I Built an AI Seller Assistant for Real Estate Agents in Claude (Full Build)

I have a confession: for months, every time a seller lead came in, I opened a brand new conversation with Claude and started over. Explained my brokerage. Explained my process. Explained who I am, from zero, every single time. It worked well enough, but it wasn’t a system: it was just a very smart chat window I kept resetting.

So I built one instead: an AI seller assistant for real estate agents that lives inside a single Claude project, and I’m walking through exactly how I built it, what it actually does on a live listing, and what I’m building into it next.

I’m William Zhang, a real estate agent in Austin, Texas, and the founder of Real Estate AI Society. I’ve been building my practice on Claude for about two years, and this seller assistant is the version I actually run on my own listings, not a demo I put together for a video.

By the end of this post you’ll know how I set up the project instructions, what happens when I send this assistant into the MLS itself, how it builds a listing website and a full listing presentation while I step away, and what I mean when I say this is only version one. If you want the actual instruction file instead of rebuilding it from scratch, comment “assistant” on the video above, or grab it through the newsletter.

Setting Up an AI Seller Assistant for Real Estate Agents Inside Claude

Everything here starts in the same place: Claude desktop, then Projects, then create a new project. I named mine Real Estate Seller Assistant, and for the description I wrote one plain sentence: help me handle seller leads from beginning to end. That’s it. That’s the whole setup step.

Once the project exists, you click Instructions, then Edit Instructions, and paste in the file that actually teaches it your business. Mine covers a handful of things: it prices out my opinion of market value, it follows fair housing rules, and it pulls MLS comps because I connected the assistant directly to my MLS data. It knows how to build a listing website and a listing presentation. And it handles a use case most agents never automate: seller updates, where it goes back into the MLS to check showing feedback and activity, finds homes nearby that dropped their price, and turns that into a report I can send my seller every week.

Once the instructions are in, I don’t touch anything else. I just open the project and ask it directly: “What can you do for me as my assistant?” It answers back in plain language, and that’s my confirmation the setup actually worked before I put it in front of a real client.

The One Rule That Makes This Assistant Trustworthy

There’s one rule in the instruction file that matters more than everything else combined: never invent a fact. Prices, dates, square footage, comps, days on market, fees, addresses, names. If it isn’t in a file I handed it, it says “not found” or it asks me. It does not estimate.

That sounds obvious until you’ve watched an AI tool confidently make up a number that isn’t real. In this business, a wrong number isn’t a typo, it’s a pricing conversation with a seller that goes sideways, or a fair housing problem you didn’t see coming. So the assistant is built to go get the actual figure from the MLS or from me directly, never to guess and sound confident about it. Everything downstream, the pricing, the comps, the market updates, only works because that rule sits underneath all of it.

Sending Claude Into the MLS for Showing Instructions

The first live piece I ran was getting Claude into the MLS itself, not asking it to summarize data I’d already pulled. I made sure the Claude plugin on Chrome could see Unlock MLS, logged in when it hit the login screen, and switched the model to Opus, since this was the initial work and I wanted the most advanced model available, not the fastest one.

From there it worked on its own. It navigated into Unlock MLS, pulled the listing information for every property my buyer wanted to see, and came back with the showing instructions attached to each one. Then it went a step further: it could schedule the showing either through ShowingTime, if that’s how the listing agent wants it handled, or by texting the seller’s agent directly to set up the tour. That’s the part I want to be specific about, because it’s easy to hear “MLS” and assume this means a comps report. It doesn’t. What it actually does is get into the MLS and pull the data itself, the showing instructions, the scheduling method, the details that used to mean a phone call or a login of my own. I’ve got the rest of my MLS stack (comps, tour routing, print sheets) written up on /tools/ if you want to see what else runs alongside this.

The Listing Website This Assistant Builds From One Prompt

Next I asked it to build a single-page property website, and this is where a reference file matters. A reference file is just an existing site with the layout and style I want, plus the listing photos I already have. I attached both, and the whole thing ran for about 11 minutes while I worked on something else. That’s the actual experience of having an AI employee: I hand it the task and walk away, and it’s done by the time I check back.

What came back had my name, my brokerage, the schedule-a-tour form, and every listing photo, laid out cleanly on both mobile and desktop. But that front end is only half the job. Click “request a showing” on a brand-new site and the form isn’t connected to anything yet, so I wire the back end through Tally.so, a free form tool I’ve taught the assistant to connect to. Now a submitted form sends me an email and adds the lead straight to my CRM. Then I host the finished site on here.now, a free hosting platform, so it’s a real, shareable web address instead of a file sitting on my computer. I went a lot deeper on this exact build, reference file and all, in how I built a listing website with Claude.

Building the Listing Presentation While I Step Away

The last piece is the listing presentation, and this is where you’re making your actual case to the seller: how you plan to market their home, how you’ll help them stand out, and where you get to ask what they actually want from an agent.

I used a prompt that covers my name, my brokerage, my years in the business, the areas I serve, my contact information, and the property address. I ran it against one of my real listings, 1117 Terrace View, stepped away, and came back 5 to 10 minutes later to 15 to 17 finished slides. It opens with outcome, not my resume: the least friction, a clean and predictable close, full transparency the whole way through. It builds a slide around the five things beyond the final number: price, preparation, marketing, negotiation, and how I manage the process start to finish. It pulls in current market conditions, months of inventory, average days on market, and because the assistant already knows my YouTube channels and my media presence, it lists that out too, as a real differentiator against agents who don’t have an audience behind them. It closes with a day-by-day launch strategy, prep before launch, the Friday go-live, an open house weekend, a review at day seven, and it ends with the close: a listing agreement ready to sign. I broke down the full build, prompt included, in how Claude builds my listing presentation.

This Is Only Version One: What I’m Building Into the CRM Next

Everything above gets you the listing. It’s marketing, comps, showings, the pitch itself. But I want to be direct about where this stops: version one helps you win the listing. Version two is what goes and finds it.

I’ve already connected this assistant to Follow Up Boss, my CRM, and here’s what I’m building next. Drip campaigns tailored to the seller leads already sitting in my database: someone who requested a home valuation gets a personalized sequence of texts, emails, and videos instead of a generic “just checking in.” Funnels that bring brand-new homeowners into my CRM in the first place, so the pipeline of future sellers keeps filling on its own. And market updates and home reports that go out without me touching anything, pulled straight from the MLS: how many homes in a seller’s neighborhood came on the market that week, which ones dropped price, which ones went under contract. I already showed you how the assistant gets into that data directly. Now it’s about turning that into something that ships to every seller automatically. I’ve documented how the CRM side already runs in how I use Claude to manage my Follow Up Boss CRM, and version two of this seller assistant is next.

If you want to build this with me, I run a four-week live cohort where we set up your AI operating system together, seller assistant included. You can join the next cohort or get on the waitlist here. And if you just want the exact instruction file behind this build, it’s free in the newsletter.