How Do You Train AI to Sound Like You in Real Estate?

By Emily Terrell — Top Coach and Speaker at Tom Ferry International. 4 years coaching at Tom Ferry, 9 years prior as a client

To train AI to sound like you, build two layers: a one-page voice file with your rules, banned words, and compliance guardrails, and a corpus of 15 to 20 pieces you wrote plus raw transcripts of you speaking. Store both in a Claude Project or Custom GPT. This guide covers each layer, the blind test, and the maintenance schedule.

Key Takeaways

• You don’t train the model, you supply context, and that context has two layers: a one-page voice file of rules and a corpus of your real writing.
• Adjectives like “warm” and “direct” produce the average of every writer online, which is exactly the generic AI sound you’re trying to escape.
• Raw transcripts of you speaking are the strongest voice source you own, and almost no agent uses them.
• Fair Housing and TREC guardrails belong inside the voice file, so compliance happens on every draft instead of at proofread.
• If your output can’t pass a blind sort, you don’t have a voice system yet. You have a preference.

What does it mean to train AI to sound like you?

Training AI to sound like you means giving a model enough of your actual writing, your actual speech, and your actual rules that it reproduces your patterns instead of defaulting to generic professional English. You aren’t changing the model itself. You’re supplying context at the start of every session, and that context has to live somewhere the tool reads automatically. It’s an input problem, not a settings problem.

Why this matters for real estate agents

Adoption is finished. Results aren’t. According to NAR’s 2025 Technology Survey (September 2025), 46% of agents use AI-generated content, most often for listing descriptions. In the same survey, only 17% reported a significantly positive impact on their business, and 46% said AI made no noticeable difference at all.

Trust is the second gap. According to Realtors Property Resource’s February 2026 survey of 225 NAR members, 82% of agents currently use AI, but 63% named accuracy of outputs as their top concern and 49% named compliance or legal issues.

Read those numbers together. Nearly everyone is producing AI content, and almost nobody trusts it or gets anything from it. The model isn’t the problem. ChatGPT, used by 58% of AI-using agents according to the same NAR survey, writes perfectly well. The problem is that most agents publish content that sounds like every other agent’s content, and that content doesn’t build a brand, doesn’t get saved, and doesn’t get remembered.

There’s research on why this happens. A Cornell study presented at the ACM CHI conference in April 2025 had 118 participants write with and without an AI assistant, and found the AI’s suggestions pulled writers toward a homogenized style that flattened their distinctive patterns. The researchers studied it as a cultural effect, but the mechanism applies to any writer with a recognizable voice: the model pulls you toward its center, and its center is the average of everything it has read.

Here’s the thing nobody wants to tell you: your differentiation is the thing being averaged away. The fix isn’t a better prompt. It’s two documents you haven’t built yet.

“Most agents are trying to prompt their way out of a documentation problem. The model isn’t guessing at your voice because it’s weak. It’s guessing because you never wrote your voice down.”
— Emily Terrell, Tom Ferry Coach

The two-layer system for training AI to sound like you

Layer one is the voice file: a one-page document of rules. Layer two is the corpus: 15 to 20 pieces of your real writing plus transcripts of you speaking. The file tells the model what to do and what never to do. The corpus shows what you actually sound like.

Most agents build one layer and wonder why the output is still off. Rules without samples produce compliant, forgettable copy. Samples without rules produce your voice with your worst habits and none of your guardrails. You need both.

Why doesn’t describing your voice work?

Describing your voice gives the model a category, not a voice. “Write in a warm, direct, professional tone” fits thousands of writers, so the model produces the average of all of them. Every agent who typed those adjectives got the same paragraph.

Samples and prohibitions are the only inputs a model can actually match against. Treat every adjective in your current prompt as a placeholder for an example you haven’t pasted yet.

What goes in a one-page voice file?

A voice file is six blocks on one page, in plain text so it pastes cleanly into any tool. Aim for 400 to 600 words: specific enough to change the output, short enough that you’ll actually maintain it.

1. Three anchor samples. The three pieces that sound most like you: a listing description you were proud of, a client text that landed, a caption that got real replies. Paste them whole. These are the reference points the model checks everything else against.
2. Your banned list. Every word and construction you never want under your name. Start with “dream home,” “nestled in,” “rare find,” and “hidden gem,” then add your own tics. Include patterns, not just words: no exclamation points, no sentences over 25 words, no rhetorical questions in listing copy. Prohibitions are checkable. Adjectives aren’t.
3. Signature phrases. The five to ten lines you actually say, like how you open a market update or how you explain your process to a seller. These are your fingerprints.
4. Market vocabulary. Your submarkets, school districts, builders, and price bands. In San Antonio that means Stone Oak, Alamo Heights, 1604 versus 281, NEISD versus NISD. A model with no local vocabulary writes national copy, and national copy is what every other agent in your market is already publishing.
5. Structural rules. Caption length, description length, one call to action and where it goes, bullets or no bullets. Then name your structural signature: leading with the number, opening with the objection, breaking a false binary. Structure is more identifiably you than vocabulary, and a model will never infer it on its own.
6. Compliance guardrails. The block almost nobody includes, and the one that protects your license. It gets its own section below.

What compliance rules belong in the voice file?

Fair Housing applies to every agent. Federal regulation makes it unlawful to publish any advertisement for the sale or rental of a dwelling that indicates a preference, limitation, or discrimination based on race, color, religion, sex, handicap, familial status, or national origin (24 CFR § 100.75). Write the instruction explicitly: never describe the people who would live in the home, never use phrases like “perfect for families,” “safe neighborhood,” or “walking distance to church,” and never characterize a school or area as suited to a type of buyer.

Texas agents add a second layer. TREC Rule 535.155 requires every advertisement to include the license holder’s or team’s name and the broker’s name at no less than half the size of the largest contact information in the ad, and TREC’s definition of advertising covers email, text messages, and social media (TREC). If AI drafts your captions and emails, broker identification belongs in the file so it never gets dropped. Agents licensed outside Texas need their own commission’s equivalent rule.

This isn’t a fringe worry. In RPR’s February 2026 survey, 28% of agents named Fair Housing a top AI concern. A guardrail in the file resolves that concern before the draft exists.

This is general information, not legal advice. Run your guardrail language past your broker or an attorney before you standardize it across a team.

How many writing samples does AI need to learn your voice?

Fifteen to twenty pieces in the corpus, on top of the three anchor samples in your file. That’s the threshold I use with every coaching client, and it’s where the two numbers agents hear, “three samples” and “twenty samples,” stop contradicting each other. The three anchors are your standard. The twenty are your range: the variety of length, channel, and mood a model needs before it can reproduce you across a caption, an email, and a listing description.

Pull them from where your voice is least filtered: sent emails, Instagram captions, client DMs, texts you’d be comfortable sharing. Skip anything a marketing company wrote, including your website copy. You want your most typical writing, not your most polished, because typical is what readers recognize.

One rule matters more than the rest: never include AI drafts you edited. Feed a model its own output and it learns to imitate itself, drifting further from you every cycle. It’s the most common way agents end up with a voice profile that sounds like nobody.

Why are transcripts your strongest voice source?

Spoken you is closer to real you than written you, because writing triggers self-editing and speech doesn’t. When you write, you correct toward what you think professional writing should sound like. When you talk, you use your actual vocabulary, your actual rhythm, and your actual way of building an argument.

If you teach, run calls, post video, or speak on panels, you’re sitting on the best voice asset you own. Pull transcripts from Zoom recordings, YouTube captions, Reels, or podcast appearances and upload them raw. Don’t clean them up. The false starts, the mid-sentence pivots, the phrases you repeat without noticing: that’s the signal. Polishing a transcript turns it back into written English, which is the thing you were trying to escape.

Where should the voice file and corpus live?

In a workspace that loads them automatically: a Claude Project or a Custom GPT. Pasting your file at the top of every chat works on any free tier, and you’ll do it inconsistently within two weeks. A stored workspace reads your instructions and knowledge files before every conversation, so the context is already in the room before you type.

Put the voice file in the instructions field and the corpus in the knowledge files. Then build one workspace per recurring job, such as listing copy, sphere follow-up, and social, rather than one workspace for everything, because the rules for MLS remarks and the rules for a casual client text fight each other. The setup steps are in What Is a Claude Project, and How Do Agents Use One? If you’ve already built an assistant and the output still disappoints, Custom AI Assistant for Real Estate: Why Yours Fails walks through the files most agents skip.

Once the file exists, drafting becomes delegable. The quality control lives in the document instead of in your head, which is the same principle behind The Social Handoff, my training on handing social media to an assistant without your followers noticing the change.

How can you tell if AI content actually sounds like you?

Run a blind sort. Mix three AI paragraphs with three you wrote, strip the labels, and hand all six to someone who knows your writing: your VA, your assistant, your spouse. If they sort them correctly, ask which word gave it away. That word becomes your next banned-list entry. Then run it again.

This is the step that separates a voice system from a voice preference. “It sounds better to me” isn’t a measurement, and you’re the worst possible judge of your own voice. Most agents I coach reach a clean sort in three or four rounds.

What are the tells that give AI writing away?

Four tells show up in almost every draft that fails the sort.

• Uniform sentence rhythm. AI drafts tend to run sentence after sentence at similar length. Human writing spikes: four words, then forty. Read the draft aloud and listen for the metronome.
• Compulsive triads. Models list three items whether the topic has three or not. If a draft is wall-to-wall triplets, some of them are padding.
• Balanced clauses. “It’s not about working harder, it’s about working smarter.” That symmetry is a model habit. Break it.
• Concept nouns where a name belongs. AI writes “leveraging technology solutions.” You’d write “I put it in Follow Up Boss.” Every abstraction you swap for a real tool, place, or number makes the paragraph more yours and more credible at the same time.

How often should you rebuild the voice file?

Every six months, and immediately after any major change: a brokerage move, a niche shift, a new team name. Rebuild rather than patch. Pull fresh samples from the last 90 days instead of adjusting old ones, because your voice drifts and the newest samples are the most accurate. A file written eighteen months ago trains your AI to sound like a version of you that doesn’t exist anymore.

How I use this in my own business

I close 70+ transactions a year in roughly five hours a week of active management, and none of that math works if I’m rewriting AI output from scratch.

My listing copy runs out of a Claude Project. The instructions hold my voice file: banned list, signature phrases, local market vocabulary, the Fair Housing guardrail, and broker identification. The knowledge files hold about twenty of my own Instagram captions and client emails, plus raw transcripts from two stage sessions and a handful of coaching calls. When I take a new listing in Stone Oak, I drop the property specs in, and the description comes back already sounding like me, already clear of Fair Housing language, and already carrying my broker identification. I edit for factual accuracy, which takes about thirty seconds.

The transcripts did more than everything else combined. My written samples had been sanded down by years of trying to sound professional. My stage audio hadn’t. Once I added transcripts, my VA stopped being able to pick my drafts out of a lineup, and that was the point where I started publishing AI-assisted content without reading it four times first.

The final pass is still mine. A model can’t produce the car line, the conversation in the driveway, or the listing that fell apart on Thursday. Those details are what make a paragraph unmistakably mine, and they’re why I rewrite the first and last sentence of every draft by hand. Those two sentences are where AI is most generic and where voice is most visible.

If you’re rolling this out to a team rather than building it for yourself, the session structure I use is in AI Training for Real Estate Agents: What It Must Cover.

Common mistakes

• Describing your voice instead of showing it. Three tone words and no examples is the whole problem for most agents. Replace every adjective with a sample.
• Building only one layer. A voice file with no corpus produces correct, forgettable copy. A corpus with no voice file produces your voice without your guardrails.
• Feeding the model its own edited output. Edited AI drafts in your sample set teach the model to imitate itself. The drift is gradual, which is why most agents never notice it.
• Cleaning up transcripts before uploading. The disfluencies are the signal. Upload them raw.
• Treating compliance as the last step. If Fair Housing screening happens at proofread, it happens inconsistently, and eventually it doesn’t happen. Build it into the file so it runs on every draft.
• Skipping the blind sort. Without a test you’re guessing, and you’ll keep guessing for months while publishing content that quietly sounds like everyone else’s.
• Never rebuilding the file. Your voice, your market, and your team change. A stale file trains the AI on a version of you that’s gone.

Frequently Asked Questions

How do I train AI to sound like me?

Build two layers. First, a one-page voice file with three anchor samples, your banned words, signature phrases, market vocabulary, structural rules, and compliance guardrails. Second, a corpus of 15 to 20 pieces you wrote yourself plus raw transcripts of you speaking. Load both into a Claude Project or Custom GPT, then test the output with a blind sort until someone who knows your writing can’t tell the difference.

How many writing samples does AI need to learn my voice?

Fifteen to twenty pieces in the corpus, plus three anchor samples inside your voice file. The anchors set the standard, and the larger set shows the model your range across captions, emails, and listing copy. Every sample has to be genuinely yours. A single AI-edited draft in the set pulls the whole profile toward generic, so leave edited AI output out entirely.

What should a brand voice file include?

Six blocks on one page: three anchor samples of your own writing, a banned list of words and constructions, your five to ten signature phrases, local market vocabulary, structural rules for length and format, and compliance guardrails for Fair Housing and your state’s advertising rule. Keep it to roughly 400 to 600 words in plain text so it pastes cleanly into any AI tool.

Can I use transcripts of myself speaking to train AI on my voice?

Yes, and they’re usually your strongest source. Spoken language skips the self-editing that makes written work sound stiffer than you actually are. Pull transcripts from Zoom recordings, YouTube captions, Reels, or podcast appearances and upload them raw. Don’t clean up the false starts or repeated phrases, because those patterns are exactly what makes the output recognizably yours.

How do I know if AI content sounds like me?

Run a blind sort. Mix three AI-drafted paragraphs with three you wrote, remove the labels, and give all six to your VA or someone who knows your writing well. If they pick out the AI paragraphs, ask which word gave it away and add it to your banned list. Repeat the test until they can’t sort them correctly.

Why does my AI writing still sound generic?

Because the model has nothing of yours to match, so it defaults to the average of everything it has read. Tone adjectives like “professional” and “friendly” land on that default. A Cornell study presented at CHI 2025 found AI suggestions pulled writers toward a homogenized style. Replace adjectives with real samples and a banned list, and store them where the tool reads them automatically.

Does training AI on my voice create compliance risk?

Voice work doesn’t create the risk, but it can hide it. Copy that sounds exactly like you can still violate Fair Housing rules or your state’s advertising requirements. Put the guardrails inside the voice file so every draft starts compliant, then keep a human review before anything publishes. This is general information, not legal advice. Consult your broker or an attorney.

How often should I update my brand voice file?

Every six months, and right after any major change such as a brokerage move, a niche shift, or a new team name. Rebuild instead of patching. Pull fresh writing samples from the last 90 days rather than adjusting old ones, because your voice drifts over time and your newest writing is the most accurate reference for the model.

Bring this to your team or event

Emily Terrell speaks at brokerage events, real estate conferences, and team trainings on AI, systems, and social media — the exact playbook in this post, delivered live to your audience. As a Top Coach and Speaker at Tom Ferry International and an active agent closing 70+ transactions a year, Emily speaks from the stage about what’s working right now, not theory. Recent stages include NAHREP and eXp Con.

Book Emily to speak at your next event:
Email: eterrell@yourcoach.com
Phone: (210) 400-9191
Web: coachemilyterrell.com

For real estate agents who want to implement this: Get the weekly real estate prompt library at weeklyrealestateprompts.com or follow @coachemilyterrell on Instagram for daily systems and AI breakdowns.