Why "write an owner update" goes wrong
Ask a chat tool to "write an owner update for 14 Elm Street" and it will. It will be polite, well organised and partly invented, because you did not give it the statement and you did not tell it to stop when it lacks a fact. The model fills gaps with likely-sounding text. That is what it is built to do.
So the fix is not a cleverer phrase. It is giving the tool the source, the job and the boundaries, the same way you would brief a new assistant on their first morning.
The five parts of a prompt that works
A prompt has five parts
| Part | What to write | What goes wrong without it |
|---|---|---|
| Source | "Using only the attached statement" | It answers from general knowledge and guesses |
| Job | "Draft a short email to the owner" | You get an essay, a list or the wrong reader |
| Rules | "Do not promise a date or mention legal action" | It makes commitments your firm never made |
| Gaps | "If a charge has no description, write NEEDS A HUMAN" | It invents a reason that sounds right |
| Check | "List anything you could not confirm at the end" | You cannot see where to look |
The gaps line does most of the work. A model told what to write when a fact is missing will usually write that instead of a guess. A model that was never told will guess every time.
Before and after
An owner update
Before: "Write an update to the owner of 14 Elm about this month."
Using only the attached closed statement for 14 Elm Street, draft a short email to the owner. First two lines: what they received and what kind of month it was. Restate the five statement totals exactly. Explain each expense only from its memo. If a charge has no memo, write NEEDS A HUMAN. Do not say anything about next month. List anything you could not confirm at the end.A maintenance reply
Before: "Reply to this tenant about their leak."
Using the after-hours policy in this Project, draft a reply to the resident message below. If it mentions water spreading, gas or burning, follow the emergency section of the policy word for word. Otherwise, confirm we received it and say when they will hear from us next, using the response times in the policy. Do not promise a visit time. Do not suggest they repair anything themselves.
[resident message]A vendor chase
Before: "Email the plumber about the late job."
Using the attached open work order, draft a follow-up email to the vendor. Name the work order number, the property and unit, and the date it was sent. Ask for one thing: a visit date and time window by [date]. Do not include the resident's name or phone number. Keep it under 80 words.None of the "after" prompts are clever. They are longer because they carry the decisions you would otherwise make while fixing the draft.
Put the standing parts in a Project
Typing your voice, approval limits and firm rules into every prompt gets old by Wednesday. Both ChatGPT and Claude let you create a Project with standing instructions and files. Put the parts that never change there.
| In the Project, once | In the prompt, each time |
|---|---|
| Your firm's voice and sign-off | The file for this job |
| Templates you already use | Which template applies |
| Rules: what it never says or promises | Anything unusual about this case |
| What to write when a fact is missing | Who the draft is for |
One Project per repeated job works better than one Project for everything. An owner update Project and a maintenance reply Project can each hold rules that would confuse the other.
Five habits that fix most bad output
- Attach the file instead of describing it. "The rent roll shows" in your words is a summary. The export is the fact.
- Ask for one thing per prompt. Draft the update, then in a new message ask for the subject line. Bundled requests get half done.
- Say who will read it. An owner, a resident and a vendor need different words for the same repair.
- When it is wrong, say what is wrong. "The second expense is the gutter clean, not a roof repair. Fix only that line" beats starting again.
- Save the prompt that worked. Paste it into the Project instructions so next month starts from the good version.
Where prompts stop being enough
A prompt cannot check its own maths. It cannot reliably notice that the expenses on a statement do not add up to the total, and it will not stop a draft going out. For jobs where a wrong number costs you an owner, you want the checks built in around the AI: re-adding the totals, refusing to continue when they disagree, and flagging what a person must look at. That is the difference between a prompt and an agent.
We build agents for property managers, and every one of them started as a prompt like the ones above that kept breaking in the same places. The free Owner Statement Agent is the owner update prompt with those checks added. Run its sample files and compare the output with what your best prompt produces. Before any real file goes in, see how to prepare an export AI can read.
Before you send a prompt
- It names the source, and the source is attached.
- It says what to produce and who will read it.
- It says what the AI must never say or promise.
- It says what to write when a fact is missing.
- It asks for a list of anything that could not be confirmed.
Questions
What is a good AI prompt for property managers?
One that names the source file, says what to produce and for whom, sets what it must never do, says what to write when a fact is missing, and asks for a list of what could not be confirmed.
Do I need special prompt templates for ChatGPT and Claude?
No. The same five parts work in both. Put the parts that never change in a Project so each prompt stays short.
Why does AI make things up in my drafts?
Usually because the prompt did not give it the source or did not say what to write when a fact was missing. Tell it to write NEEDS A HUMAN instead of guessing.
When should I use an agent instead of a prompt?
When a mistake is expensive and needs a check the AI cannot do reliably on its own, such as re-adding statement totals before anything reaches an owner.