up” comes back with your real week, your instructions and connectors are wired correctly. If it comes back generic, go fix the instruction file before you try anything bigger.

The Real Demo: Christine Ran Transaction Coordination on 1117 Terrace View

Now the task I actually wanted to show, because catching up on email is table stakes.

We all lose track of deadlines. Contracts, receipt of option money, receipt of earnest money, and every important date buried across a dozen emails. So I opened the executive assistant project, picked a model, hit the microphone button because I was tired of typing, and said: understand the deal at 1117 Terrace View Drive in Georgetown, Texas, gather all of the contracts, read through them, and put all the important dates like the end of the option period and the closing date on my calendar.

Before it starts, Claude asks how you want to interact with it while it works: manually approve, automatically approve, or skip all approvals. Manually approve means you sit there clicking yes to every single action, which defeats the point of automating anything. Skip all approvals is only for a workflow you’ve run many times and know cold. Automatically approve is the balanced default, and it’s what I recommend to start. I picked that, chose Opus since this involved reading legal documents, started it, and walked away from my desk.

Here’s what she’d done when I came back.

She went into my inbox, found the contracts, and read the PDFs. Then she did something I never asked for and never programmed: she researched TREC, the Texas Real Estate Commission, on her own so she’d understand how these contracts actually function. How the option period works. How the end date is calculated. How the third-party financing addendum works. Nobody taught her Texas contract mechanics. She went and learned them because she needed them to finish the job.

Then she derived the deadlines out of the documents and wrote them to my calendar. The 3rd, flagged as the lender approval deadline and last day to terminate under third-party financing. Closing on the 6th at 10 a.m. I opened Google Calendar and they were sitting there.

That’s a transaction coordinator. I just taught my executive assistant how to do it instead, and now it runs on every deal going forward. This is the same idea I built my AI executive assistant workflow around, except Cowork makes it something you can set up in an afternoon without touching a terminal.

The Honest Part: Claude Can Be Confidently Incorrect

Check the work. That’s the whole caveat and I’m not going to soften it.

Claude will occasionally hand you a beautiful, well-formatted, completely wrong output. It’s confidently incorrect, and a well-formatted wrong closing date is worse than no closing date because you’ll believe it. We are not turning an AI employee loose to send client communication unsupervised, which is exactly why the approval rule sits in the instruction file and why write and delete stay on needs approval.

Mine does very well on dates and deadlines, comps, and net sheets, but that’s because I’ve been training this assistant a long time. Yours gets better the same way: you use it, you catch things, you tighten the instructions.

What Comes After Cowork: Scheduled Tasks, Skills, Plugins, and Claude Code

Everything above is the beginner layer. Here’s what you’re growing into.

Scheduled tasks run a job at a set time without you asking. I run one at 8:00 a.m. that goes through my email and drafts all the replies, and another at 8:00 p.m. that wraps up the day and sets up tomorrow. I wake up and the work is already done.

Skills are how you build a repeatable workflow that produces a consistent output every time. This is the concept that matters most after projects, and it’s most of what we build together in the bootcamp. I broke down the ones I actually run in the 5 best Claude skills for realtors.

Plugins package skills up so you can hand them to someone else on your team.

Claude Code is the more powerful version of what you’ve been using, and despite the name, I don’t use it because I’m writing software. I use it because it works without the restrictions Cowork has, and it’s where most of my own work happens now. Claude in Chrome extends the same thing into your browser.

Start Here

Download the desktop app, get on Pro, create one project, write a real instruction file with real rules, connect Gmail and Google Calendar, set read-only to always allow and write and delete to needs approval, and run “catch me up” every morning for a week. That one loop will teach you more about what an AI employee feels like than any other tutorial you can watch, including mine.

When you’re ready to build the rest of it, I run a live cohort where we set up the executive assistant, the content assistant, and the CRM assistant together, wired into your actual CRM and MLS. You can grab a seat or get on the waitlist at /cohort/. And the free Claude Starter Guide, including the executive assistant instruction file I use for Christine, is in the newsletter. Before any of that, the Claude for real estate agents starter guide covers the seven moves worth making in your first week.