The first time I let an AI draft a listing description from rough notes, it gave the house a fireplace. Also a "sun-drenched breakfast nook." The house had neither. Nothing in my notes mentioned them — the model just decided a listing sounds better with them, and it was right about the sounding better part. That's exactly the problem.
Real-estate copy is a liability surface. An invented amenity isn't a style flaw; it's a misrepresentation a buyer can screenshot. So I rebuilt how I prompt for listing work, and the fix turned out to have nothing to do with clever tone instructions. It was three structural rules.
Rule 1: A fact boundary, stated like a contract
Every prompt opens by fencing the model in: write using only the facts I provide; do not invent amenities, schools, prices, or measurements. Then the facts go in as labeled fields — beds, baths, square footage, key features, outdoor features, updates — not as a paragraph the model can riff on. Labeled fields read as data. Paragraphs read as inspiration.
Rule 2: A defined output shape
"Write a nice description" gets you adjectives. "Write 2 short paragraphs, end with a neutral call to action to schedule a showing" gets you a draft you can actually paste. Same for social posts: three captions, three named tones (warm / short and punchy / question-based), and an explicit ban on misleading scarcity claims — because "won't last long!!" about a house that has sat for 60 days is the kind of line that ends careers.
Rule 3: Test on a fictional listing first
Before any prompt touches a real property, I run it on invented facts. The fictional run shows you the structure's failure modes with zero risk: Where does it pad? What does it do when a field is empty? (The correct behavior, which I now state explicitly for market content: write "not provided" — never fill a data gap with a plausible number. An AI-generated market statistic going out under my name is a hard no.)
Here's one of the two prompts I run most, exactly as it ships:
You are my real estate copy assistant. Write a polished listing description for this property using only the facts I provide. Do not invent amenities, schools, prices, or measurements.
Property facts:
- Address/area: [PROPERTY AREA]
- Property type: [HOME TYPE]
- Beds/baths: [BEDS] bed / [BATHS] bath
- Square footage: [SQFT]
- Key features: [FEATURE 1], [FEATURE 2], [FEATURE 3]
- Outdoor features: [OUTDOOR FEATURES]
- Updates: [RECENT UPDATES]
- Target length: [WORD COUNT] words
- Tone: [TONE: warm / modern / luxury / concise]
Write 2 short paragraphs. End with a neutral call to action to schedule a showing.
Nothing in it is magic. That's the point: the reliability comes from the fence, the shape, and the test run — not from finding the one perfect phrasing.
If you want the full set I use (50 prompts across listing copy, client email, market explainers, social posts, and objection handling), there's a free 5-prompt mini pack here: https://promptandskillco.gumroad.com/l/free-real-estate-prompts-mini — and the full pack is $29 one-time: https://promptandskillco.gumroad.com/l/evjxf
Disclosure: I sell the full pack. Every example in this post is fictional.
Question for agents who use AI on live listings: what's your verification step before a draft goes out — do you re-check every claim against the MLS sheet, or have you built the check into the prompt itself?
Top comments (0)