The prompt trap
Open the chat. Type a question. Get an okay answer. Close the tab. Tomorrow, type the same question slightly differently and get a different okay answer. Repeat forever. Never get better.
Free class · Content
How to make AI write real estate content in your voice.
Your AI captions don't sound generic because the AI is generic. They sound generic because it gets a question with no context. In this build-along you write four small files once, then stack them to produce any piece of content in your voice.
Ryan Wanner, founder of AI AccelerationInnovation Lab North, Compass Lakeside, HendersonvilleEvery class on this page is free to read and watch. No account needed.
What you'll learn.
- 01Why the same model gives one agent generic copy and another agent copy that sounds like them.
- 02How to write the plan yourself before you open AI, and why the observing is always yours.
- 03How to pull your real voice out of three things you've already written.
- 04The four plain-text files that make up the machine: self.md, persona.md, subject.md, and machine.md.
- 05How to stack the files for any output, from a listing description to a newsletter.
- 06Why every correction goes on the file, not the prompt, and the two safety checks before anything posts.
Part 1
Files beat prompts.
The file system
Build four files one time. Paste them as context every time you ask. Each correction goes on a file, so each output gets sharper. You stop starting over.
Write the plan before you touch AI.
Two minutes, pen on paper: what you're building, what done looks like, what you already know cold, and the smallest version that ships today. Then paste it into AI as a second pair of eyes. AI can help you make sense of things and draft the work. Observing your business is yours.
Own your context. Don't rent it.
If everything the AI knows about you lives in one app's memory, that app has you locked in. Files on your laptop go anywhere. When a better model comes out, you paste the same files and keep going.
Part 2
Build the four files.
One folder on your laptop. Plain-text markdown files, no code. Pick one kind of content you'd actually ship this week and build for that first. The same files work for the next one.
- 01
self.md
Who you are, how you sound, what you know cold, your signature phrases, and your do-not-say list. The file that makes every output sound like you.
- 02
persona.md
The audience you're writing for right now: who they are, what they want most, what they fear, the words they use, and where they are in the decision. One file per audience.
- 03
subject.md
The thing you're writing about: a listing, a market topic, a recipient, or a listing event. Only verified facts. Anything you're not sure of gets marked [verify].
- 04
machine.md
The runner. It tells the AI how to read the other three files, which output format to produce, and to check the do-not-say list before returning anything.
Start with self.md: pull your voice from what you've already written.
Don't describe your voice with adjectives. Show it. Paste this prompt with three pieces of your own writing. Pick your normal Tuesday-morning writing, not your best work. That's what your voice actually is.
Voice extraction
Analyze the writing style across these 3 pieces of writing. Tell me: 1. Typical sentence length and rhythm 2. Signature phrases or openers I actually use 3. Words I lean on. Words I avoid. 4. What makes this distinctly mine, including patterns I might not see myself Be specific. Cite examples from the text. I'll use this output to build a voice card. --- [Paste piece 1: a listing description you wrote] --- [Paste piece 2: an email to a past client] --- [Paste piece 3: a text to a client]
Use the analysis to fill in five sections:
- 01
Who you are
Your name, market, brokerage, and what you actually sell. Two sentences at most.
- 02
How you sound
Five tone bullets written as contrast pairs: warm but not cheesy, data first but not dry. Generic bullets produce generic output.
- 03
What you know cold
Three things AI can't look up: the school-zone lines in your area, how an HOA really works, the detail you explain every week.
- 04
Signature phrases
Three to five things you actually say. If you've never said it out loud, it doesn't go on the list.
- 05
Do not say
Three to ten phrases you never use: the clichés you hate, the Fair Housing landmines, anything that makes you cringe.
Then run the style-match test. Paste your finished self.md into a fresh chat and ask for a three-sentence caption about one of your listings. If it sounds like you, the file works. If it sounds robotic, a tone bullet is too generic. Sharpen it and run it again.
persona.md stops generic output.
Without an audience, AI writes to nobody. Give it who they are, what they want most in order, what they worry about, the words they actually use, and how close they are to deciding. Name the file after the person, like "Downsizer" or "First-time buyer newsletter subscriber," not a buyer category. One subject plus five personas is five different pieces.
subject.md holds only verified facts.
Four variants, same shape: a listing, a market topic, a recipient, or a listing event like a price change. Every fact comes from the MLS, the seller, public record, or what you saw yourself. If you don't know it, write [verify]. The AI sees the brackets and doesn't invent.
Part 3
Stack the files. Ship one piece.
machine.md is the runner. Copy this into a new file, keep the output format you built for, and save it next to the other three.
machine.md: the runner
# machine.md: Content Output Machine ## Role You are a content production assistant for a real estate operator. Always read the attached self.md, persona.md, and subject.md before writing anything. ## Voice Match self.md exactly. Use signature phrases naturally. Never use anything from the do-not-say list across any of the three cards. ## Audience Speak directly to the persona in persona.md. Use their vocabulary. Match their stage. ## Facts Only use facts present in subject.md. If a fact is marked [verify], write [VERIFY: <fact>] in the output and move on. Never invent. ## Output formats (keep the one you built for) - listing_mls: 180 words max, no exclamation - listing_caption_ig: 150 words, hook in line 1 - email_just_listed: 200 words, subject + body, 1 CTA - sms_sphere: 3 variants, under 160 chars, no links - blog_neighborhood: 800-1000 words, H2 sections - cold_email_3touch: 3 emails, under 125 words each - video_script_short: 30-45 sec, hook + body + CTA + thumbnail brief - newsletter_monthly: 3 sections, 600 words total, 1 CTA ## Compliance pass Before returning output, scan against every do-not-say in self.md and subject.md. Replace any flagged phrases. ## End every output with A [VERIFY] block listing every fact from subject.md so the operator can confirm before publishing.
machine.md plus self.md. Every stack starts here.
When you're writing about a specific listing, topic, recipient, or event.
When you're writing for a specific reader. Most content needs both.
Open a fresh chat. Paste machine.md, then self.md, then the others your job needs. Type one sentence: "Write [output format] for [audience or subject]." Read it against your do-not-say list, clear the [VERIFY] block, edit a line or two, and ship it.
When the machine gets it wrong, fix the file.
Robotic output means self.md needs a sharper tone bullet or another signature phrase. Wrong vocabulary means persona.md needs the words your audience really uses. An invented fact means subject.md needs the verified version. A banned phrase goes on the do-not-say list. The prompt is gone after this chat. The file stays in your folder. One more trick: once a long back-and-forth finally gets the output right, ask the AI what you'd need to give it to get there in one try, and put that in the file.
AI doesn't understand consequences. You do. The machine drafts. You're the publisher, and the publisher carries the responsibility.
Free with a student account
Get the workbook and prompts, free.
Create a free student portal account to get the workbook, prompts, templates, and downloads for every class.
Create my free account →Already a student? The same link signs you in.
- Realtor, Client, and Channels context card kits with AI builder prompts
- The full class slide deck
- The Content Creation Machine skill card for your AI
Questions agents ask.
How do I make AI write in my voice?
Show it, don't describe it. Paste three things you've actually written, like a listing description, a client email, and a text, and ask the model to analyze your sentence length, signature phrases, and word choices. Build a self.md file from that analysis, then test it: ask for a short caption and see if it sounds like you.
Why does my AI content still sound robotic when I write detailed prompts?
Because a detailed question with no context still gets a generic answer. The model doesn't know your voice, your audience, or your facts. Put those in reusable files and paste them every time. When output drifts, fix the file that caused it instead of rewriting the prompt.
What is a markdown file, and do I need to code?
A markdown file is a plain-text file that ends in .md. You write it in plain English with a few headings. No code. It's a clean format for AI to read, and because the files live on your laptop, you can take them to any model you want.
Which AI tool should I use to build this?
Any of them. Web apps like ChatGPT, Claude, Gemini, and Grok are the fastest start: you paste the files into a chat. Desktop apps can read the folder on your laptop and keep projects between sessions. Agentic coding tools have the highest ceiling and the steepest ramp. Pick the one where you can finish a build today.
Is AI-written listing copy safe for Fair Housing?
Only after you review it. You're responsible for everything you publish, not the AI. Put Fair Housing landmines on your do-not-say list so the machine avoids them, then read every output for protected-class language and verify every fact before it goes out.
Keep going.
Content
How to Fix Generic AI Real Estate Pages
Improve an AI-built page without starting over: a blind-spot pass, fixes folded back into your context, honest cuts, and one clear call to action.
Content
How to Stop AI Writing From Sounding Robotic
Build a Voice Kit from messages you actually sent, rewrite any draft in your voice, and run four checks before it goes out.
Get found
How to Get Recommended by AI for Real Estate
Check what answer engines say about you, verify every source they cite, and publish one sourced answer worth citing.
Library
Every free class
The whole library of recorded classes, by topic.
Research sources
Check the current rules, capabilities, and source basis.
- 01HUD: Fair Housing guidance for digital advertising ↗
Government guidance on Fair Housing risk in digital housing advertising and delivery.
- 02NAR: 2026 Code of Ethics ↗
Current professional standards requiring a true picture in advertising and prohibiting misleading images and manipulations.
- 03Google: Generative AI content guidance ↗
Google's guidance on accuracy, quality, relevance, metadata, disclosure context, and avoiding low-value scaled generation.
- 04FTC: Advertising and marketing guidance ↗
Government guidance on truthful, non-deceptive advertising and substantiation.
Product access, limits, local listing rules, platform policies, and laws can change. Re-check the linked authority before important or public work.
Keep building
Put AI to work in your real estate business.
AI Acceleration: Real Estate is a private, self-directed curriculum for practical AI work across content, communication, lead follow-up, CRM, client research, operations, and relationship management. Build it. Review it. Keep what works.
Explore the Skool community ↗Follow Ryan + AI Acceleration
No ranking, revenue, closing, or time-saving promise. The classes teach the work and the review systems around it.
Record: innovation-lab-public-classes. Approved record hash: add45e5ab0ee95b80ad70a4ecb50c8b8317c6997119c13fd8a727225815aff73. Claims: public-classes-free-to-watch, public-classes-portal-unlock, public-classes-taught-live.