Field notes / AI systems · Level 1 of 5

Getting Real Value Out of a Chat Window

Most people ask a four-word question, get a mediocre answer, and decide the technology is overhyped. The problem is not the model. It is the four words.

This is the second post in the Five Levels series, and it covers the bottom rung: you, a browser tab, and a model. No automation, no connectors, nothing to install. If you read the first post you know I called this the biggest single jump on the whole ladder. Twenty dollars a month, and for an operator who has never done it properly, it changes what a Tuesday looks like.

The catch is the phrase "done properly." Almost nobody does. I have watched smart investors, people who can walk a property and price the rehab in their head, type "is this a good deal" into ChatGPT with no address, no numbers, and no context, then show me the useless answer as proof that AI is not there yet.

The model is there. Here is how to meet it halfway.

The one rule: context is the product

A language model is not a search engine. A search engine wants three words. A model wants everything you have: the whole email thread, the whole contract, your situation, your constraints, what you plan to do with the answer. The quality of what comes back tracks the quality of what you put in, almost linearly.

So the one rule is this: paste more than feels reasonable. The whole seller thread, not your summary of it. The full inspection report, not the three items you remember. Models read 100 pages in seconds and do not get tired or skim. The thing that feels like oversharing to you is the thing that makes the answer specific instead of generic.

I use Claude for this (I build on it professionally, so I know its edges). ChatGPT works the same way for everything in this post. Pick one and stop tool-shopping.

The five habits

1. Say who you are and what you are doing. One sentence changes everything: "I buy 3-6 bed value-add houses in Texas using seller finance and subject-to." Now every answer is filtered through your actual business instead of a generic one.
2. Paste everything, then ask. Context first, question second. Full documents, full threads, photos of paper. Do not pre-summarize; your summary is where information dies.
3. Specify the output format. "Give me a table," "give me three bullets for my VA," "write it as a text message I can send." You will use the answer somewhere. Ask for it in the shape of that somewhere.
4. Push back on the first answer. The first response is a draft, not a verdict. "You missed the balloon payment." "Shorter." "What would a skeptical lender say about this?" The second and third turns are where the good stuff lives.
5. Verify anything with a number in it. Models occasionally misread or round numbers, and they do it confidently. For anything that feeds an offer, make the model quote its source: "for every number, quote the exact line from the document you got it from." Then spot-check.

Worked example: the rambling seller email

A distressed seller sends you 900 words at 11pm. Divorce, a roof they maybe fixed, a tenant who maybe pays, and somewhere in there, what they actually want. Paste the whole thing under this:

I'm a real estate investor. Below is an email from a seller.

Give me:
1. Their situation in two sentences.
2. Every number mentioned (asking price, debt, payment, arrears), quoted exactly as written.
3. What they seem to actually want, which may differ from what they're asking for.
4. Motivation signals: timeline pressure, emotional language, flexibility hints.
5. Three questions I should ask next, ordered by what unlocks the deal.

Email:
[paste the whole email]

Thirty seconds later you are negotiating from a brief instead of a feeling. Run it on the whole thread when they reply, not just the last message.

Worked example: the 40-page contract

This is not legal advice and the model is not your attorney. It is the fastest first-pass reader you will ever hire, and it makes the call with your attorney shorter and sharper.

I'm the buyer in this purchase agreement, closing with seller financing. Read the full contract below and give me:

1. The 5 clauses most likely to hurt me, each with the exact quoted text and a plain-English explanation of the risk.
2. Every date and deadline in the document, as a table: what, when, what happens if missed.
3. Anything unusual compared to a standard purchase agreement in this state.
4. Anything that contradicts another part of the same document.

Quote exact text for every finding. If you're unsure about something, say so rather than guessing.

Contract:
[paste all of it]

The "quote exact text" instruction matters. It forces the model to ground every claim in the document, which kills most hallucination, and it lets you verify in five seconds by searching the quote.

Worked example: the rent roll you did not build

A wholesaler sends a blurry photo of a rent roll spreadsheet. Both Claude and ChatGPT read images. Paste the photo with this:

This is a rent roll for a property I'm evaluating. From the image:

1. Rebuild it as a clean table: unit, tenant, rent, lease end, notes.
2. Total monthly and annual gross rent.
3. Red flags: below-market rents, month-to-month leases, vacancies, any unit where the numbers look inconsistent.
4. Three questions the seller should answer about this rent roll.

If any cell is unreadable, mark it [unreadable] instead of guessing.

The last line is habit five in action. An unreadable cell guessed wrong is how a bad number gets into your underwriting.

Where Level 1 breaks

Do this for two weeks and you will hit the ceiling I described in the first post. You will notice you are pasting the same paragraph about your buy box into every conversation. You will notice your VA cannot reproduce your results, because the magic lives in how you happened to phrase things. You will notice Tuesday's analysis contradicts Friday's because nothing persists.

That is not failure. That is the signal that you have outgrown the rung. The fix is moving your context out of your head and into a place the model reads every single time, which is Level 2, and it takes about an hour to set up. That is the next post: a buy box document the model actually follows, plus the follow-up move most people miss, matching deals to your buyers' buy boxes on the dispo side.

The series

1. 00The Five Levels of AI in a Real Estate BusinessThe map
2. 01Getting real value out of a chat windowYou are here
3. 02Building a buy box that the model actually followsPublished
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