- ◆Most bad AI responses are not a tool problem. They are a prompt problem. This post covers the four mistakes that cause most of the frustration.
- ◆Mistake 1: the task is too vague. If your prompt means ten different things, you will get ten different kinds of wrong.
- ◆Mistake 2: not enough context. AI knows nothing about you, your clients, or your agency. If you do not tell it, it guesses. Generic in, generic out.
- ◆Mistake 3: stacking tasks. One prompt, one job. Asking for three things at once produces three mediocre outputs instead of one good one.
- ◆Mistake 4: burying the task at the end. AI weights what comes first. Lead with the task, follow with the context. Not the other way round.
- ◆Includes a fully annotated prompt example, before and after examples for estate agents, letting agents, and mortgage brokers, and a ten-second check that catches most problems before they happen.
Most people who give up on AI do not have a tool problem.
They have a prompt problem.
You type something in. You get something back that is almost right but not quite. You edit it for ten minutes. At the end you have spent more time than if you had written it yourself. You close the tab and decide AI is not for you.
The fix usually comes down to one thing they missed.
Tell AI exactly what you want, who it is for, how long it should be, and what to leave out. That structure is what separates a prompt that produces something useful from one that produces something you have to fix.
The rest of this post goes through the four specific mistakes that break that structure, with before and after examples for UK property professionals. If you want the full five-part formula first, the AI Foundations prompt guide covers it from scratch.
The Four Mistakes That Produce Bad AI Responses
Mistake 1: The Task Is Not Specific Enough
AI does not know what a good vendor update email looks like at your agency. It knows what a generic one looks like. Ask for a generic one and that is what you get.
The task is specific when it means one thing. If your prompt produces ten different correct answers, it is not specific enough.
“Write an email to a vendor.”
“Write a two-paragraph update email to a vendor whose buyer had their mortgage offer confirmed yesterday. We are waiting on searches expected back within ten working days. Tone: reassuring and professional. No more than 100 words.”
The after version takes thirty extra seconds to write. It saves five minutes of editing.
Mistake 2: Not Giving AI Enough Context
Every time you open a new AI chat, it knows nothing about you. Not your business, not your clients, not how you write. You are briefing a capable person who has never met you.
The output you get reflects that. Generic tone. Generic structure. Something that reads like it was written by a competent stranger, because it was.
“Write a property description.”
“Write a property description for a three-bedroom end-of-terrace in Nottingham. Large rear garden, recently fitted kitchen, catchment for the local outstanding primary school. Target buyer is a young family. 120 words. Warm, factual tone. Avoid words like stunning or beautiful.”
Build a context block you reuse every time
Write two or three sentences that describe your business, your clients, and your tone. Paste them at the start of any prompt. One minute to write once, saves time on every prompt after that. Post 2 in this series covers exactly how to build one.
Mistake 3: Asking for Too Many Things at Once
One prompt. One job.
The more tasks you stack into a single prompt, the worse AI does each one. It spreads itself thin. You get three outputs that are each about 70% there instead of one that is 95% there.
People do this to save time. It does not save time. It creates more editing.
“Write me a vendor update, chase up the solicitor on the Bedford Road sale, and sort a reply for the Rightmove lead from this morning.”
That is three separate briefs with three separate tones, three separate recipients, and three separate purposes. AI will attempt all three. None of them will be right.
Split it. Pick the first task and prompt it properly.
“Write a vendor update email. Buyer’s mortgage offer confirmed yesterday. Searches ordered, expected back within ten working days. No issues to flag. Tone: reassuring and professional. Do not give any specific exchange dates. Under 100 words.”
Then do the solicitor chaser. Then do the portal reply. Three prompts, each taking sixty seconds to write, each producing something close to usable. Compare that to twenty minutes untangling a combined output that got none of them right.
Mistake 4: Burying the Task at the End
AI reads everything in a prompt but weights what comes first. Lead with four lines of context and add the actual task at the end, and the output reflects that structure. The context dominates. The task gets less attention than it deserves.
“My vendor has been on the market six weeks with no offers. One second viewing booked Thursday. She is feeling deflated. Write her an update email.”
“Write a warm, honest vendor update email. Vendor is deflated after six weeks with no offers. Twelve viewings total, one second viewing booked Thursday. Acknowledge the situation honestly, give context about the second viewing, end with a suggestion to speak briefly this week. Under 120 words.”
Task first. Context second. Every time.
What a Good Prompt Actually Looks Like
Before and after examples are useful. What is more useful is understanding why the after version works.
The five elements of a prompt that works
Task: what you want AI to produce. Context: the specific situation. Tone: how it should sound. Constraints: what to leave out. Format: length and structure. You do not need all five every time. You do need the first two every time.
Here is a fully annotated example using a letting agent arrears chaser.
“Write an email chasing late rent.”
“Write an arrears chaser email to a tenant who is twelve days late on their monthly rent payment. This is the first time they have been late in two years of tenancy. Tone: firm but understanding. Do not use threatening language or reference legal proceedings. Include the amount owed and a clear instruction on how to pay. Ask them to get in touch if there is a problem we should know about. Under 100 words.”
Why each part is in there:
“Write an arrears chaser email” — the task stated first and clearly. AI knows exactly what type of output to produce.
“Twelve days late, first time in two years of tenancy” — context that changes the tone. Without it, AI defaults to a standard late payment template with no memory of the relationship.
“Firm but understanding” — tone instruction. Without it, AI picks one or the other. Usually the wrong one.
“Do not use threatening language or reference legal proceedings” — a constraint. This is the most underused part of a prompt. Telling AI what to leave out is often more effective than telling it what to include.
“Under 100 words” — format instruction. Without a word count, AI writes however long it wants. Usually too long.
That structure (task, context, tone, constraint, format) is the difference between an output you send and an output you rewrite.
Before and After: Three Property Roles
Plain English definitions, one search
If any terms in this post are unfamiliar, the Jargon Buster covers AI language in plain English. One search, one sentence answer. No sign-up needed.
Estate Agent: Portal Lead Reply
“Reply to a Rightmove enquiry about a three-bed house.”
“Write a brief, friendly reply to a Rightmove enquiry about a three-bedroom semi in Hucknall at £240,000. The enquirer asked about the garden size and parking. Confirm both: 60-foot rear garden and a driveway for two cars. Ask one question back about their timeframe for moving. Under 80 words. Sign off with your first name.”
For a full set of portal reply templates, the guide to replying to Rightmove and Zoopla enquiries faster has six ready-to-use prompts covering every enquiry type.
Letting Agent: Rent Reminder
“Write an email chasing late rent.”
“Write a polite but clear rent reminder to a tenant who is seven days overdue on their monthly payment. First contact on this. Tone: professional and friendly, not threatening. Include the amount outstanding and payment instructions. Do not reference legal action. Under 80 words.”
Mortgage Broker: Case Update
“Update my client on their mortgage application.”
“Write a brief case update email to a first-time buyer. Application was submitted eight days ago. Lender has confirmed receipt. We are currently awaiting the valuation booking, expected within five working days. Tone: reassuring and clear. Do not speculate on timelines beyond what is confirmed here. Do not mention specific products or rates. Under 80 words.”
The Check That Takes Ten Seconds
Before you send any prompt, ask yourself one question.
Does this mean more than one thing?
If yes, add detail until it does not.
Go back to the letting agent example above. Every element in that prompt meant one thing. There was only one logical response. That is the test. One question, ten seconds, catches most of the problems on this page before they become output you have to fix.
Where to Go From Here
This series has four more posts after this one. Each covers a specific technique that changes what AI produces.
Most people who work through all five land on the same conclusion: the quality of their output was never a tool problem.
Post 2 covers context in full. What to include, how to structure it, and how to build a reusable business introduction that improves every prompt you write from here.
Post 3 covers role prompting. Adding five words to a prompt changes what comes back. Post 3 shows you which five words and why they work.
Post 4 covers tone and format. How to get the right length, the right register, and the right structure without rewriting half the output.
Post 5 covers what to do when the first draft is close but not quite there. The three follow-up moves that fix most outputs without starting again.
Want the full prompt formula first?
AI Foundations Blog 3 covers the full structure if you want to go deeper before continuing. This post focuses on the mistakes that happen before the formula.
Ten Prompts Worth Stealing
Ten property-specific prompts ready to copy. Portal lead replies, vendor updates, tenant chasers, viewing confirmations. Formatted and ready to use.
Get the Free Prompt Pack → Only Email Signup NeededQuestions People Actually Ask
The prompt is too vague. AI produces the most likely answer to the question it receives. Add a specific task, context, tone, and word count, and the output changes. Vague in, vague out.
Ask whether your prompt means more than one thing. If it does, narrow it down. Add the situation, the tone, the word count, and one or two things to leave out.
Not giving enough context. AI does not know who you are or what your business sounds like. Include your role, the specific situation, and who the output is for.
No. Politeness does not change the output. Specificity does.
Yes. Save any prompt that produces a result close to what you would have written. Change the specific details next time, keep the structure. That is how a prompt library builds up.
Tasks with clear structure, a defined output, and no need for verified facts suit AI well. Emails, summaries, descriptions. Tasks that need accurate data, regulated advice, or local knowledge need a human checking the output before it goes anywhere.
How to Give AI the Context It Needs to Actually Help You
AI does not know who you are, what you do, or what you need. Post 2 covers exactly what context to include in every prompt and how to build a reusable business introduction that makes every prompt you write from here better.
Read Part 2 →