Every morning you re-explain your own job to a chatbot. Unit mix, ILS contracts, which property is in lease-up, the fact that "The Maddox" is the one with the rooftop and "Maddox Row" is the townhomes eight miles away. Then you ask for a renewal email. Then you close the tab and the whole thing evaporates.
That's the tax. Not the writing, the briefing.
Projects fix it. Every major assistant now has some version: ChatGPT has Projects, Claude has Projects, Gemini has Gems. Same idea in each. You create a container, you load it once with the context and instructions that never change, and every conversation you start inside it begins already knowing your portfolio. You stop being a prompt writer and start being someone with a very fast, very literal junior marketer who read the whole binder.
What a project actually is
Three parts, and only one of them takes real work.
Custom instructions. A block of standing rules that gets prepended to every chat in that project. Tone, brand voice, banned words, formatting defaults, who you're writing for. Write it once.
Files. Uploaded documents the assistant can read on demand. Your brand guide, current concession sheet, floor plan list with square footages and rents, your ILS contract terms, last quarter's traffic report, the eight review responses you're actually proud of.
Chats. Individual conversations that live inside the container. Monday's renewal campaign, Wednesday's Google Ads copy, Friday's asset management update. They share the context but don't share each other's mess.
The mental shift: a chat is a conversation, a project is a job description.
Set up your first one in twenty minutes
Don't try to build one project for your whole portfolio. Build one per property, or one per property group if you're running forty and they share a brand. Then build a couple of function projects that cut across everything.
Here's the structure that holds up.
One project per lease-up or per repositioning asset. These need the most context and change the most. Instructions cover the property's positioning, the competitive set by name, current pricing, the concession you're running, the leasing team's names and how they sign emails.
One project per stabilized property group. Renewals, resident comms, review responses, social. Lower intensity, but the assistant still needs to know that Property A allows big dogs and Property B has a 40-pound limit before it writes a pet-friendly post.
One "Performance" project. No property-specific copy. This is where you paste CRM exports, ILS lead reports, and spend data, and ask questions. Load your source-tracking taxonomy so it stops inventing categories. Load last year's numbers so it can compare without you re-uploading.
One "Asset Management" project. Instructions: write for people who read numbers first, lead with occupancy and leased percentage, no adjectives, three bullets max per section. This one earns its keep the week before every ownership call.
What goes in the instructions
Most people write custom instructions like a horoscope. "Be professional and helpful." Useless. Instructions should be the stuff you'd otherwise correct on the second draft, written down once.
Actual examples worth stealing:
- "Never use the words luxury, oasis, or nestled. We are a value brand and residents can tell."
- "Our floor plans are named, not lettered. A1 is 'The Ridge.' Use the name."
- "Emails to current residents go out under Marisol's name, Community Manager. Emails to prospects go out under the leasing team generally, no individual name."
- "We do not say 'act fast' or 'limited time' unless there is an actual deadline I have given you in the chat."
- "Every piece of resident-facing copy ends with the property phone number and the office hours, which are in the uploaded fact sheet."
- "When I give you a lead number, do not congratulate me on it. Tell me what it means."
That last one saves more sanity than it should.
Fair housing lives in the instructions, and this is the part to get right
Ask a model for apartment copy and it will describe the resident instead of the unit. "Perfect for young professionals." "Ideal for a growing family." "A great fit for students." Every one of those describes a protected class or a proxy for one, and every one of those is a fair housing problem sitting in your ad account with your company's name on it.
The fix is a standing instruction, and it belongs in every project that produces public-facing copy: describe the unit, the community, and the amenities. Never describe who should live there. No references to family status, age, national origin, religion, disability, or any characteristic of a person. If the copy would still be accurate with any resident in the unit, it's fine. If it implies who belongs there, rewrite it.
Put that in the instructions and it applies to every chat you start. That's the real argument for projects over one-off prompts: your compliance guardrail runs automatically instead of depending on you remembering at 4:50 on a Friday.
Still read the output. The instruction reduces the failure rate a lot. It does not make the model incapable of the mistake, and you're the one whose name is on the ad.
The files that matter more than you'd think
Upload the boring documents. A one-page property fact sheet with unit mix, square footages, current rents, amenities, pet policy, parking, and utility structure will improve output more than any clever prompt you'll ever write. Add your concession sheet and update it when it changes. Add your five best-performing emails so it has a voice target instead of a vibe.
One warning: files go stale silently. The assistant will confidently quote a concession you stopped running in March, because you told it that was true and never told it otherwise. Date every file in the filename. "Fact Sheet - Maddox - Updated Feb 2025." Refresh monthly. Ten minutes, first Monday, done.
What you can do tomorrow morning
Open the Maddox project. Type: "Write three renewal offer emails for residents expiring in April, tiered by lease term." No context paragraph. No re-explaining the brand. It already knows the property, the voice, the signature, and that you hate the word nestled.
That's the whole difference. You went from thirty minutes of setup and a mediocre draft to twenty seconds and a good one, and the gap gets wider every week you keep the files current.
Where this breaks
Projects don't share context with each other, which is a feature until you paste something into the wrong one and get renewal copy for the wrong property. Check the project name before you type. Long-running chats inside a project can also drift off the instructions; if output starts feeling generic, kill the chat and start a fresh one in the same project rather than arguing with it.
And the assistant will never tell you a file is out of date. It will just be wrong, pleasantly, in your voice.