When an AI answer comes back useless, the cause is almost never the tool — it is a missing part of the prompt, and prompt anatomy is how you find it. A useful answer contains four things because the request contained four things: a role, the context, a precise task, and the shape of the output. Take one away and the answer sags in a predictable way. Drop the context and it turns generic. Drop the task boundary and it wanders. Drop the format and you spend ten minutes reformatting. The reason two brokers on the identical subscription get gold and garbage is not luck and it is not a better model — it is that one of them wrote a complete request and the other wrote four words. This piece takes the anatomy apart from the output side: how to read a bad answer, name the part that is missing, and put it back.
If you have already met the basics — what a prompt is and why writing one is the first skill to learn — this is the next question that matters: not what is a prompt, but why did mine come back useless, and which part do I fix. For the ground floor, our explainer on what a prompt is and how to write one covers the mechanics. Here the subject is diagnosis, which is the skill a busy operator actually spends their day using, because bad answers arrive before good ones do.
Why the Same Tool Gives One Broker Gold and Another Garbage
Sit two people from the same firm in front of the same tool — ChatGPT, Claude, Gemini, or Microsoft Copilot, it does not matter which — and hand them the same lease to summarize. One gets back a clean table they drop into the deal file after a check. The other gets a rambling paragraph that buries the rent and misses the renewal option, shrugs, and decides AI is not ready for real work. Same software, same document, opposite result. The variable was the instruction.
This is the fact most firms miss when they shop for AI. They treat a disappointing answer as evidence about the tool, when it is almost always evidence about the prompt. The general-purpose tools are close enough in ability that, for the language work a brokerage does daily, the gap between any two of them is small next to the gap between a complete request and a bare one on any of them. That is the whole reason a lean shop can out-operate a much larger competitor: the lever that changes output is a learnable skill, not a bigger software budget.
Once you accept that the answer is a mirror of the request, the useful question stops being “which tool” and becomes “what was missing from what I asked.” Answering that quickly is the skill. It starts with knowing the parts.
The Four Parts a Useful Answer Is Made Of
A useful answer is built from four things you put in the prompt. Think of briefing a sharp new analyst who is fast and literal but knows nothing about your specific deal until you tell them.
Role — who you want the tool to be. “You are a commercial real estate analyst.” “Act as a leasing broker writing to a prospective industrial tenant.” Naming the role sets the register and the working assumptions, and it is the difference between a generic business paragraph and one that reads like your profession wrote it.
Context — what it needs to know. The deal facts, the audience, the constraints. “This is a 10-year lease with three months of free rent and a 3% annual escalator.” “The recipient is a logistics company we have not worked with before.” The tool knows nothing about your situation beyond what the prompt contains. Missing context is the single most common reason an answer comes back useless.
Task — the exact thing to produce. Not “help with this lease” but “list the base rent, term, renewal options, CAM structure, and assignment clause in a table.” Precise verbs, a precise deliverable, and a clear boundary around what to include and what to leave out.
Format — the shape you want back. “In a five-row table.” “As three short bullets.” “In two paragraphs under 150 words, professional tone.” Specifying the shape saves the reformatting and hands you output you can use as it stands.
Role, Context, Task, Format. You do not need all four on every job — a one-line rewrite needs less than a full market write-up — but these four parts are the anatomy, and every useless answer is missing at least one of them. The framework version of this, built for a team to standardize on, is laid out in our prompt skills framework for CRE professionals; what follows is how to run it backwards, from a bad answer to its cause.
Reading the Signature of a Useless Answer
The diagnostic move is this: a useless answer fails in a specific way, and each way points at a specific missing part. Learn the four signatures and you stop guessing.
Generic and hollow → missing context. The answer reads like it could have been written for any firm, any building, any tenant. It states the obvious, hedges, and tells you nothing you did not already know. That is the signature of a prompt that named a task but starved it of facts. The fix is not a better tool; it is the deal specifics the request left out.
Wandering and unfocused → missing task boundary. The answer covers clauses you never asked about, skips the ones you needed, and makes you hunt for the number that mattered. It “summarized the lease” the way a first-day intern would — everything, in no order, prioritizing nothing. The fix is a task that names exactly what to include and, just as important, what to ignore.
Wrong shape → missing format. The content is fine but it arrived as five paragraphs when you needed a table, or as a table when you needed a two-line email. You spend the saved time reformatting. The fix is one sentence naming the shape.
Wrong register → missing role. The answer is competent but tone-deaf — too casual for an investor update, too stiff for a broker-to-broker note, generically “professional” in a way your recipient would clock as machine-written. The fix is telling the tool who it is and who it is writing to.
There is a fifth failure that looks like the others but is not a prompt problem at all — the confidently wrong answer, where the tool states a fabricated figure or a misread clause with total assurance. No missing part explains that one, and no added part fixes it. We come back to it below, because knowing which failures anatomy cannot solve is as important as knowing which ones it can.
A Useless Answer and a Useful One, Dissected
Nothing makes the anatomy clearer than the same job done twice. Take a task every brokerage runs weekly: summarizing a lease for an acquisitions review.
The useless request:
“Summarize this lease.”
Paste the lease under it and you get a summary, even a plausible one. But it decided for you what mattered, wandered across clauses you do not care about, skipped the renewal option you did, and handed you prose you have to re-read to find the rent. Run the diagnosis: it is generic and it wanders, so it is missing both context (the purpose and audience) and a task boundary (which terms, in what order). Two parts absent, two failure signatures — exactly as the anatomy predicts.
The useful request:
“You are a commercial real estate analyst. Summarize the lease below for an acquisitions review. Return a table with these rows: tenant, premises size, base rent and rent schedule, lease term and commencement, renewal options, CAM/operating-expense structure, assignment and subletting rights, and default provisions. Flag anything unusual in a short note under the table. If a term is not stated in the lease, write ‘not specified’ rather than guessing.”
Same tool, same lease, ninety seconds more thought. This version pins the role, states the purpose and audience, names the exact eight terms to extract, fixes the format as a table, and — the line most people leave out — tells the tool what to do when a fact is absent so it does not invent one. What comes back is something you check and drop into the file, not a paragraph you rebuild. The gap between those two outputs is not a better model. It is the same model, driven by someone who put the four parts in.
The Two Parts Everyone Forgets
Two ingredients separate an answer that is merely adequate from one you would send without touching, and both are easy to omit because they feel optional.
The audience, named explicitly. “Summarize this for an acquisitions review” and “summarize this for a first meeting with a nervous seller” should produce different documents, and they will — but only if you say which. The audience is context, and it is the context people skip most, because in their own head it is obvious. It is never obvious to the tool.
The instruction for missing facts. The single most valuable line in the useful lease prompt above is the last one: if a term is not stated, write “not specified” rather than guessing. Without it, a tool asked to fill eight rows will quietly invent the two it cannot find, and an invented CAM structure in a deal file is worse than a blank one. Telling the tool how to handle absence is the difference between a draft you trust and one you have to audit line by line.
Both of these are prompt anatomy too — they live inside context and task — but they are the parts a self-taught user discovers last and a trained one uses from the start. That gap is most of what separates a firm that dabbles from a firm that runs on the skill.
Where Anatomy Stops Helping
Honesty about the limits is what separates useful training from a sales pitch. A complete prompt improves how a tool handles language. It does not turn the tool into something it is not, and three failures survive any anatomy.
Arithmetic and exact figures. These tools predict plausible text, not verified calculation. Ask one to compute effective rent across a term with free months and escalators and it may hand back a confident wrong number no matter how well you phrase the request. Keep the math in the spreadsheet; use the prompt for the words around it.
Fabricated fact. No prompt stops a tool from occasionally stating something false with full assurance — an invented figure, a misremembered statute, a citation that does not exist. That behavior, called hallucination, is a property of how the technology works, not a phrasing gap. This is the confidently wrong signature from earlier, and its only fix is a human who owns the facts. The tool drafts; a person verifies every number, legal term, and market claim.
Judgment. Whether the deal is good, whether to push on a term, what the client actually needs — a prompt cannot outsource that, and you would not want it to. The anatomy makes routine drafting faster so your judgment goes where it earns its keep.
A firm that trains the skill trains the limits alongside it. Knowing which four failures a better prompt fixes and which three it never will is what keeps AI a tool your team controls rather than one that quietly puts a wrong number in front of a client.
The Prompt Is Also Where Your Data Goes
There is one more reason the prompt deserves care, and a careful principal raises it first. The prompt is not only your instruction — it is the doorway your firm’s information walks through to reach the tool. When you paste a lease, an offering memorandum, or a rent roll into the box to give the tool its context, that content now sits in the prompt, and the question of what is allowed to go there is a real one.
So writing a good prompt and deciding what belongs in a prompt are the same lesson taught from two sides. A fluent user writes the sharp instruction and, in the same motion, classifies the material: public and internal facts go in freely, NDA-bound or personal data is redacted or kept out, and the whole team works from one paid business tier where the vendor states it does not train on your input by default. The parts of a custom setup that can hold a firm’s context safely are worth understanding too — our explainer on what a custom GPT is and what CRE teams build with them covers where reusable context lives. Getting the instruction right and getting the input right are two halves of one skill, and prompt anatomy is where both are practiced.
Building the Diagnostic Habit in a Small Firm
The encouraging part is how small the lift is. This is not a certification or a new hire. It is a repeatable structure — Role, Context, Task, Format — plus the reverse move of reading a bad answer for its missing part, applied to the two or three documents your firm produces most and practiced until it is automatic. A team of ten can reach genuine fluency in an afternoon of guided practice and a couple of weeks of doing real work with it.
The mistake to avoid is the common one: buy the subscription, hand it to the team, and hope. A powerful tool given to people who have not learned to drive it produces a few cautious experiments and quiet disuse — the exact pattern behind “we tried AI and it did not do much.” What changes that is teaching the anatomy and the diagnosis together, so that when an answer comes back useless nobody concludes the tool failed; they ask which part was missing and add it. The ordered version of that program, sequenced so the skill lands before the tooling budget, is the CRE AI training playbook, which starts here — with the skill — rather than with a shopping list.
Where to Start
You do not need to become technical to read a bad answer and fix it, and you do not need to guess which of your firm’s tasks are the right ones to practice on. A free AI-readiness assessment is a short working session that looks at your actual mix of brokerage, management, and acquisitions work, finds the handful of document-heavy tasks where a complete prompt pays off fastest, and shows your team the four-part anatomy — and the diagnosis that goes with it — on your own documents. Book a free AI-readiness assessment and your team leaves able to tell, at a glance, why an answer came back useless and exactly how to make the next one useful.
Frequently Asked Questions
What is prompt anatomy?
Prompt anatomy is the set of parts that make up an effective request to an AI tool: a role (who you want the tool to be), context (the deal facts and audience it needs), a task (the exact deliverable), and a format (the shape of the output). A useful answer contains these because the prompt did; a useless answer is missing at least one. Understanding the anatomy lets you both write a complete request and, just as usefully, read a disappointing answer and name the part that was left out.
What actually separates a useless AI answer from a useful one?
The completeness of the request, not the choice of tool. On the same subscription, a four-word prompt returns a generic, wandering answer, while a request that names the role, context, task, and format returns work you can nearly use as-is. Two people at the same firm get opposite results from the same software for exactly this reason. The answer is a mirror of the prompt, so the difference lives in what you asked, not in which tool you asked.
How do I tell which part of my prompt was missing?
Read the failure signature. A generic, hollow answer means missing context — the deal facts the request starved it of. A wandering answer that covers the wrong clauses means a missing task boundary. Content in the wrong shape means a missing format instruction. A competent but tone-deaf answer means a missing role. Match the way the answer failed to the part that causes that failure, add the part, and re-ask.
Why do two people at the same firm get different results from the same AI tool?
Because they wrote different prompts. The tool is identical; the instruction is not. One person supplied the role, the deal context, a precise task, and the output format, and got a usable draft. The other typed a bare request and got a generic paragraph. The gap in output is the gap in input. This is why AI results are a training question rather than a tool-purchasing question for most small firms.
Can a better prompt fix every bad AI answer?
No, and knowing the exceptions is part of the skill. A complete prompt fixes answers that are generic, wandering, wrongly shaped, or tone-deaf. It does not fix three things: arithmetic (keep effective-rent math in your spreadsheet), fabricated facts stated with confidence (a human must verify every figure and legal term), and judgment (whether a deal is good is yours to decide). A prompt improves how the tool handles language; it does not make the tool reliable at math, fact, or decision-making.
What is the most commonly forgotten part of a prompt?
Two tie for it. The first is the audience — “summarize this for an acquisitions review” and “summarize this for a nervous first-time seller” should produce different documents, but only if you say which. The second is the instruction for missing facts: telling the tool to write “not specified” rather than inventing a value it cannot find. That single line is the difference between a draft you trust and one you have to audit clause by clause.
Do I need technical skills to diagnose a bad AI answer?
No. The whole method is in plain English and closer to briefing a new analyst than to programming. You read how the answer failed, name the missing part in ordinary terms — it lacked the deal facts, it did not know who it was writing for, it came back in the wrong shape — and you add that part. No coding, no settings, no IT department. That low barrier is exactly why a small firm can build the skill without hiring anyone technical.
Is it safe to put lease or deal data into a prompt?
It can be, with the right handling, because the prompt is where your data enters the tool. Two things govern it: whether the vendor trains on your input (the business and enterprise tiers of the major AI products state they do not, by default) and whether your NDA allows sharing the material at all. The safe practice is to put the team on one paid business tier and classify each document before pasting — public and internal facts go in freely, NDA-bound or personal data is redacted or kept out.
How long does it take a small firm to learn this?
Faster than most expect. The anatomy — Role, Context, Task, Format — plus the reverse habit of reading a bad answer for its missing part takes an afternoon of guided practice to grasp and a couple of weeks of real work to make automatic across a team of ten. The barrier is not difficulty; it is the common mistake of buying a subscription and handing it over with no training, which produces a few experiments and then disuse. A short, structured push on your own documents turns curiosity into a shared skill.
Dirk Jan van Veen, PhD
