Field notes / AI systems · Level 2 of 5

Building a Buy Box That the Model Actually Follows

Write your criteria once, put them where the assistant reads them every single time, and deal screening stops depending on which version of you showed up that morning. Then run the same trick backwards on your buyers.

Level 1 ends with you pasting the same paragraph about what you buy into every conversation. Level 2 is one move: that paragraph becomes a document, the document lives inside a project, and every deal you paste gets judged against it automatically. Same verdict on Tuesday as on Friday. Same verdict when your VA runs it as when you do.

That last part is the point. This is the first rung where AI works for your business instead of for you personally, because it is the first rung where the judgment is written down instead of living in your head.

This post has two builds. First, acquisitions: screening incoming deals against your buy box. Second, the one almost nobody does: dispo, matching a deal you control against your buyers' buy boxes, so you stop blasting 400 people with every contract and start sending five buyers something they actually want.

Step 1: Write the buy box document

Open a doc. Call it Buy Box v1, and put today's date in it. This document is about to become the single source of truth for what you buy, so treat it like one: when your criteria change, you change the document, not the prompt.

Here is the skeleton, filled in with example numbers. Replace them with yours.

BUY BOX v1 (updated 2026-08-17)

WHO I AM
Creative finance investor. I buy with seller finance, subject-to,
and hybrid structures. Exit strategies: coliving conversion,
midterm rental, or wrap.

MARKETS
Primary: San Antonio, Corpus Christi TX metro areas.
Secondary: Southeast US, case by case.
Hard no: rural (over 30 min from a metro), flood zone A/V.

PROPERTY
Single family residential, 3-6 bedrooms, 1,200+ sqft.
Built after 1960. No manufactured, no condos.
Value-add welcome: cosmetic to medium rehab. No foundation
or fire-damage projects.

NUMBERS
Purchase price: $200K-$600K.
Subject-to: existing rate under 5.5%, PITI leaves at least
$400/mo spread against market rent (or $700 against
coliving/MTR income).
Seller finance: 10% down max, rate 4% or lower, balloon
5+ years out.
Cash-to-seller at close: under $40K all-in including arrears.

DEALBREAKERS (automatic PASS)
- Litigation, open probate without authority, or unclear title
- Balloon under 3 years
- Negative spread at market rent with no coliving upside
- Seller wants full retail AND all cash

VERDICTS
PURSUE: fits the box, numbers work on stated terms. Move today.
DIG: could work if 1-2 specific unknowns break right. List them.
PASS: fails a dealbreaker or the numbers. Say which, in one line.

OUTPUT FORMAT
Verdict first. Then a 5-line summary: address, structure,
the key numbers, the one thing that makes or kills it,
next action. Never bury the verdict.

Two notes on why this works. The verdict labels give the model a decision to make instead of an essay to write, and forcing "verdict first" means your VA can process thirty of these without reading thirty essays. The dealbreakers section does more work than the wishlist; models are agreeable by nature, and an explicit "automatic PASS" list is what gives yours permission to say no.

Step 2: Load it into a project

1. In Claude, create a new Project. Name it "Acquisitions Screening."
2. Add your Buy Box doc to the project's knowledge. Every conversation inside the project now reads it automatically.
3. In the project instructions, paste: "You screen real estate deals against the Buy Box document. Always follow its VERDICTS and OUTPUT FORMAT sections exactly. When information is missing, list what's missing under DIG rather than assuming values."
4. Optional but powerful: add 2-3 past deals as files, each labeled with the verdict you actually gave and one line on why. Real examples teach the model your judgment faster than any instruction.

On ChatGPT the same build is a Project (or a Custom GPT) with the doc in its knowledge and the same instructions. Identical idea, different menu.

Step 3: Screen deals

Now the daily motion is one paste. Wholesaler blast, agent email, text from a bird dog, whatever:

Screen this deal.

[paste the entire email/listing/text thread, numbers and all]

That is the whole prompt. The buy box, the verdict system, and the output format are already in the room. This is what Level 2 buys you: the intelligence moved from the prompt into the infrastructure.

Step 4: Test it before you trust it

Do not skip this. Before the project touches live deal flow, feed it three deals where you already know your answer:

1. An obvious PURSUE you did or wish you did.
2. An obvious PASS that violates a dealbreaker.
3. A genuine edge case that made you think for a day.

If the model gets one wrong, the fix is almost never a better prompt. The fix is a clearer buy box. When it called my edge case a PURSUE and I would have said DIG, it was because my document said nothing about arrears counting toward cash-to-close. I fixed the document, bumped it to v1.1, and the verdict flipped. Your buy box doc will get better precisely because the model is bad at reading your mind. So is your VA. That is the same problem, and this fixes both.

Step 5: Hand it to your VA

Invite your VA to the project (both Claude and ChatGPT support shared projects on team plans; worst case, they log a shared account). Their instruction set is one line: paste every incoming deal, forward you everything marked PURSUE immediately, batch the DIGs for your Friday review, log PASSes to a sheet with the one-line reason.

Review the log weekly for ten minutes. You are looking for two failure types: PURSUEs you would have passed (your box is too loose somewhere, tighten the doc) and PASSes you would have chased (a dealbreaker is written too broadly). Every correction goes into the document with a version bump. After a month of this the doc is sharper than what was in your head, because it has been hit with live fire.

The dispo flip: buyer buy-box matching

Here is the follow-up move, and if you wholesale or JV at all, it might be worth more than the acquisitions side.

Every operator I know keeps their buyers list as names and phone numbers. Almost nobody keeps it as criteria. Which means every contract goes out as a blast to the whole list, response rates rot, and your serious buyers learn to ignore you.

The fix is the same trick pointed the other direction. One spreadsheet, one row per buyer:

Name | Markets | Property types | Beds | Price band |
Strategy (flip / rental / coliving / MTR / wrap) |
Financing (cash / hard money / DSCR / takes subto) |
Max rehab appetite | Proof of funds on file (Y/N) |
Last closed with you (date) | Notes

1. Create a second project: "Dispo Matching."
2. Add the buyers sheet (export it as CSV) to the project knowledge, plus a short instructions doc: "Match deals to buyers from the list. Rank matches. Never recommend blasting the full list."
3. When you have a deal to move, paste it in with the prompt below.
4. Keep the sheet honest: after every dispo, update who bought, who ghosted, who said "more like this." The list is a living document, same as the buy box.

I have this deal under contract. Match it against my buyers list.

Give me:
1. Top 5 buyers ranked, each with 2 lines: why they match,
   and the one thing in their criteria that might make them pass.
2. Buyers who look close but aren't. One line on why,
   so I don't waste their attention.
3. A 3-sentence pitch for the #1 buyer, written for a text
   message, using their strategy and numbers, not generic
   deal-speak.
4. If fewer than 3 buyers match, say plainly that this deal
   doesn't fit my list, and describe the buyer profile I'd
   need to find.

Deal:
[paste address, numbers, terms, photos summary, rehab estimate]

Item 4 is the honest-broker clause. Some deals do not fit your list, and knowing that before you blast is the difference between a dispo operation and a spam operation. Item 3 quietly solves the other half: the first message the buyer sees is written against their criteria, which is why they answer it.

Where Level 2 breaks

Everything above still runs on a human finger. Somebody pastes the deal in, somebody copies the verdict out, somebody updates the sheet. When deal flow is five a week, that is fine. When it is fifteen a day across three inboxes, the paste itself becomes the job, and things start slipping through on the days nobody pastes.

That is the Level 3 conversation: connecting the assistant to the inbox and the sheet directly, so the reading happens without you. Next post, I build my actual morning triage: every email across multiple Gmail accounts, aggregated, screened against this same buy box, and prioritized before I have had coffee. And I will be blunt about the new failure mode you inherit the moment nothing forces a human to look.

The series

1. 00The Five Levels of AI in a Real Estate BusinessThe map
2. 01Getting real value out of a chat windowPublished
3. 02Building a buy box that the model actually followsYou are here
4. 03Wiring an assistant into your inbox, sheets, and CRMPublished
5. 04Getting it off your laptop: hosting, logging, and alertsComing next
6. 05Evals and cost control, or how to know it is still rightComing soon