Prompt Engineering: How to Actually Get AI to Do What You Want

A real estate team we spoke with tried Claude for writing listing descriptions, got a bland, generic paragraph back, and decided "AI doesn't work for us." The tool wasn't the problem. The request was. This is the entire idea behind prompt engineering — and it's a skill, not a gift some people are born with.

The Anatomy of a Request That Actually Works

A vague request gets a vague answer. "Write a listing description" gives the AI almost nothing to work with, so it falls back on generic real estate language. A specific request gets a specific answer: the square footage, the three features that actually sell the house, the tone (warm, not salesy), and the length you need for the MLS field.

The pattern that works every time: say what you want, give the specific details that matter, and say how it should sound. Skip any one of those three, and you get a mediocre first draft.

Asking Cold vs. Showing It Examples First

There are two basic ways to ask for something. You can ask cold, with no examples — that works fine for simple, common tasks. Or you can show the AI two or three examples of what "good" looks like before asking for a new one — this works dramatically better for anything with a specific house style, like your listing voice, your firm's letter format, or your intake note format.

For anything your business does the same way every time, showing examples first is worth the extra minute. It's the difference between a new hire guessing at your style and a new hire who's seen five examples of exactly what you want.

Asking It to Show Its Work

For anything more complicated than a single paragraph — say, drafting a counteroffer summary that needs to walk through several numbers correctly — you get noticeably better results by asking the AI to work through it step by step before giving the final answer, instead of jumping straight to a conclusion. It's the same reason a person double-checks their math by writing it out instead of doing it in their head.

Instructions That Persist vs. the Question You're Asking Right Now

There's a difference between the standing instructions a workflow always follows — your firm's tone, your formatting rules, your do-not-say list — and the one-off question or task for today. Standing instructions get set once, when we build the workflow. The one-off request is whatever your team types in that day. Keeping those two separate is why a well-built workflow stays consistent even as ten different employees use it.

Getting the Output in the Format You Actually Need

If you need the output as a specific format — a table, a specific email structure, a set number of bullet points for an MLS listing — ask for that explicitly, and it'll hold to it consistently. This is what turns "AI wrote something" into "AI wrote something I can paste directly into our system without reformatting it."

Testing and Improving Requests Like You'd Test Anything Else

The first version of a prompt is rarely the best one. The teams that get the most value treat their requests the way they'd treat a new employee's first few tries — review the output, note what's off, adjust the instructions, and try again. After a handful of rounds, the request is dialed in and keeps producing solid results without more tweaking.

Why This Matters for Your Business

This is precisely the part of AI adoption most businesses get stuck on — not whether the technology works, but not knowing how to ask it well. We do that work for you. During a build session, we don't hand your team a chatbot and a "good luck" — we write, test, and refine the actual requests behind your workflow until the output is consistently something your team is happy to send, then hand over the finished thing.