Stop Starting From Scratch Every Time You Open ChatGPT
Every new chat starts with an assistant that knows nothing about you. A Project fixes that, and you only set it up once.
Every new chat starts blank
Open ChatGPT or Claude and you're talking to an assistant that has never heard of you. It doesn't know your market, how you write, or that you never say "stunning." So you type the background again, or you skip it and get something generic.
Even a long chat that went well has the same problem. Somewhere past the fiftieth message, the things you said at the start get fuzzy, and the answers drift back toward the average agent.
What a Project is
Both apps have a feature called Projects. A Project is a folder with two things in it: instructions the AI reads before every chat, and files it can look at. Every new chat you start inside it begins already knowing who you are.
You set it up once. After that, "write a reply to this buyer" works on the first try, because the instructions already say how you write and what you never promise.
One Project per job
Don't build one giant Project for your whole business. Instructions that cover everything end up covering nothing. Split them by the kind of work:
Client messages, with your voice file and the rules for what you never promise. Listings, with your Fair Housing rules and the words you ban. Seller conversations, with how you handle price and what your market looks like right now.
Each one stays short. If a Project's instructions run past a page, it's trying to do two jobs.
What goes in the instructions
Who you are, who you work with, and the rules. The most useful rules are the ones that stop mistakes: use only facts I give you, never state a market number I didn't provide, flag anything I should double-check. The free AI guide has a full set you can copy, along with the voice file that goes with it.
Rules about what not to do are often more useful than rules about what to do. "Never promise a sale price or a timeline" and "if you don't have the fact, ask me instead of guessing" stop the mistakes that actually cost you. The AI will otherwise try to be helpful in exactly the ways you don't want. The words and habits that make it sound like every other agent belong on that list too, and this post covers the six worth banning.
Fix it once, not every time
The part most agents skip is the update. The first draft from a new Project will be good, not perfect. When it gets something wrong, don't just fix that one draft. Fix the instructions, so the mistake doesn't come back next week.
After a draft that was off
That draft was wrong in these ways: [e.g. "too formal," "promised a timeline I never gave," "opened with a greeting I never use"] Write one or two short lines I can add to my Project instructions so you don't make this mistake again. Make them specific and give me only the lines, nothing else.
Paste what it gives you into the instructions. Do this a few times over the first month and the Project starts to sound like a person who's worked with you for a year.
What it won't do
A Project remembers your rules. It doesn't remember to write to the client you closed with three years ago, because it only works when you open it. Your past clients are most of next year's business, and they only come back if someone keeps in touch. That part is what I build.