Why "write me a caption" gets you a brochure
Large language models are prediction engines. Given a vague request, they predict the most average answer to that request, and the most average "new listing caption" on the internet is a cliché-stuffed brochure paragraph. You did not get a bad answer. You got the statistical middle of every bad listing post ever written.
The way out is not a secret trick. It is giving the model the same thing you would give a new assistant on their first day: who they are working as, what you need, the facts, the shape you want it in, and how it should sound.
A prompt is a brief, not a wish. Treat it like one and the output changes immediately.
The five parts of a prompt that works
Here is the formula. Each part is one sentence you can define in your head, with an example you can adapt tonight.
1. Role
The role puts the model in the right head before it writes a word. It tells the AI which version of itself to be, so it pulls from the right vocabulary and priorities.
Example: "You are a top-producing listing agent and copywriter in Austin."
Without a role, you get a generalist. With one, you get someone who already knows what an MLS description is for.
2. Task
The task is one clear verb and one clear deliverable. Write, summarize, rank, rewrite, draft. One job per prompt gets you a cleaner answer than three jobs mashed together.
Example: "Write a 120-word MLS description."
Notice the number. "A description" is vague. "120 words" is a task the model can actually complete.
3. Context
Context is the facts only you have, and it is the part the model cannot make up without lying. Bedrooms, baths, price, the renovation, the school, who the buyer is. Every fact you leave out is a fact the AI will invent.
Example: "3 bed, 2 bath, 1950s brick ranch, renovated kitchen, half-acre yard, $612K, near Oak Hill Elementary. Buyers: young families relocating."
This is where most agent prompts fall apart. It takes 20 seconds to type and it is the difference between copy about your house and copy about a house.
4. Format
Format is the shape of the answer, described so precisely you can paste it without editing. Word count, paragraphs or bullets, what to include, what to leave out.
Example: "Two paragraphs, no bullet points, no exclamation marks."
If you do not say the shape, you get headings, emoji and a bulleted feature list you now have to strip out by hand.
5. Tone
Tone is how the writing should sound, named in plain words plus a short list of banned words. Models love adjectives. Banning the worst ones does more than any amount of "make it sound natural."
Example: "Warm, specific, plain English. No 'nestled,' 'boasts' or 'oasis.'"
Weak prompt versus strong prompt
Put the five parts together and the difference is obvious.
Weak: "Write a caption for my new listing."
Strong:
You're my social media manager. Write an Instagram caption for a new listing: {{2-bed condo, Riverside, low HOA, 10-minute walk to the train, $389K}}. Audience: {{first-time buyers}}. 60 words, warm, no exclamation marks, end with "DM {{RIVER}} for the tour."
Role, task, context, format, tone. It is one sentence longer than the weak version and it produces something you can post from your phone 30 seconds later.
Eight prompt patterns worth memorizing
The five parts are the foundation. These eight patterns are the moves you layer on top once you have the basics down.
- Give it examples. Paste two of your best captions and say "match this style." Two real examples beat any adjective you could use to describe your voice.
- Ask for options. "Give me three versions: safe, bold, playful." You pick instead of edit, and picking is faster.
- Ask it to ask. "Before writing, ask me any questions you need." Fewer invented facts, because the model asks for the square footage instead of guessing it.
- Constrain hard. Word counts, banned words, required phrases. Constraints are not limits on creativity; they are what keep the output usable.
- Go step by step. "First list the features, then rank them by buyer appeal, then write the description." The ranking step is where the good copy comes from.
- Critique, then rewrite. "List five weaknesses in this draft, then fix them." The model is a surprisingly honest editor of its own work.
- Swap the role. "Now read it as a skeptical buyer. What would make you scroll past?" This catches the salesy lines you have stopped noticing.
- Demand paste-ready format. "Return only the text. No preamble, no headings, no 'Here's your caption!'" Saves a delete on every single output.
Here is a prompt that uses four of them at once:
You are a listing copywriter. Here are two captions I'm proud of: {{paste caption 1}} and {{paste caption 2}}. Match that style. Before writing, ask me any questions you need about the listing. Then give me three Instagram captions for {{listing facts}}: one safe, one bold, one playful. Each under 70 words. Return only the captions, no preamble.
Iterate like it is a conversation, because it is
The first answer is a draft. Nobody's first draft is the final draft, including the AI's, and the mistake most agents make is treating the first output as a verdict on whether "AI works."
Reply instead. "Shorter." "More specific about the yard." "Less salesy." "Now give me a LinkedIn version." Each reply builds on the last one because the whole chat is the context the model is reading from. It already knows the listing, the buyer and the tone. You are just steering.
Starting a new chat to make a small tweak throws all of that away. You will have to re-explain the listing from scratch and you will get a different draft, not a better one. Stay in the thread for tweaks. Start fresh only when the thread itself has gone wrong (more on that below).
A quick aside on what the model actually sees
Ask an AI model how many r's are in the word "strawberry" and it may confidently tell you two. It is not being lazy. It sees words as chunks called tokens, not as letters, so counting characters is genuinely hard for it. It can write you a sonnet about a split-level in Plano and still miscount to three.
The lesson for agents: AI is great at language and shaky at precision. Prices, square footage, HOA fees and dates need to come from you, and you should check them in the output every time.
The seven-step fix when the output still misses
Even with a five-part prompt, sometimes the answer is off. Diagnose it instead of giving up. Each problem has a specific fix.
- Too generic. Add context and paste an example. Generic output almost always means missing facts.
- Wrong tone. Name the tone you want and ban the specific words you hate. "No 'dream home,' no 'stunning,' no 'must see.'"
- Invented facts. Add this line: "Use only the facts I gave you. If something is missing, ask me instead of guessing."
- Too long. Give a word count, then add "cut anything that does not add new information."
- Wrong shape. Specify format: paragraphs, bullets, number of lines, what goes first.
- Missed the point. Ask it to restate the task in one sentence before it writes. If the restatement is wrong, you found the problem before wasting a draft.
- Still off after all that. Start over with a better prompt. Do not argue with a bad thread. A chat that went sideways early keeps dragging that mistake forward, and a clean restart with a sharper five-part prompt takes 30 seconds.
Keep this list somewhere you can see it. Most "AI doesn't work for real estate" complaints are one of these seven, and every one of them is a prompt fix, not an AI problem.
The no-prompt way
If writing a five-part brief every time sounds like one more thing on the list, there is a shortcut. Tools built specifically for agents bake the role, format and tone in so you only supply the context. In Agentree's AI Studio, for example, you pick the listing, the tool already knows it is writing an MLS description for a real estate agent in your voice, and you get the paste-ready result. No prompt box, no "nestled."
Either way, the formula is the same under the hood. Knowing it means you can get good copy from any AI, any time, and you will recognize a lazy prompt the moment you type one.