Lesson 02 · 14 min read
Prompting for Commercial Real Estate Work
A seven-part prompt structure for commercial real estate tasks, how to ask for page citations, how to iterate, and a reusable prompt library with ready-to-adapt examples.
The difference between a useless AI answer and a useful one is usually the prompt. "Summarize this OM" gets you a brochure in paragraph form. A prompt that names your role, the deal, the exact fields you need, the format, and the rule that every number must cite a page gets you a deal sheet you can check in ten minutes.
This lesson gives you a seven-part prompt structure, shows how to supply documents and demand citations, explains how to iterate when the first answer is weak, and walks through five prompts you can adapt. It ends with a worked example on a Winter Park office lease where one follow-up question caught a wrong rent date.
Why structure matters
A language model fills gaps with whatever is most likely. If you do not say who the reader is, it writes for a general audience. If you do not say what format you want, it writes paragraphs. If you do not say "only from the document," it mixes the document with its general training, and you cannot tell which is which.
A structured prompt removes those gaps. It also makes the output easier to verify, because you asked for it in a shape you can check against the source.
The seven parts of a good prompt
You will not need all seven every time. For anything that feeds a client, an offer or an ad, use them all.
| Part | What it does | Example wording |
|---|---|---|
| Role | Sets the expertise and point of view | "Act as an analyst for a commercial real estate brokerage reviewing a deal for a buyer client." |
| Context | Gives the situation the model cannot know | "The buyer is a private investor considering a small-bay flex building in Sanford, Florida, using bank financing." |
| Task | States exactly what you want done | "Extract the deal terms below into a table." |
| Source material | Supplies the facts to work from | "Use only the attached OM and rent roll." |
| Output format | Defines the shape of the answer | "A two-column table, then a bulleted list of missing items. Under 300 words." |
| Constraints | Sets the rules | "Do not estimate or fill gaps. Cite the page number for every figure. Label any projection as pro forma." |
| Permission to not know | Gives the model an approved way to say no | "If the documents do not answer a question, write NOT IN SOURCE." |
The last part matters more than it looks. Models are trained to be helpful, so without permission to say "not found," they tend to produce something. Giving them a fixed phrase like NOT IN SOURCE makes gaps visible and searchable.
A reusable template
ROLE: Act as [role, e.g. an analyst at a commercial real estate brokerage].
CONTEXT: [Who the work is for, the property type and location, the decision being made.]
TASK: [The one thing you want done.]
SOURCE: Use only the material below or attached: [list each document by name].
Do not use outside knowledge for facts or numbers.
FORMAT: [Table / bullets / email / formula], [length limit], [columns or headings].
RULES:
- Cite the page or section for every fact and number, like (OM p. 12) or (Lease §4.2).
- Label anything that is a projection, estimate, or seller claim.
- Do not do math in your head. Show each calculation.
- If the source does not answer something, write NOT IN SOURCE.
When done, list anything you were unsure about.
Save this as the first entry in your prompt library. Most of the other prompts in this course are variations on it.
Supplying source material
The best way to reduce invented facts is to give the model the facts.
- Upload or paste the actual document. A rent roll, T-12, lease, OM or report works far better than a description of it.
- Name each document. "Document A is the OM. Document B is the rent roll dated June 30." Then the model can cite by name.
- One job per document when accuracy matters. Asking about five leases at once invites crossed wires. Do them one at a time, then compare.
- Check that it can read the file. Scanned PDFs, photos of pages, and complex tables may come through as garbage or not at all. Ask: "List the section headings and the last page number you can read." If the answer does not match the document, fix the file before going further.
- Strip confidential details when you can. Lesson 7 covers what never goes into a general tool.
Ask for page and section citations
Citations turn an answer you have to trust into an answer you can check. Ask for them in a specific format and then use them.
- Page numbers for OMs and reports: "(OM p. 14)"
- Section numbers for leases and contracts: "(Lease §6.3, Amendment 2 §1)"
- Exact quotes for anything that matters: "Quote the sentence that states the renewal option."
A cited page is a claim, not proof. Models can cite the wrong page or paraphrase a quote. For every money term, date, and right that affects value, open the page and read it. If a quote does not appear in the document word for word, treat the whole answer with more suspicion.
Iterate instead of starting over
The first answer is a draft. Good follow-ups often fix it faster than rewriting the whole prompt.
| Problem with the answer | Follow-up to send |
|---|---|
| A number looks wrong | "Show the calculation for that figure, with the source page for each input." |
| It used outside knowledge | "Remove anything not stated in the documents. Mark removed items." |
| It missed something you know is there | "Check page 22 of the OM. Does it change your answer?" |
| It is too long or too vague | "Rewrite as a table with these columns only: [list]." |
| It sounds too sure | "What assumptions did you make? What would change your answer?" |
| It agreed with your framing | "Now argue the opposite. What are the three biggest risks?" |
| The conversation is long and drifting | Start a new chat, paste the documents again, and use your template |
Long conversations can degrade. The model may start mixing earlier documents with later ones or forget an instruction you gave at the start. When a session gets long or you switch properties, start fresh.
Five prompts to adapt
Each prompt below uses square brackets for the parts you fill in. Edit them to your firm's style and save the versions that work.
1. OM summary into a deal sheet
ROLE: Act as an analyst for a commercial real estate brokerage representing a buyer.
CONTEXT: My client is evaluating [property type] at [city], Florida. They will finance with a bank loan.
TASK: Read the attached offering memorandum and build a deal sheet.
FORMAT: A table with these rows: asking price, NOI as stated, cap rate as stated,
whether NOI is in-place or pro forma, building size, land size, year built,
each tenant with suite size, lease expiration and options, lease type,
expense recoveries, any debt assumption terms. Columns: Item | Value | Source page.
RULES:
- Use only the OM. Write NOT IN SOURCE where the OM is silent.
- Label any figure the OM calls projected, pro forma, or "potential."
- After the table, list the documents I should request to verify these figures.
Lesson 4 builds this into a full workflow.
2. Tour follow-up email
ROLE: Act as a commercial real estate agent writing to a prospective tenant.
CONTEXT: We toured [suite] at [address] on [date]. Facts to include:
[asking rent as quoted by landlord], [size], [availability date], [two features they liked],
[one concern they raised and what I learned about it].
TASK: Draft a short follow-up email.
FORMAT: Under 150 words. Plain, friendly, professional. One clear next step.
RULES:
- Use only the facts above. Do not add features, rates, or promises.
- Do not describe the rent as "below market" or similar unless I said so.
- End with my signature block placeholder: [signature with licensed brokerage name].
3. Lease question with citations
ROLE: Act as a lease analyst. You are not giving legal advice.
SOURCE: Document A is the base lease. Document B is the First Amendment.
Document C is the Second Amendment. Later amendments control over earlier documents.
TASK: Answer these questions:
1. When does the current term expire?
2. What is base rent today, and when and how does it change?
3. Does the tenant have renewal options? On what notice?
RULES:
- For each answer, quote the controlling sentence and cite document and section.
- If an amendment changed the answer, say which one and what it replaced.
- If unclear or not stated, write NOT IN SOURCE or UNCLEAR and explain why.
Lesson 5 covers lease abstraction in depth, and Reading a NNN Lease covers the clauses themselves.
4. Spreadsheet formula help
I use [Excel / Google Sheets]. My layout:
B2 = purchase price, B3 = loan-to-value as a percent, B4 = annual interest rate,
B5 = amortization in years, B6 = year-1 NOI.
TASK: Write formulas for:
- loan amount in B8
- annual debt service in B9 (monthly payments, fully amortizing)
- DSCR in B10
- debt yield in B11
FORMAT: For each, give the formula, then one sentence on what it does.
Then give me a test case with simple numbers I can check by hand.
Do not change my cell layout.
Asking for a test case is the habit that catches formula errors. Lesson 6 uses it. If you are new to model layout, start with Pro Forma Structure.
5. Market report summary
ROLE: Act as a research assistant for a commercial real estate agent.
SOURCE: The attached [publisher] [property type] market report for [market], [quarter/year].
TASK: Summarize the report for a client update.
FORMAT: Two sections.
"Reported facts": bullets of measured figures (vacancy, absorption, deliveries,
asking rents, sales volume) exactly as the report states them, each with page number.
"Forecasts and opinions": bullets of anything the report predicts or interprets,
each with page number.
RULES:
- Do not combine figures from different time periods or geographies.
- Note the report's definition of the market area if given.
- Do not add any numbers that are not in the report.
Splitting facts from forecasts keeps a publisher's opinion from turning into your stated fact. To build your own brief from several sources, see Building a 30-Minute Market Brief.
Build a reusable prompt library
Prompts that work are assets. Keep them.
- One place. A shared document, a note app, or a folder of text files. Many assistants also let you save custom instructions or reusable projects; check what your tool currently offers.
- Name by task. "OM deal sheet v3," "Tour follow-up," "Lease Q and A with cites."
- Bracket the variables. Anything that changes per deal goes in square brackets so you do not leave a prior client's details in the next prompt.
- Version and date. When you improve a prompt, keep the old one until the new one has worked on a few deals.
- Note the failure you fixed. "Added NOT IN SOURCE rule after it invented a tenant." That note teaches the next person on your team.
- Keep confidential examples out. The library should never contain a real client's numbers.
A worked example: the Winter Park office rent bump
Illustrative numbers. Not market data.
An agent represents a buyer of a small Winter Park office building. One tenant occupies 3,200 SF. The agent uploads that tenant's lease and asks a quick question with no structure:
"When does this tenant's rent go up and to what?"
The answer: "Rent increases 3% to $18.54 per SF on January 1, 2026."
The figure looks plausible. The agent follows up with the lease-question template, asking for the controlling sentence and section. The second answer quotes Lease §4.2: base rent of $18.00 per SF, increasing 3% "on each anniversary of the Commencement Date." It then quotes §1.4: the Commencement Date is March 1, 2025.
So the first answer had the rate and the new rent right but the date wrong. The model assumed a calendar-year bump. The correct schedule:
| Period | Rent per SF | Annual rent (3,200 SF) | Monthly rent |
|---|---|---|---|
| Mar 1, 2025 to Feb 28, 2026 | $18.00 | $57,600 | $4,800 |
| Mar 1, 2026 to Feb 28, 2027 | $18.54 | $59,328 | $4,944 |
Two months of difference sounds small. But if the buyer's model assumed the bump in January and closing prorations were set from it, the rent figures on the closing statement and the first months of the buyer's cash flow would both be off. The agent also asks one more follow-up: "Does the lease say how rent per SF is rounded after year 2?" The answer is NOT IN SOURCE, so year 3 ($18.54 × 1.03 = $19.0962) goes on the question list for the tenant estoppel.
One structured follow-up, with a demand for the controlling sentence, found the error. That is the pattern for the rest of the course.
Common mistakes
- One-line prompts for high-stakes work. Brevity invites the model to fill gaps.
- No "not found" instruction. Without it, gaps get filled with guesses.
- Mixing documents without labels. The model cannot cite what it cannot tell apart.
- Accepting the first answer. Ask for the calculation, the quote and the assumptions.
- Copying a prior deal's prompt with its details still in it. Use brackets.
- Asking leading questions. "Why is this a strong deal?" gets cheerleading, not analysis.
What to take away
- Structure prompts with role, context, task, source, format, constraints, and permission to say NOT IN SOURCE
- Supply the actual documents, labeled by name, and confirm the tool can read them
- Ask for page or section citations and exact quotes, then open the source and check them
- Iterate with targeted follow-ups: show the math, list assumptions, argue the other side
- Start a new chat when a session gets long or you change properties
- Ask for a hand-checkable test case whenever AI writes a formula
- Separate reported facts from forecasts when summarizing any report
- Keep a short, versioned prompt library with bracketed variables and no client data
Next lesson: AI for Prospecting and Owner Research.