Matt Manser: SI for real estate agents

Where I'd start tomorrow: a simple four-step plan for real estate agents learning Claude

Matt Manser — Sun, 01 Nov 2026 09:00:00 +0000

If I had to start again tomorrow, this is exactly what I'd do.

Twenty days of this series is a lot to take in. If you've followed every episode, good. If you've dipped in and out, or you're landing here first, that's fine too. This post is the short version: the order I'd do things in if I were starting from nothing.

There are only four steps. And the most important part isn't the steps themselves. It's the pace.

The four steps, in order

Here's the whole plan in one line: chat for one job, then give it one folder to work in, then connect your email and set up the morning debrief, then save your first skill. One step a week, not all at once.

Each step builds on the one before. Skip ahead and you end up with something powerful you don't trust yet, which usually means you stop using it.

Step 1: Chat for one job

Start with a plain conversation in Claude. No files, no connections. Pick one real job from your week and work through it in the chat.

The point of this step isn't the result. It's learning how Claude responds when you give it clear instructions instead of vague questions. You'll notice quickly that the better you brief it, the better the work comes back.

Step 2: Give it one folder to work in

This is the shift from Day 1. Instead of asking Claude about your work, you hand it a folder and give it a job.

One narrow folder. Copies of files you're comfortable with it reading. A clear instruction about what you want back, and "show me your plan first" at the end. This is where Claude stops being an adviser and starts being an assistant.

Step 3: Connect your email and set up the morning debrief

Once you trust how it handles a folder, connect your email. Claude can link to some email and calendar services through its connectors, which you'll find in its settings. The exact options depend on which email you use, so check what's available for yours.

With email connected, the morning debrief becomes possible: one prompt you run at the start of the day that reads what came in and tells you what needs your attention. For an agent, that's the difference between starting the day buried in the inbox and starting it with a short list.

This is also where the safety habits matter most. Email means Claude could, in principle, send something. Keep approvals on for anything that sends, spends or deletes. Let it read and draft. You decide what goes out.

Step 4: Save your first skill

By now you'll have a job you ask Claude to do the same way over and over. That's your first skill.

A skill is a set of saved instructions Claude can reuse, so you're not re-explaining the same job every time. Take the instructions that work, save them, and the next time you need that job done it starts from your standard rather than from scratch.

This is where the system starts to feel like yours. Not a tool you use, but a way of working you've built.

Why one step a week

It's tempting to do all four this afternoon. I'd hold off.

Each step changes how much you're trusting Claude with. A chat can't touch anything. A folder can, but only what's in it. Email reaches your clients and contacts. A skill runs your way of doing things on repeat. Every step up is worth getting comfortable with before you take the next one.

A week per step gives you time to use each one on real work, notice where it gets things right and where it needs a tighter brief, and build the habits before the stakes go up. Four weeks from now you'll have something that actually runs part of your week, and you'll understand how it works.

Rushing it usually goes the other way. Too much connected too fast, a few surprises, and it gets switched off.

The Day 21 prompt

There's no new prompt today. This one's a recap.

Prompt

(No prompt, recap.)

That's deliberate. By Day 21 the job isn't to learn another prompt. It's to go back and use the ones you already have, in order. Every prompt from the series is there for you to reuse, and the four steps above tell you which to start with.

If you want something to type today, go back to the Day 1 prompt. It's the right starting point for Step 2, and it doesn't change anything.

Try it yourself

Here's the four-week plan, laid out so you can follow it.

Week 1: Chat for one job. Open Claude and work through one real task from your week in a plain conversation. Write instructions, not questions: what to look at, what you want back, and what not to touch.
Week 2: Give it one folder. Make a new folder with copies of a handful of work files. Add just that folder to Claude, run the Day 1 prompt, and let it do one of the jobs it suggests. Ask for its plan first.
Week 3: Connect your email and set up the morning debrief. Look in Claude's settings for the connector that matches your email. Keep approvals on. Run a morning debrief each day this week and see what it surfaces.
Week 4: Save your first skill. Pick the job you've repeated most over the last three weeks. Save the instructions that work as a skill, and use it for the rest of the week.

At the end of each week, ask yourself one question: do I trust this step enough to build on it? If yes, move on. If not, give it another week. There's no prize for finishing fast.

The habits that carry through every step

Whichever step you're on, the same three habits keep you safe.

One narrow folder. Give Claude the files for one job, never your whole computer. As agents we hold a lot of personal client information, and the rules about how that's handled depend on where you work. If it isn't in the folder, Claude isn't reading it.
Approvals on for anything that sends, spends or deletes. Especially once email is connected. Nothing should go out or disappear without you saying yes.
"Show me your plan first." End your jobs with it. You see what Claude intends to do before it does it, and you catch misunderstandings while they're cheap to fix.

You can loosen the reins once you trust it. Start tight.

Where to go from here

That's the end of 21 Days of SI. If you take one thing from the whole series, make it the order: chat, folder, email and debrief, skill. One step a week.

You don't need to build everything at once to get value from it. You need one job done well, then the next.

Missed yesterday? Day 20: My whole back office

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Hi, I'm Matt Manser. I specialise in super intelligence (SI) and artificial intelligence (AI), and what they can do for real estate agents. I'm figuring it all out so you don't have to. I can help with anything from automating parts of your business to setting up your own AI CEO and a team of AI agents working behind the scenes to support you. You can do as little or as much as you like. Whether you're just getting started or ready to see what's possible, I'm here to guide you.

My whole back office: how a real estate agent can organise Claude into watchers, doers and a gate

Matt Manser — Sat, 31 Oct 2026 09:00:00 +0000

This is everything that runs my business while I'm at an open home.

Not a single clever app. A team. Watchers notice things, doers do the work, and a gate checks anything before it reaches a client. Twenty-eight agents, one chief of staff, and me.

For nineteen days this series has been about single jobs. One folder, one prompt, one finished result. Today is about what happens when you have enough of those jobs that they need organising. That's when it stops being a tool you use and starts being a back office.

A quick note on words, as always. I call it SI, super intelligence. Most people say AI. When I say SI I mean tools like Claude, made by Anthropic.

Why it needs a shape

Once you've handed over a few jobs, you hit a new problem. Which job runs when? Who checks the output? What happens if two of them want to do the same thing?

If you don't answer those questions, you end up with a pile of separate tasks you have to babysit. You've swapped doing the admin for managing the admin. That's not much of a win.

The fix is the same one any business uses when it grows past one person. You give everyone a role.

The three roles

Every agent in my back office fits into one of three roles. That's the whole idea. It's simple on purpose.

Watchers notice things

A watcher's job is to keep an eye on something and speak up when it changes. It doesn't act. It just notices.

In an agent's world, that might be an inbox, a set of listings, a calendar, or a list of follow-ups. The watcher's only question is "has something happened that someone should know about?"

Keeping watchers separate matters. Something that only looks can't accidentally send, change or delete anything.

Doers do the work

A doer gets handed a job and produces the finished thing. A draft reply. A tidied document. A summary. A report.

This is the part most people picture when they think of SI. But on its own, a doer just waits. Pair it with a watcher and it has something to work on without you starting it.

The gate checks before it reaches a client

This is the most important role, and the one people skip.

Anything a doer produces that's going to a client goes through a gate first. The gate checks it against your rules. Is it accurate? Is it the right person? Does it say anything it shouldn't? Has it already been sent?

If it passes, it goes. If it doesn't, it stops and waits for a person. That person is me.

The chief of staff and me

Sitting across all of them is one chief of staff. In my case, that's Stirling. Its job is to coordinate the others and keep me across what matters, so I'm not checking twenty-eight things myself.

And then there's me. I set the rules, I make the calls the gate sends my way, and I do the part of the job that should always be a person: the relationships.

Why this matters for an agent's week

Our week doesn't happen at a desk. It happens at open homes, appraisals, in the car and on the phone. The admin doesn't stop while we're out. It just piles up.

A back office with this shape keeps moving while you're away from the desk. Watchers keep watching. Doers keep drafting. And the gate means you're not trusting it blindly. Anything that reaches a client has been checked first, and anything uncertain comes back to you.

That's the difference between automation that worries you and automation you can actually leave running.

The Day 20 prompt

Prompt

(No prompt today, walk through the org chart.)

There's no prompt to copy today, and that's deliberate. This episode isn't a job to hand over. It's the map that all the other jobs sit on.

Here's what each part of that line is telling you:

"No prompt today." Before you add more jobs, step back and look at the ones you already have. You don't need Claude for this part. You need a pen.
"Walk through." Go through it slowly, one role at a time. Watchers, doers, gate, chief of staff, you.
"The org chart." Draw it like you'd draw a team. Who notices, who does, who checks, who decides. If you can't draw it, it isn't organised yet.

Try it yourself

You can sketch your first version in one sitting.

List your recurring admin. Every job that comes up week after week. Don't sort it yet, just write it down.
Mark each one W, D or G. Is it something to notice (watcher), something to produce (doer), or something to check before it goes out (gate)? Most jobs turn out to be a watcher and a doer joined together.
Write your gate rules. A short list of what must be true before anything reaches a client. Plain sentences are fine. Accurate details, right person, nothing sent twice, anything unsure comes to you.
Pick one pair to build first. One watcher and one doer that go together. Set it up the way earlier episodes showed: one narrow folder, a clear brief, and "show me your plan first".
Put yourself in as the gate to start. Until you trust it, every output comes to you before it goes anywhere. Loosen that later, one rule at a time.
Ask Claude to review your chart. Add the folder with your sketch in it and ask Claude to point out gaps, overlaps, or anything that reaches a client without a check. Tell it not to change anything.

Draw it on paper first. It's much easier to fix a box on a page than a job that's already running.

Keep the safety habits

The bigger the system, the more these matter, not less.

One narrow folder per job. Each agent gets access to what it needs and nothing else. As agents we hold a lot of personal information about clients, and the rules about how that's handled depend on where you work.
Approvals on for anything that sends, spends or deletes. That's what the gate is for. Nothing goes out, gets paid for, or disappears without passing it.
"Show me your plan first." Every new job starts with a plan you can read before anything happens.

Watchers that only look, doers that only draft, and a gate that only lets checked work through. Each piece is safe on its own, which is what makes the whole thing safe to leave running.

Where to go from here

You don't need twenty-eight agents. You need one watcher, one doer and a gate, with you in the middle. Get that working, then add the next pair. The org chart grows one box at a time.

Tomorrow is the last day of the series, and we'll pull it all together.

Missed yesterday? Day 19: Texting my AI from the car

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Texting my AI from the car: what a real estate agent's SI assistant looks like once the basics are in place

Matt Manser — Fri, 30 Oct 2026 09:00:00 +0000

I changed a listing price with six words from the car.

That's not a trick, and it's not a special app. It's where all the steps in this series lead once the basics are in place. You have one assistant you can text, that knows your business, can make the change, and checks its own work before it tells you it's done.

Today's episode is a bit different. There's no new prompt to copy. Instead, I'm showing you my real text thread with Stirling, my assistant, so you can see what the finished system looks like from the agent's side: a phone, a short message, and a reply.

What "texting my AI" actually means

Most people picture an assistant on their phone as a chatbot. You ask something, it answers, and you still go and do the work.

This is the other half from Day 1: giving it a job instead of asking it a question. The difference now is where you're sending it from. You're not at your desk with a folder open. You're parked outside an appointment, or between opens, and you send a short message.

For that short message to work, three things have to be true.

It knows your business

Six words only work if the assistant already has the context. It needs to know which listings are yours, what "the price" means on each one, and how you like things done. That's what yesterday's episode was about: giving it a brain. Without that, a short text gets you a string of questions back instead of a finished job.

It can actually make the change

Knowing what you want isn't enough. The assistant needs access to the place the change happens, whether that's your CRM, your website or a document. Without that access, the best it can do is tell you what to change. With it, it makes the change for you.

It checks its own work

This is the part I care about most. A reply that says "done" means nothing unless it's true. A good assistant goes back, looks at the result and confirms it before reporting back. If something didn't work, it should say so plainly, not paper over it.

Why it matters for an agent's week

Our week doesn't happen at a desk. It happens in the car, at appraisals, at open homes and on the phone. The small jobs pile up in the gaps: a price change a vendor has just agreed to, a note to add after a call, an update that needs to go out today.

Usually those jobs wait until you're back at the computer, and by then there are ten of them. Being able to hand one over the moment it comes up, and get a checked answer back, means it's handled while it's still fresh.

The point isn't the phone. The point is that the job gets done properly without you stopping what you're doing.

A quick word on driving

"From the car" means parked. Australian road rules on phone use while driving are strict, and no job is worth it. Pull over, send the message, and get back on the road. The reply will be there when you next stop.

The Day 19 prompt

Prompt

(No prompt today, just show the real thread.)

There's no prompt to copy today, and that's deliberate.

"No prompt today" is the lesson. Once the basics are in place, you stop writing long, careful prompts for everyday jobs. The context, the access and the habits are already set up, so a short, plain message is enough.

"Just show the real thread" is the proof. A finished system is easier to understand by seeing it than by reading about it. Look at how short my messages are, how specific the replies are, and how each reply confirms what was checked, not just that something was done.

Try it yourself

You won't build the whole thing today, and you shouldn't try. But you can take the first step towards it.

Install the Claude app on your phone. Sign in with the same account you use on your computer, so your projects and conversations carry over.
Use the project from yesterday. Open the project where you gave Claude a brain, with your notes on how you work. That's what lets a short message make sense.
Send it a small job that doesn't change anything live. Something like tidying your notes from an appointment you've just left, or drafting a follow-up message for you to check. Keep it to one job.
Ask it to check its own work. Finish with "When you're done, check the result and tell me exactly what you checked." Read how it answers.
Only connect live systems once you trust it. If you later connect Claude to a tool where it can change real listings, start with one tool and keep approvals on so nothing changes without your yes.

Then look at the gap between that and your usual week. Every small job you currently save for later is a candidate.

Keep the safety habits, even on your phone

Short messages make it easy to forget the habits from earlier in the series. They matter more here, not less, because you're not sitting in front of the screen watching it work.

One narrow job, one narrow area. Only give it access to what that job needs. Not your whole computer, not every system you use.
Approvals on for anything that sends, spends or deletes. And for anything that changes a live listing. A price change is a yes you give, not one it gives itself.
"Show me your plan first." For anything new, have it tell you what it intends to do before it does it. It costs you one extra message.
Ask what it checked. "Done" isn't an answer. "Done, and here's what the listing says now" is.

You can loosen the reins as you build trust. Start tight, especially with anything clients or buyers will see.

Where to go from here

This is what the system looks like from the outside: a short text and a checked reply. Everything underneath it is what we've been building one day at a time. That means giving it jobs instead of questions, giving it the right folders, giving it a brain, and keeping the approvals on.

Two days to go. If you've followed along, you already have most of the pieces.

Missed yesterday? Day 18: Give it a brain

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Give it a brain: how to make Claude answer like you by reading your own notes first

Matt Manser — Thu, 29 Oct 2026 09:00:00 +0000

My AI reads my notes before it answers me.

That's the whole idea of today's post. Out of the box, Claude knows a lot about real estate in general. It doesn't know anything about how you do it. It doesn't know how you price, what you say when a vendor pushes back, or what you've learnt the hard way over the years.

So when you ask it something, you get a sensible, general answer. The kind of answer any agent could give. Useful, but not yours.

Give it your notes first, and that changes. It answers the way you would.

Why generic answers aren't enough

Most of what makes a good agent good isn't in a textbook. It's judgement. How you read a vendor. How you explain the evidence. Which words you use and which ones you avoid. What you've seen go wrong before.

That knowledge usually lives in your head, or scattered across old emails, notebooks and the notes app on your phone. Claude can't see any of it. So when you ask it for help with a tricky conversation, it fills the gaps with general advice.

General advice is fine for a first draft. But if you then have to rewrite it to sound like you and match how you actually work, you haven't saved much.

What "giving it a brain" means

It's simpler than it sounds. Nothing to install, no technical setup.

You keep your notes in one place. Plain documents, in plain English. Then you point Claude at that folder and tell it to read your notes before it answers.

Now when you ask a question, it isn't guessing what a typical agent would do. It's working from what you've written down about how you do it.

This builds on Day 1. There, you gave Claude a folder of files to do a job with. Here, the folder isn't the job. It's the background knowledge Claude uses for every job.

What to put in your notes

You don't need a polished manual. Rough notes in your own words work well, because that's the voice you want back. A few ideas for where to start:

How you price. How you read the evidence, how you explain a price range, what you do when the comparable sales don't line up neatly.
How you talk to vendors. Phrases you use, the order you explain things in, how you handle difficult conversations.
What you've learnt. Lessons from past campaigns, written as general rules, not client stories.
How you like things written. Your tone, words you never use, how formal or casual you are.
Your process. The steps you follow from appraisal to settlement.

One note per topic keeps it easy to find and update. A short, clear page beats a long, rambling one.

A word on privacy. Write these as your methods and lessons, not as records about particular clients. No names, addresses or personal details. The rules about client information depend on where you work, so if it isn't needed, leave it out.

The Day 18 prompt

This is the prompt from today's episode. It uses a conversation most agents know well.

Prompt

Read my notes in this folder before you answer. Using only what's in my notes, how would I handle a vendor who wants to list $100,000 above the evidence?

Here's what each part does.

"Read my notes in this folder before you answer." This tells Claude where your knowledge is and that it comes first. Without it, Claude may answer straight from its general knowledge and never open your files.

"Using only what's in my notes" is the important bit. It stops Claude from blending your approach with generic advice from the internet. If your notes don't cover something, you want to see the gap, not have it quietly filled with someone else's method.

"How would I handle..." asks for your approach, not the best approach in general. It frames the answer around how you work.

"A vendor who wants to list $100,000 above the evidence" is a real, specific situation. Specific questions get specific answers. A vague question like "how do I talk to vendors?" gets a vague reply, even with your notes.

That "only what's in my notes" line is also a quick test. If the answer comes back thin, that tells you something useful: your notes on that topic need more detail.

Try it yourself

Make a new folder called something like "My notes". Keep it just for this.
Write two or three short notes. Start with how you price and how you handle a vendor whose expectations are above the evidence. Plain documents are fine. Your own words, no client details.
Open the Claude desktop app and add that folder. Under the message box, choose to work in a project or folder and pick the one you just made. Only that folder.
Paste the Day 18 prompt and send it.
Read the answer against your notes. Does it sound like you? Did it stick to what you wrote? Did it make anything up?
Fill the gaps. If it said your notes didn't cover something, add a line or two to the right note, then ask again.

Then try the comparison. Start a new conversation with no folder added and ask the same question. Put the two answers side by side. One is a general answer. The other is yours.

Keep it a reference, not a free-for-all

This folder is something Claude reads, not something it should be changing. A few habits help.

One narrow folder. Just your notes. Not your whole computer, not your client files. If it isn't in the folder, Claude isn't reading it.
You own the edits. If you want Claude to help tidy or add to a note, ask it to show you the change first. End with "show me your plan first" and say yes only when it's right.
Keep approvals on. Anything that sends, spends or deletes should still need your yes. Reading your notes doesn't change that.
Check what's connected. Any folder you've added shows up in Claude's settings, so look now and then to make sure the list is only what you meant to share.

It gets better as your notes do

The nice thing about this is that it improves without any extra setup. Every time you learn something, add a line to your notes. Every time Claude gives an answer that doesn't sound like you, work out what was missing and write it down.

Over time, your notes become a written version of how you run your business. That's useful well beyond Claude. It's something you can hand to a new team member, or reread yourself before a hard conversation.

And when you pair this with what you've already learnt in the series, giving Claude jobs, having it check its own work, it isn't just doing tasks anymore. It's doing them your way.

Start with two or three notes and today's prompt. See how close it gets. Then keep adding.

Missed yesterday? Day 17: Make it check its own work

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Make it check its own work: the one line that catches Claude's mistakes before you do

Matt Manser — Wed, 28 Oct 2026 09:00:00 +0000

A confident answer and a true answer aren't the same thing.

Claude writes well. It writes clearly, in full sentences, with numbers and dates in all the right places. That's exactly why its mistakes are easy to miss. A wrong figure in a tidy paragraph looks just as believable as a right one.

The fix is simple, and it takes one line. When Claude finishes a job, tell it to go back over what it just made, check every number against the source, and tell you what it couldn't verify. That one step catches most mistakes before they reach you, let alone a client.

Why confident isn't the same as correct

SI (super intelligence; most people say AI) doesn't sound unsure when it's unsure. It doesn't hedge the way a nervous junior might. It writes the same steady sentence whether the number came straight from your file or got mixed up along the way.

Mistakes usually creep in at the joins. A figure copied from one file into another. Two dates that look alike. A name spelled one way in your notes and another way in the contract. A total worked out from numbers that were never quite the same. None of these are dramatic. They're the same small slips a busy person makes, and they're just as easy to skim past.

The good news is that Claude is much better at catching these when you ask it to look for them. Writing something and checking something are two different jobs. If you only ever ask for the first, you only ever get the first.

Why this matters for an agent

Most of what we hand over in real estate has facts in it that people rely on. Prices, dates, addresses, names, land sizes, settlement terms, inspection times. A vendor report, a property summary, a buyer follow-up, a market update. Get one of those wrong and it isn't just embarrassing. It can confuse a client, cost you trust, or worse.

So the more of your week you hand to Claude, the more this matters. Day 1 was about giving it a job instead of asking it questions. Today is the other half of that: once it's done the job, make it check the job.

What checking actually looks like

When you ask Claude to check its own work against your files, it does something different from writing. It goes back to the source, finds where each fact came from, and compares. If the report says one thing and the file says another, it should flag it. If it can't find where a number came from at all, it should say so.

That last part is the most useful. A list of "things I couldn't verify" tells you exactly where to look. Instead of re-reading the whole document yourself, you check the handful of items it wasn't sure about.

It's not a guarantee. Claude can still miss something, and you're still the one who signs off. But it turns "read every line hoping to spot a problem" into "check these few things it flagged", which is a much better use of your attention.

The Day 17 prompt

Use this straight after Claude has made something for you from your files.

Prompt

Go back over what you just wrote. Check every number, date and name against the source files. List anything you couldn't verify, and fix anything that's wrong.

Here's what each part does.

"Go back over what you just wrote." This switches Claude from writing mode into checking mode. It's now reviewing, not producing.
"Check every number, date and name against the source files." These are the three things most likely to go wrong and most likely to matter. "Against the source files" is the key part. It tells Claude to check against your actual documents, not against what sounds right.
"List anything you couldn't verify." This is the line that does the heavy lifting. It makes Claude own up to gaps instead of smoothing over them, and it gives you a short list to check yourself.
"Fix anything that's wrong." If it finds an actual mismatch, it corrects it rather than just pointing at it.

If you'd like to see the fixes before they're made, change the last part to "show me what you'd fix first". That's a good habit while you're still getting a feel for it.

Try it yourself

Set up one narrow folder. Copy in a few files that have real facts in them, such as a property's details, some appointment notes and a price list. Copies, not originals.
Open the Claude desktop app and add that folder. Just that one folder, nothing else.
Give it a job. Ask it to write something from those files, like a one-page property summary or a short follow-up email. Ask it to show you its plan first, then let it go ahead.
Paste the Day 17 prompt as your next message, in the same conversation.
Read the "couldn't verify" list. Open the source files and check those items yourself.
Look at what it changed. Compare the corrected version with the first one so you can see what it caught.

If you want to test it, put a small deliberate mistake into one of your copied files, like two different prices for the same property, and see whether it flags the mismatch. It's a quick way to learn how much to trust the check.

Make it a habit, not a one-off

The real value comes when checking becomes part of every job. A few ways to build that in:

Add it to the end of your instructions. Instead of a separate message, finish a job request with "when you're done, check every number, date and name against the files and list anything you couldn't verify."
Use it before anything goes out. Anything heading to a client, a vendor or a buyer should get a check first. It costs you a minute.
Keep approvals on for anything that sends, spends or deletes. Checking is a safety net, not a replacement for you saying yes. Nothing should go out the door until you've seen it.
Keep the folder narrow. The fewer files Claude is working from, the easier it is for it to trace every fact back to its source, and the easier it is for you to check it too.

And if the "couldn't verify" list keeps coming back long, that's useful information. It usually means the job needs better source files, or a clearer instruction about where to find things.

Where to go from here

Claude will keep sounding confident. That's just how it writes. Your job isn't to trust it less, it's to build the check into how you work, the same way you'd ask a new assistant to double-check a contract before it goes out.

One line after each job. Every number, date and name against the source. A short list of what it couldn't confirm. That's most of the mistakes caught before anyone else sees them.

Missed yesterday? Day 16: The handbrake

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

The handbrake: how to stop Claude sending anything to a client without your say-so

Matt Manser — Tue, 27 Oct 2026 09:00:00 +0000

The day my AI nearly texted my whole database twice.

That's the moment this rule stopped being a nice idea and became the most important part of my system. Today's post is about the handbrake: one rule that sits between Claude and anyone outside your business.

By this point in the series you've given Claude jobs, connected it to your files and saved jobs as Skills. It's doing real work now. And the more real work it does, the more it matters what happens at the edge, where something leaves your desk and lands in a client's phone or inbox.

What the handbrake is

It's simple. Anything that reaches a client goes through one gate.

Not three different checks in three different places. One gate, every time. Before a message goes out, the gate asks a few plain questions:

Is this a duplicate? Has the same message already gone to this person?
What time is it? Is this a reasonable hour to be contacting someone?
Has this been sent before? Is it repeating something that already went out?

If the answer to any of those is wrong, it doesn't go. Mine has stopped 8,760 texts that would have been duplicates.

That number isn't there to impress anyone. It's there to show how often a well-meaning system will try to do the same thing twice. Software repeats itself. A job runs again, a list gets loaded twice, a step gets retried. None of that is a problem until the thing being repeated is a text to a real person.

Why it matters for an agent

Most of what we send is personal. An appraisal confirmation. A price update to a vendor. A follow-up to someone who came through an open home. A market update to a database of past clients.

Get one of those wrong and it's awkward. Send it twice and it looks careless. Send it to the whole database at the wrong hour and you've spent years of goodwill in one evening.

Clients don't see your system. They only see what arrives. So the rule has to be about what arrives.

The good news is you don't need a complicated setup to start. You need one instruction, written down, that Claude follows every time.

Draft, never send

The rule written into my instructions comes down to this: never send, only draft.

Claude can write the message. It can pick who it should go to, check the details and lay it all out. What it doesn't do is press send on its own. It stops, shows me exactly what it's about to do, and waits.

That one change turns Claude from something that acts on your behalf into something that prepares and hands over. You stay the last set of eyes on anything with your name on it.

Over time you can build checks on top of that, like the duplicate and time-of-day checks above. But the foundation is the same. Nothing outward-facing happens without a yes.

The Day 16 prompt

Prompt

Before you send, post or change anything that a client or the public will see, stop and show me exactly what it is. Never send anything on your own.

Here's what each part is doing.

"Before you send, post or change anything" covers the three ways something can go wrong in public. Sending is texts and emails. Posting is social media and websites. Changing is the quiet one, like editing a listing or updating something a client can already see.

"that a client or the public will see" draws the line in the right place. Claude can still tidy your files, draft notes and sort your lists without stopping every two minutes. The handbrake only applies where it counts.

"stop and show me exactly what it is" means you see the real thing, not a summary. The actual words, the actual recipient, the actual post. If it's going to ten people, you see that it's ten people.

"Never send anything on your own" closes the door. Even if Claude thinks it has permission, even if it's sure, it waits for you.

Try it yourself

Pick one narrow folder. Use the folder for a single job, like drafting follow-ups after an open home. Copies of files, not originals, and not your whole computer.
Put the prompt where it will always apply. If you're working in a project, add it to the project's instructions so every conversation in that project starts with it. If you're not, paste it as the first line of the conversation.
Give Claude a job that ends in a message. For example: "Draft a follow-up text for each person in this list. Show me your plan first."
Read what comes back. Check the wording, the names and the number of messages. Look for anything doubled up.
Approve or correct. Say yes to the ones that are right, fix the ones that aren't, and send them yourself.
Check your settings. If you've connected Claude to anything that can send email, post or message on your behalf, look at how it asks for permission in the app's settings. Keep approvals on for anything that sends, spends or deletes.

Do this for a week and you'll start to notice the kinds of mistakes it would have made. That's useful. It tells you which checks to add to your own gate.

Building your own gate

Once the basic rule is working, you can ask Claude to add simple checks before it shows you anything. A few that make sense for most agents:

Duplicate check. "Before showing me any message, check whether the same person already has an identical message in this list."
Time check. "Flag anything that would go out early in the morning or late at night."
History check. "Compare against the list of messages already sent in this folder and tell me if any of these repeat."

Keep a simple log in the folder too, a running list of what was drafted, what was approved and what was stopped. It's an easy way to see what the gate is doing for you, and it gives Claude something to check against next time.

Start with the handbrake on. You can decide later which jobs, if any, have earned a little more freedom. Most outward-facing ones never should.

Where to go from here

The more Claude does for you, the more one rule matters. Draft, never send. Show me exactly what it is. Wait for my yes.

Add the Day 16 prompt to your instructions today, give Claude one job that ends in a message, and read what it shows you before anything goes anywhere.

Missed yesterday? Day 15: Skills: save a job once

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Claude skills for real estate agents: save a job once and run it by name

Matt Manser — Mon, 26 Oct 2026 09:00:00 +0000

If you're typing the same instructions twice, you're doing it wrong.

By now you've probably noticed a pattern. You give Claude a job, it does it well, and a few days later you need the same job done again. So you scroll back, find the instructions you wrote last time, copy them, paste them and tweak them. Or you type them out again from memory and forget half of it.

That's the problem skills fix.

A quick reminder on words. I call it SI, super intelligence. Most people say AI. When I say SI here I mean tools like Claude, made by Anthropic.

What a skill actually is

A skill is a job you explain once and then just name.

You write out the instructions properly, one time. Claude saves them as a skill with a name. From then on, you don't repeat the instructions. You say the name, give it what it needs for this particular job, and it follows the rules you already set.

Think of it like training a new assistant. The first week, you explain how you like things done. After that, you don't explain it every time. You just say "can you do the listing description for this one?" and they know what you mean.

Under the hood, a skill is a small set of saved instructions, sometimes with example files alongside. Claude knows which skills it has and pulls one in when you ask for it by name, or when the job clearly matches.

Why this matters for an agent's week

Most of the writing in a real estate business isn't one-off. It's the same handful of jobs over and over:

Listing descriptions
Appraisal follow-up emails
Open home recap notes
Price update messages to vendors
Social captions for a new listing

Every one of those has rules. How long it should be. What tone. What to always include. What to never say. Those rules live in your head, and every time you hand the job to Claude without them, the result drifts a bit.

A skill locks the rules in. The listing description you get on a Friday afternoon follows the same standard as the one you carefully briefed on a Monday morning. And you're not spending the first five minutes of every job re-explaining yourself.

It also makes the rules easier to improve. If something comes back wrong, you fix the skill once, and every future job gets the fix.

Start with a job you've already briefed well

Yesterday was about briefing Claude properly. A skill is that brief, saved.

So the best time to make a skill is straight after a job has gone well. You've just written clear instructions, Claude has followed them, and you've checked the result. That conversation already holds everything the skill needs.

For listing descriptions, your instructions might cover things like:

Length. A rough word count or number of paragraphs.
Structure. What goes first, what goes last.
Tone. How it should sound, and words to avoid.
Accuracy. Only use features in the notes. Never invent a room, a view or a distance.
Spelling. Australian spelling throughout.
What to send back. For example, a headline plus the description, ready to paste.

Those are your rules, not mine. Write the ones that match how you work and what your office requires.

The Day 15 prompt

Use this in the same conversation where you've already given Claude your listing description instructions.

Prompt

Turn the instructions I gave you for listing descriptions into a reusable skill called Listing Description, so next time I just say 'listing description' and paste my notes.

Here's what each part is doing.

"Turn the instructions I gave you for listing descriptions" points Claude at the brief that's already in the conversation. You don't have to write it all again. It gathers up what you said.

"into a reusable skill" tells it the output isn't a description. It's a saved set of instructions it can use again later.

"called Listing Description" gives it a clear name. Pick names you'll actually remember and say. Plain names beat clever ones.

"so next time I just say 'listing description' and paste my notes" describes how you'll use it. That shapes the skill: it knows to expect raw notes as the input and to turn them into a finished description using your rules.

Try it yourself

Check skills are turned on. In the Claude desktop app, look in Settings for the area that covers capabilities or skills. Depending on your plan and version, it may be on already. If you can't find it, ask Claude where skills live in your version of the app.
Brief one listing description properly. Start a conversation, write out your rules in full, and paste notes from a property. Use notes with no client personal details in them.
Check the result. Fix anything you don't like by adjusting the rules, not just the description. You want the instructions right before you save them.
Paste the Day 15 prompt. Add "show me the skill before you save it" on the end, so you can read the instructions it's about to keep.
Read what it shows you. Make sure every rule you care about is in there and nothing has been added that you didn't ask for.
Save it. Depending on your version, Claude will either save the skill for you, or give you a skill file to add through Settings. Follow what it tells you.
Test it in a fresh conversation. Start a new chat, type "listing description", paste a different set of notes, and see if the result follows your rules without any extra explaining.

If the test result is off, go back to the skill and tighten the instructions. That's normal. Most skills get better over the first few uses.

Keep the same safety habits

Skills make jobs faster, which is exactly why the habits from earlier in the series still matter.

One narrow folder. If a skill works with files, give it the folder for that job only. Not your whole computer.
Approvals on. Keep approvals on for anything that sends, spends or deletes. A skill that drafts a vendor email is fine. A skill that sends it should still wait for your yes.
"Show me your plan first." Especially the first few times you run a new skill. You'll see whether it's following your rules before it does the work.
Read before you use. A listing description still goes out under your name. Check it against the property before it goes anywhere.

What to save next

Once the first skill works, you'll start noticing other jobs you repeat. A good rule of thumb: if you've briefed the same job twice, it's ready to be a skill.

Build them one at a time. Brief the job well, check the result, then save it. A small set of skills you trust is worth more than a long list you've never tested.

Missed yesterday? Day 14: How to brief AI properly

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

How to brief AI properly: the 6-part brief every real estate agent should use with Claude

Matt Manser — Sun, 25 Oct 2026 09:00:00 +0000

Bad results from AI are almost always a bad brief.

When Claude hands back something bland, generic or just plain wrong, the easy conclusion is that the tool isn't very good. Usually that isn't what happened. It was given a vague job and it did a vague job. It filled every gap you left with the safest, most average guess it could make.

The fix isn't a cleverer tool or a magic phrase. It's a better brief.

Same quick note as always: I call it SI, super intelligence. Most people say AI. When I say SI here I mean tools like Claude, made by Anthropic.

Why the brief matters so much

Think about briefing a new assistant on their first day. If you said "write something about the new listing", you wouldn't be surprised when they came back with the wrong length, the wrong tone and none of the details you cared about. They'd have done their best with what you gave them. Claude is the same.

Everything you don't tell it, it has to guess. And its guesses are built to be safe and middle-of-the-road. That's exactly why so much AI writing sounds the same.

For an agent, this matters every day. Listing copy, appraisal follow-ups, vendor updates, social captions, emails to past clients, notes after an open. Each one has a person reading it, a tone that suits them, and things you'd never want said. If none of that is in the brief, none of it ends up in the result.

The six parts of a good brief

A good brief answers six questions. Leave any of them out and Claude has to guess.

Who you are. Your role and where you work. A real estate agent in your area writes differently from a generic copywriter, and Claude adjusts when it knows.
The job. What you actually want made, stated plainly. One job, not three tangled together.
Who it's for. The reader. A first home buyer, a downsizing vendor and a busy investor all want different things from the same information.
What good looks like. An example, or a clear description of the standard. This is the part most people skip, and it does the most work.
What to avoid. The words, claims and habits you don't want. For agents this includes anything you can't back up.
The format. How you want it handed back. Length, headings, dot points, a ready-to-send email, a table.

None of these are complicated. The difference is that you've written them down instead of hoping Claude reads your mind.

Same job, two briefs

Take one everyday job and brief it two ways.

The lazy brief

Something like: "Write a listing description for a 3 bedroom house."

What comes back is usually what you'd expect. It's the right sort of shape, but it's generic. It reaches for the same tired real estate words everyone has seen a hundred times. It may invent features to fill the space, because you didn't give it any. It doesn't know the suburb, the buyer or the length you need. You end up rewriting most of it, which is the job you were trying to hand over in the first place.

The 6-part brief

Now the same job, briefed properly:

I'm a real estate agent selling a family home in [suburb].
The job is a listing description for the online portals, using only the details in the attached property notes.
It's for young families who want space and are close to their budget limit.
Good looks like [paste a past description you were happy with], warm, specific and easy to read.
Avoid clichés, exaggeration and any feature or claim that isn't in the notes.
Give it to me as a headline plus three short paragraphs, under 200 words.

Same Claude, same job, a completely different result. It writes for a real reader, in the style you've shown it, with the details you gave it and nothing made up. You're editing a draft that's close, not starting again.

The Day 14 prompt

Here's the brief as a fill-in-the-blanks prompt. Keep it somewhere handy and fill in the brackets each time.

Prompt

I'm [who you are]. The job is [task]. It's for [audience]. Good looks like [example]. Avoid [things]. Give it to me as [format].

Here's what each part does.

"I'm [who you are]" sets the point of view. Claude writes as someone in your role and your market, not as a stranger.

"The job is [task]" gives it one clear thing to do. If you have three jobs, write three briefs.

"It's for [audience]" tells it who will read the result, which shapes the words, the detail and the tone.

"Good looks like [example]" shows it the standard. A real example of your own work is the strongest thing you can give it, because it doesn't have to guess your style.

"Avoid [things]" closes off the habits you don't want. Clichés, overstatement, jargon, anything you can't stand behind.

"Give it to me as [format]" means it comes back in a shape you can use straight away, not one you have to rearrange.

Try it yourself

Pick one job you do every week and give it the full brief.

Choose a real job. A listing description, a vendor update, a follow-up email, a social caption. Something you'd normally write yourself.
Send the lazy version first. One line, no detail. Save what comes back.
Start a new conversation and fill in the six parts. Take your time with "good looks like". Paste in an example of your own that you were happy with.
If the job needs your files, add one narrow folder. Only the files for that one job, and copies rather than originals.
Add "show me your plan first" to the end of the brief, so Claude tells you how it will approach the job before it starts.
Compare the two results side by side. Note what the better brief fixed, and what you'd still change.
Tighten the brief and save it. Whatever you had to correct, add to "avoid" or "good looks like" next time. A brief you reuse gets better every time you use it.

Keep the safety habits

A better brief gets you a better result, but it doesn't replace checking. A few habits still apply.

One narrow folder. If the job needs files, give Claude the folder for that job only, never your whole computer. If it isn't in the folder, Claude isn't reading it.
Approvals on. Keep approvals on for anything that sends, spends or deletes, so nothing goes out or disappears without your yes.
Read before it's public. Anything going to a client or onto a portal gets read by you first. Check every fact against your own notes. You're the one responsible for what's said about a property.

Where to go from here

When a result disappoints you, look at the brief before you blame the tool. Ask which of the six parts you left out. Most of the time the answer is right there.

Start with one job this week. Brief it properly, save the brief, and reuse it. Over time you'll build a small set of briefs for the jobs you do most, and the results will be closer to right on the first go.

Missed yesterday? Day 13: One listing, a week of posts

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

One listing, a week of posts: how real estate agents can use Claude to plan seven days of Instagram content

Matt Manser — Sat, 24 Oct 2026 09:00:00 +0000

One listing should be a week of content, not one post.

Most agents put up a "just listed" post, maybe a carousel of photos, and move on. Then the listing sits there for the rest of the campaign while the feed goes quiet or fills up with something unrelated.

But a listing has a lot more in it than one post can hold. There's the home itself, the best feature, the street it's on, the suburb around it, the open home, the buyers you're hoping to reach, and eventually the result. That's seven different angles, and each one is worth its own day.

Today's job is handing that planning to Claude. Give it the listing and it plans seven days: just listed, a feature, the street, the suburb, open home, a buyer tip, and the result.

Why this matters for an agent's week

Posting about a listing isn't hard. Deciding what to post, every day, in the middle of everything else, is the hard part.

That's the bit that usually gets skipped. Not because agents don't know social media helps, but because sitting down to think up a fresh angle each morning loses out to phone calls, appraisals and paperwork.

A plan fixes that. If you've got seven posts mapped out on the day the listing goes live, each one with a hook and a caption ready, the daily decision is gone. You just check it, add the right photo and post.

It also helps the listing. A week of posts with different angles gives buyers more reasons to stop scrolling than one post repeated. Someone who skips "just listed" might stop for a post about the street, or a buyer tip that happens to answer the question they had.

What Claude actually does here

This is the same shift from Day 1: you're not asking Claude for social media tips, you're giving it a job.

You paste in your listing details. Claude reads them, picks seven different angles, and writes one post for each day. Each post gets a hook as its first line and a short caption underneath.

What you get back is a plan you can read in a couple of minutes, then use straight away or tweak.

What it doesn't do is post anything. That stays with you. Claude drafts, you check, you post.

The Day 13 prompt

Prompt

Here's my new listing: [paste details]. Plan 7 days of Instagram posts about it, one per day, each with a different angle, a caption under 80 words and a hook as the first line. Don't invent features.

Here's what each part is doing.

"Here's my new listing: [paste details]." This is the raw material. Paste in your listing copy, the property details, the key features, and anything else you'd normally put in the ad. The more real detail you give it, the better the posts. Leave out anything private, like the vendor's personal details or their reasons for selling.

"Plan 7 days of Instagram posts about it, one per day." This sets the shape of the job. Seven separate posts, not one long post or a vague list of ideas.

"Each with a different angle." This is what stops you getting the same post seven times with the words shuffled. It pushes Claude to look at the listing from different directions: the home, a feature, the street, the suburb, the open home, a buyer's point of view, the result.

"A caption under 80 words." Short captions are easier to read on a phone and easier for you to check. Giving it a limit also stops it padding.

"A hook as the first line." On Instagram, the first line is what people see before they decide whether to read on. Asking for a hook makes Claude lead with the most interesting thing, not the address.

"Don't invent features." This is the most important line in the prompt. If you don't say it, Claude may fill gaps with things that sound right, like a renovated kitchen or a north-facing yard, when your listing never said so. In real estate that's not a small slip. Advertising a feature a property doesn't have is a real problem. This line tells it to stick to what you gave it.

Even with that line in, read every post against the listing before it goes up. You're the one responsible for what gets published.

Try it yourself

Pick one current listing. Gather the listing copy and property details you'd normally use in the ad.
Open a new conversation in Claude. You don't need a folder for this one, since you're pasting the details in. If you'd rather work from a file, add one narrow folder that holds only that listing's copy, nothing else.
Paste the prompt and replace [paste details] with your listing information.
Read the plan. Check every post against the listing. Is each feature real? Is anything overstated? Does each day have a genuinely different angle?
Ask for changes. If a caption is too salesy, too long or not how you'd say it, tell Claude. Something like "make day 3 sound more relaxed" or "rewrite day 5 for the Saturday open home at 11am" works fine.
Save the plan. Copy it into your notes, or if you're working in a folder, ask Claude to save it as a file there for you to open.
Post it yourself, one day at a time. Add the right photo for each post, and check it once more before it goes live.

A note on the last day. The result post only makes sense once there is a result, so treat day seven as a draft you come back to. Before posting anything about a sale, check with your vendor and your office about what you can share.

Keep the habits tight

The same safety habits from Day 1 apply here, and one of them matters more than usual.

One narrow folder. If you use a folder, make it one listing, not your whole listings drive. Claude only needs that property's details.
Approvals on for anything that sends, spends or deletes. Publishing is sending. If you ever connect Claude to anything that can post for you, keep approvals on so nothing goes live without your yes.
"Show me your plan first." If you'd like to see the seven angles before Claude writes all the captions, add that line to the end of the prompt. You can swap an angle out before any writing is done.

Where to go from here

Once you've done this for one listing, you've got a pattern you can reuse for every listing after it. Same prompt, new details, a week of posts planned in one go.

You can also adjust it to suit you. Change Instagram to Facebook, ask for a shorter caption, or give it the tone you like to use. The parts that matter stay the same: real details in, different angles, a clear hook, and nothing invented.

Missed yesterday? Day 12: Your market update in 5 minutes

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Your market update in 5 minutes: how real estate agents can use Claude to find sourced suburb numbers and write the script

Matt Manser — Fri, 23 Oct 2026 09:00:00 +0000

I used to dread my market update video.

The video wasn't the hard part. The hard part was everything before it: finding current numbers for each suburb, working out which ones were worth mentioning, and turning a page of figures into something a homeowner would watch for more than five seconds.

Today's job hands that whole front end to Claude. You ask it for the latest public numbers for your suburbs, have it turn them into a short script, and make it show you where every number came from.

That last part matters most, and it's what this post is really about.

Why a market update is a good job to hand over

A market update is one of the most useful things an agent can put out. Homeowners want to know what's happening on their street, and a short, regular update keeps you in front of them between the times they're thinking about selling.

It's also the kind of job that slips. The research is fiddly, the numbers move, and by the time you've pulled it together there's no energy left for the part that counts, which is getting in front of the camera.

Claude is good at the fiddly middle. It can look up published figures, line them up, and draft a script in plain English. What it can't do is make a number true. That's why the prompt is built the way it is.

The rule: no source, no number

When you ask Claude for market figures, there are two ways it can go wrong.

It can't look anything up. If web search isn't available in your conversation, it's working from what it learned in training. Those figures will be out of date, and it may not say so clearly.
It fills a gap. Ask for a number it can't find and it may give you something that sounds right rather than telling you it came up empty.

Neither is acceptable in something you're going to say to homeowners under your name. So you don't just ask for numbers. You ask for the source and the date beside every one, and you tell it plainly not to use anything it can't source.

That changes the job. Instead of "give me some figures", it becomes "find published figures and prove it". You get a list you can check, not a list you have to trust.

Different sources, different numbers

One thing worth knowing before you start: different data providers publish different figures for the same suburb. They use different sales, different time periods and different methods, and some report monthly while others report quarterly.

None of that means one is wrong. It means you should know which one you're quoting, and stick with the same source from one update to the next so your numbers compare properly. Asking Claude for the source and date of each figure makes that easy to see.

The Day 12 prompt

Prompt

Find the latest published median house price, change over 12 months and days on market for [suburb], with the source and date for each. Then write me a 40-second market update script for local homeowners. Don't use any number you can't source.

Here's what each part is doing.

"Find the latest published median house price, change over 12 months and days on market for [suburb]" sets exactly three numbers to look for. Three is enough for a short video and few enough to check. "Published" points it at figures someone has actually released, not its own estimates. Swap in your suburb name.

"with the source and date for each" is the part that makes this safe to use. Every figure comes with where it came from and when, so you can open the source and confirm it yourself.

"Then write me a 40-second market update script for local homeowners" turns the research into something you can record. Forty seconds keeps it short enough for social video. "Local homeowners" tells it who's listening, so it writes for people living in the suburb, not for other agents.

"Don't use any number you can't source" closes the gap. If it can't find a figure, the honest answer is to say so and leave it out. This line gives it permission to do exactly that.

Try it yourself

Check Claude can search the web. In the Claude app, web search is usually a setting or tool you can switch on for a conversation. If it's off, Claude can't find current figures, so turn it on before you start.
Start a new conversation and paste the prompt. Replace [suburb] with one suburb you work in. Start with one, not your whole patch.
Read the numbers before the script. Look at each figure and its source and date. Click through to the source and check the number is really there and says what Claude says it does.
Check the dates. "Latest" can mean last month or last quarter depending on the source. If anything looks old, ask Claude to look for a more recent release.
Read the script out loud. Time it. If it runs long, ask for a tighter version. If it doesn't sound like you, tell Claude how you'd normally say it and ask for another pass.
Ask for an on-screen source line. Something short like the provider name and month, to show in the video or put in the caption. Homeowners trust a number more when they can see where it came from. Check your office's guidelines on how market figures should be quoted, as rules on this vary.
Record. You've skipped the part that used to stop you, so this is the bit to put your energy into.

Once one suburb works, run the same prompt for your others. If you want to keep the results, give Claude one folder for your market updates and ask it to save each suburb's figures, sources and script as a file there. Same habit as always: one narrow folder for the job, not your whole computer, and ask it to show you its plan first before it creates or changes any files.

Keep the checking step

It's tempting, once the first few come back clean, to stop clicking through to the sources. Don't.

The point of the source-and-date rule isn't that Claude will always get it wrong. It's that you're the one saying these numbers to people who might list with you. A wrong figure in a market update costs more trust than the video earns. The check takes a minute, and it's the minute that lets you post with confidence.

The same goes for anything that leaves your hands. If you later connect Claude to tools that can post, send or schedule, keep approvals on for anything that sends, spends or deletes. Today's job doesn't need any of that. It finds, drafts and hands back to you. You decide what goes out.

Where to go from here

The market update is a good example of the pattern running through this week. Claude does the slow research and the first draft. You check the facts and bring the voice. It's not about the video making itself. It's about getting from blank page to ready-to-record without the part you used to dread.

Start with one suburb and the Day 12 prompt. Check every number against its source. Then record.

Missed yesterday? Day 11: The Friday vendor update

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

The Friday vendor update: how real estate agents can use Claude to draft it every week

Matt Manser — Thu, 22 Oct 2026 09:00:00 +0000

Vendors don't leave because the house didn't sell. They leave because nobody told them anything.

That's the uncomfortable truth about vendor relationships. A slow campaign can be explained. A quiet agent can't. When a vendor hasn't heard from you in a week, they fill the gap themselves, and what they fill it with is rarely kind to you.

Today is about closing that gap with one scheduled job. Every Friday, Claude reads the week's notes for a listing and drafts the vendor update. You check it and send it.

A quick reminder on words. I call it SI, super intelligence. Most people say AI. When I say SI I mean tools like Claude, made by Anthropic.

Why Friday, and why every week

The vendor update is one of those jobs everyone agrees matters and nobody enjoys. It isn't hard. It just arrives at the end of the week, when you're tired, and it means digging back through inspection notes, buyer comments and call logs to piece together what actually happened.

So it slips. Or it gets written in a rush and says very little.

The fix isn't trying harder. It's making the first draft turn up on its own. If the draft is already sitting there on Friday afternoon, the job shrinks from "write the update" to "read the update and press send". That's a much easier thing to keep doing every week.

Friday works because the week is done, the weekend open homes are coming, and a vendor who hears from you before the weekend goes into it knowing where things stand.

What the job actually is

This builds on the idea from Day 1: don't ask Claude about the job, give it the job.

Here the job is:

Read this week's notes for one property.
Pull out the inspection numbers, what buyers said and how strong the interest is.
Add your recommendation for next week.
Write it as a short, honest, warm update.
Save it as a draft, not send it.

The only new part is the timing. Instead of you remembering to ask on Friday, the task is set to run at the same time each week.

Where the notes come from

Claude can only report on what's in your notes. If the week's buyer feedback lives in your head, the draft will be thin.

That's not a reason to skip this. It's a good reason to keep a simple notes file per listing. After each open or private inspection, add a few lines: how many groups came through, any comments worth passing on, anyone who asked for a contract or a second look. Rough is fine. Claude does the tidying.

Keep those notes in one folder for that property. That folder is what you'll give Claude.

The Day 11 prompt

Prompt

Every Friday at 2pm: read this week's notes for [address] and draft my vendor update: inspections, what buyers said, interest level, and my recommendation for next week. Honest, warm, under 200 words. Save as a draft.

Here's what each part is doing.

"Every Friday at 2pm" sets the rhythm. It's early enough that you can read and send before the end of the day.
"read this week's notes for [address]" tells it exactly what to look at. One property, this week only. Swap in the real address.
"inspections, what buyers said, interest level, and my recommendation for next week" is the structure. Vendors want the same four things every time: how many came, what they thought, how keen they are, and what happens next.
"Honest, warm" sets the tone. Honest matters most. A vendor update that only reports good news isn't an update, and vendors can tell.
"under 200 words" keeps it readable on a phone. Vendors read these between other things.
"Save as a draft" is the safety line. Claude prepares it. You send it.

That last part is the one not to change. The update goes out in your name, to a client who is trusting you with their biggest asset. You read every one before it leaves.

Try it yourself

Make a folder for one listing. Put that property's notes file in it, plus anything else from the week you'd normally refer to. Copies are fine.
Open the Claude desktop app and add that folder. Just that one folder, the same way you did on Day 1.
Run it once by hand first. Paste the prompt without the "Every Friday at 2pm" part, with the real address filled in, and add "show me your plan first" at the end.
Read the draft closely. Check the inspection numbers match your notes. Check the buyer comments are accurate and fairly put. Check the recommendation is actually what you'd say.
Adjust the prompt if needed. If it's too formal, say so. If you always sign off a certain way, add that. Get the one-off version right before you schedule it.
Set it to run weekly. Use Claude's scheduling feature to run the full prompt every Friday at 2pm. Menu names vary between versions, so look for scheduled or recurring tasks in the app. Check how it works on your setup, as scheduled tasks generally need your computer on and the app running at that time.
On Friday, read and send. Open the draft, make any changes, and send it yourself.

Start with one listing. Once you're happy with the drafts it produces, set up the same task for your other listings, each with its own folder.

Keep it tight

The same habits from earlier in the series apply here, and they matter more when a task runs on its own.

One narrow folder per listing. Vendor notes contain personal information about your clients and about buyers. Only the notes for that property should be in that folder. If it isn't in the folder, Claude isn't reading it.
Approvals on for anything that sends, spends or deletes. If you've connected Claude to your email, keep it set so nothing goes out without you saying yes. This task saves a draft. It shouldn't be sending anything.
"Show me your plan first" while you're setting it up. Once the drafts are consistently right, you can drop it from the scheduled version.
Check the numbers every time. A wrong inspection count in a vendor update does more damage than a late one.

The point of it

Vendors don't need a long report. They need to hear from you, regularly, with the truth and a plan. That's what keeps them with you through a slow week.

A scheduled draft doesn't replace the phone call when the news is hard. It makes sure the weekly update never quietly drops off your list in the first place.

Missed yesterday? Day 10: Every enquiry gets a reply draft

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Every enquiry gets a reply draft: how real estate agents can use Claude to answer buyers and sellers faster

Matt Manser — Wed, 21 Oct 2026 09:00:00 +0000

A buyer who waits a day for a reply has already called someone else.

Every agent knows this. And every agent has had the week where enquiries pile up anyway. You're at an appraisal, then an open, then a contract signing, and the inbox keeps filling. By the time you sit down at night, some of those messages are two or three days old.

Today's step is about closing that gap. Instead of writing every reply from scratch, you have Claude find the enquiries you haven't answered and draft a reply to each one. It drafts. You read it and press send.

Why this one matters

Most enquiries aren't hard to answer. Someone wants to know if the property is still available, when the next open is, whether there's a price guide, or what you'd think their place is worth. The answer usually takes a minute to work out. Writing it well, in a tone that sounds like you, takes longer. Multiply that by a full inbox and it's the job that keeps getting pushed to tomorrow.

That's what makes it a good job to hand over. The thinking is light, the writing is repetitive, and the cost of leaving it is real.

The goal isn't to take you out of it. It's to change your part from writing every reply to checking every reply. Thirty seconds each instead of thirty minutes.

Drafts, not sends

This is the most important line in today's post. Claude writes the drafts. It doesn't send them.

Email is outward-facing. Once something goes to a buyer or a seller, you can't take it back. So for this job, Claude's role ends at the draft. You're still the one who reads each reply, fixes anything that's off, and decides when it goes.

That's also why the prompt ends the way it does. More on that below.

What you need set up first

For Claude to read your inbox, it needs a connection to your email. In the Claude apps this is done through connectors, which you'll find in Settings. The episode shows Gmail, and the setup is similar for other email services where a connector is available. You sign in to your email account, give Claude permission, and from then on it can search and read your mail when you ask it to.

A couple of things to check when you connect:

What it can do. Look at what the connector is allowed to do. Reading and searching is what this job needs. If your setup can also save drafts, Claude can put them straight into your drafts folder. If it can't, it will show you the drafts in the conversation and you copy them across.
Approvals. Keep approvals on for anything that sends, spends or deletes. If Claude ever tries to do one of those things, you'll be asked first. For this job, the answer is no.

The Day 10 prompt

Prompt

Find every buyer or seller enquiry in my inbox from the last 3 days that I haven't replied to. Draft a reply to each one in my voice, answering what they actually asked. Save them as drafts. Do not send anything.

Here's what each part is doing.

"Find every buyer or seller enquiry" tells Claude what kind of email to look for. Not newsletters, not supplier invoices, not the office group chat. Just people asking you about a property or about selling.

"From the last 3 days" keeps the job narrow. It's the inbox version of adding one folder instead of your whole computer. A short window means fewer emails to work through and fewer chances of digging up old conversations you've already dealt with another way.

"That I haven't replied to" is the filter that makes this useful. Claude checks whether you've already answered, so you're not doubling up on replies.

"In my voice" asks it to write the way you write. It does this best when it has something to go on, like your sent emails or a sample of how you normally reply. More on that in the steps below.

"Answering what they actually asked" stops it writing a generic "thanks for your enquiry" email. If someone asked about parking, the reply should mention parking. If they asked about price, the reply should deal with price the way you normally would.

"Save them as drafts. Do not send anything." draws the line. Claude prepares, you decide.

Try it yourself

Connect your email. In the Claude app, go to Settings and find connectors. Add your email account and sign in. Check what it's allowed to do while you're there.
Give it a sample of your voice (optional, but worth it). Make a new folder and put in a short document with three or four replies you've sent before that you're happy with. Remove client names and details first. Add just that one folder to the conversation. This gives Claude a clear model to copy.
Start with a plan. Paste the Day 10 prompt, and add "show me your plan first" on the end. Claude will tell you how it's going to find the enquiries and what it's going to do with them before it does anything.
Check the list. Ask it to list the enquiries it found before drafting. Make sure they're the right ones, and that it hasn't missed any or included something that isn't an enquiry.
Let it draft. Once the list looks right, tell it to go ahead. If it asks for permission to save drafts, approve that. If it ever asks for permission to send, say no.
Read every draft. Open your drafts folder (or read them in the conversation). Check names, addresses, times and anything factual. Claude only knows what's in the email thread and whatever you've given it, so if a reply needs a price guide or an open time it can't see, fill that in yourself.
Send the ones you're happy with. Edit anything that doesn't sound like you. Then send.

The first time through, expect to change a few things. Tell Claude what you changed and why, so it can do better next time. "Shorter." "Less formal." "Always offer a call back." Small notes like that add up quickly.

A few habits to keep

Keep the window short. Three days is plenty. If you've let it go longer, run it on the most recent days first.
Never let it send. Even once you trust the drafts, the send button stays yours. It's the last check before something reaches a client.
Watch for personal information. Enquiries come with names, phone numbers and sometimes details about someone's circumstances. Only connect what you're comfortable with, and follow whatever rules apply where you work.
Check your connections now and then. Look in Settings to make sure the only accounts and folders connected are the ones you meant to share.

Where to go from here

The enquiries were always going to need a reply. The only question was how long people waited for it. Having the drafts ready means the slow part is done before you sit down, and all that's left is reading, tweaking and sending.

Run the prompt once today. Read what comes back. Send the ones that are right.

Missed yesterday? Day 9: The 6am morning debrief

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

The 6am morning debrief: how to set up a daily Claude summary of your emails, leads and appointments

Matt Manser — Tue, 20 Oct 2026 09:00:00 +0000

Every morning at 6am I get one message that tells me what matters today.

It isn't a newsletter or a dashboard I have to log in to. It's a short summary that Claude writes for me before the day starts: what's urgent, who needs a reply, which leads are waiting, and what's in the calendar.

I set it up once. It runs every weekday on its own.

A quick reminder on words. I call it SI, super intelligence. Most people say AI. When I say SI in this series I mean tools like Claude, made by Anthropic.

What the morning debrief actually is

Yesterday was about connecting your email and calendar to Claude. Once that's done, Claude can read them when you ask. Today is about not having to ask.

A scheduled task is a job you give Claude once, with a time attached. You write the instructions, pick when it should run, and Claude does the job at that time without you opening a chat and typing anything.

The morning debrief is the simplest useful scheduled task there is. Claude looks through the last 24 hours of email and today's calendar, sorts what it finds into a few groups, and gives you a short summary.

Why it matters for an agent

Most agents start the day by opening their inbox and scrolling. Somewhere in there is a buyer asking about a contract, a vendor wanting an update, a portal enquiry from overnight, and a dozen newsletters and notifications burying all of them.

The real cost isn't the reading. It's the sorting. Working out what's urgent, what can wait, and what's a lead that will go cold if nobody answers it this morning.

That sorting is a job, not a question, which is exactly the shift from Day 1. And it's a job that needs doing every single weekday, which makes it a good fit for a schedule.

A debrief also helps with the calendar side. An appraisal at 10am is easier to walk into when you've been reminded the night before's emails included a question from that same vendor.

The Day 9 prompt

This is the instruction you give the scheduled task. Paste it in as it is to start, then adjust it once you've seen a few mornings of results.

Prompt

Every weekday at 6am: review my emails from the last 24 hours and my calendar for today. Give me 1) urgent, 2) needs a reply, 3) leads, 4) today's appointments with what to prep. Keep it under 200 words.

Here's what each part does.

"Every weekday at 6am" sets the timing. Weekdays only, so it isn't writing you summaries on a Sunday unless you want it to. When you create the scheduled task you'll also set the time in the schedule settings, so make sure the two match.
"Review my emails from the last 24 hours and my calendar for today" tells it what to look at, and just as importantly, what not to. It isn't trawling your whole mailbox history, just the window that matters this morning.
"1) urgent" goes first, because it's the thing you most need to see before your coffee.
"2) needs a reply" catches the emails that aren't urgent but will become a problem if they sit for three days.
"3) leads" pulls enquiries out into their own list, so a new buyer or a possible appraisal doesn't get lost among the admin.
"4) today's appointments with what to prep" turns your calendar into a to-do list. Not just "11am, appraisal", but what you might need for it.
"Keep it under 200 words" is the part people skip and shouldn't. A debrief you can read in a minute gets read. A two-page report doesn't.

Notice what the prompt doesn't ask for. It doesn't ask Claude to reply to anything, file anything, or delete anything. It reads and reports. That's deliberate.

Try it yourself

Do this when you've got ten quiet minutes, not first thing on a busy morning.

Check your connections. Your email and calendar need to be connected to Claude first. If you haven't done that yet, start with Day 8.
Run it once by hand. Open a normal chat in the Claude desktop app and paste the prompt, but start it with "Show me your plan first, then do this once now:". Read the plan, let it run, and check the summary against your actual inbox. Did it miss anything urgent? Did it call a newsletter a lead?
Adjust the wording. If it got something wrong, tell it. For example, "anything from a portal enquiry address counts as a lead" or "ignore emails from these senders". Add those lines to the prompt.
Create the scheduled task. In the Claude desktop app, find the option for scheduled tasks and create a new one. Paste in your adjusted prompt, then set it to run on weekdays at 6am. The exact menu wording can change between app updates, so if you can't spot it, ask Claude in the app where scheduled tasks are.
Check how it runs. Look at whether your computer and the Claude app need to be on and awake at 6am for the task to run. If your laptop is closed in a bag at that hour, pick a time when it won't be, or leave it on.
Read the first week closely. For the first few mornings, compare the debrief with your inbox. Once it's consistently catching what matters, you can start trusting it as your first read of the day.

Keep it read-only

The morning debrief runs while you're asleep. That's the point, but it also means nobody is watching it run. So the safety habits from Day 1 matter more here, not less.

Read, don't act. Keep the scheduled version to reading and summarising. If you want Claude to draft replies, have it draft them for you to check and send yourself, rather than sending anything on its own.
Approvals on for anything that sends, spends or deletes. Leave those approvals switched on. A summary that runs at 6am shouldn't be able to send an email at 6am without you saying yes.
Only connect what the job needs. For this task, that's your email and your calendar. It doesn't need access to every folder on your computer. If you want it to save each debrief as a file, give it one narrow folder for that and nothing else.
Mind the client details. Your inbox is full of personal information. Make sure the way you're using Claude fits the privacy rules where you work, and keep the summary somewhere only you can see it.
Show me your plan first. Any time you change the instructions, run it once by hand with "show me your plan first" before you let the schedule take over again.

You can give it more to do later, once you've seen it behave. Start with read-only.

Where to go from here

The morning debrief is a small thing. One short message, five days a week. But it's the first job in this series that happens without you starting it, and that's a different way of working with Claude.

Once you're comfortable with one scheduled task, you'll start noticing other jobs that happen on a rhythm: a weekly check of listings, a Friday wrap-up, a reminder list before open homes. Each one works the same way. Write clear instructions, test it by hand, then put it on a schedule.

Missed yesterday? Day 8: Connect your email and calendar

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Connect your email and calendar: the day Claude stops being a chatbot for real estate agents

Matt Manser — Mon, 19 Oct 2026 09:00:00 +0000

This is the day Claude stopped being a chatbot for me.

Up to now in this series, Claude has only known what you've shown it. A folder of files, a pasted document, a photo of a receipt. That's useful, but it means you're always the one carrying information to it.

Connecting your email and calendar changes that. Claude can look at what's actually happening in your business, your appointments and the messages coming in, and tell you what needs your attention.

A quick reminder on words. I call it SI, super intelligence. Most people say AI. When I say SI I mean tools like Claude, made by Anthropic.

What connecting actually means

Claude has a feature called connectors. A connector links Claude to another service you already use, so it can read from it when you ask.

Today we're adding two: Gmail and Google Calendar.

Once they're connected, you don't need to copy emails into the chat or type out your week. You ask, and Claude looks for itself. It can search your inbox, read threads, check what's in your calendar, and pull it together into something you can act on.

Here's the part that matters most. It reads, and it drafts, but it never sends anything without you. You stay the one who decides what goes out.

Why this matters for an agent's week

Our week runs on two things: appointments and messages.

Appraisals, open inspections, buyer follow-ups, vendor updates, conveyancer emails, settlement dates. Most of it lives in the inbox and the calendar, and the hard part usually isn't doing any single task. It's keeping track of everything at once.

That's what this step helps with. Instead of scrolling back through two days of email wondering what you've missed, you can ask Claude to look and tell you. Instead of opening tomorrow's appointment and trying to remember what it's about, Claude can find the related emails and remind you.

It's the difference between an assistant who waits for you to hand them things and one who can check the diary and the inbox themselves.

How to connect Gmail and Google Calendar

Menu names shift a little between versions, but the steps look roughly like this.

Open Claude and go to Settings. You'll find it from your name or profile icon.
Find Connectors. This is where Claude lists the services it can link to.
Choose Gmail and connect it. You'll be sent to a Google sign-in page. Sign in with the account you use for work and read what Claude is asking permission to do before you allow it.
Do the same for Google Calendar.
Check they're switched on in your chat. In a new conversation, look in the tools or connectors menu near the message box and make sure both are enabled.

If your work email isn't Gmail, have a look through the Connectors list for your provider. What's available can depend on your plan and setup.

The Day 8 prompt

This is the prompt for today. Like Day 1, it doesn't change anything. It only reads and reports back.

Prompt

Look at my calendar for this week and my emails from the last 2 days. What do I need to prepare for, and which emails still need a reply?

Here's what each part is doing.

"Look at my calendar for this week" tells Claude where to start and keeps it to a sensible window. Not your whole year, just the days you're actually about to work.

"and my emails from the last 2 days" puts a limit on the inbox too. Two days is recent enough to be relevant and small enough to read properly. You can widen it later.

"What do I need to prepare for" asks Claude to connect the two. It looks at what's coming up and tells you what each appointment needs from you, often using details it found in your emails.

"and which emails still need a reply?" turns a scroll through your inbox into a short list. Claude checks which threads are still waiting on you, so you can see what's outstanding at a glance.

Notice the prompt is a job, not a question about how to manage email. It says what to look at and what to give back, which is the habit we've been building since Day 1.

Try it yourself

Connect Gmail and Google Calendar using the steps above.
Start a new conversation and check both connectors are switched on.
Paste the Day 8 prompt and send it.
Read what comes back. Check it against what you know. Did it catch the appointments you expected? Are the emails it flagged really the ones waiting on you?
Pick one email from its list and ask Claude to draft a reply. Read the draft, change what you'd say differently, and send it yourself if you're happy with it.
Pick one appointment and ask Claude to pull together everything it can find about it from your recent emails, so you walk in prepared.

Do this at the start of a working day and you'll get a feel for how useful it is quickly.

Keep the reins tight

Your inbox holds a lot of personal information about clients, so the same safety habits from earlier in the series apply, and they matter more here.

Connect only what you need. The same idea as giving Claude one narrow folder instead of your whole computer. Connect the work accounts this job needs, and nothing else. You can disconnect anything from Connectors at any time.
Keep approvals on for anything that sends, spends or deletes. Reading and drafting is fine. Anything that leaves your account or removes something should wait for your yes.
Ask for the plan first. For anything bigger than a summary, end your request with "show me your plan first" so you see what Claude intends to do before it does it.
Know your office's rules. How client information can be handled depends on where you work. Check before you connect a work account.

It's worth going back into Connectors now and then to check the list is only what you meant to share.

Where to go from here

Once your email and calendar are connected, Claude isn't working blind any more. It can see your week and your inbox, and it can help you stay on top of both without you carrying everything to it.

Start small. Run today's prompt once a day for a week, check its answers against what you know, and get used to reviewing its drafts. Trust builds from there.

Missed yesterday? Day 7: Receipts to a spreadsheet

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Receipts to a spreadsheet: how real estate agents can use Claude to sort tax time

Matt Manser — Sun, 18 Oct 2026 09:00:00 +0000

Tax time used to cost me a weekend.

Not because any one receipt is hard. A receipt is a date, a supplier, an amount and some GST. The problem is there are a lot of them, they come in every shape, and someone has to read each one and type it into a spreadsheet. That someone was me.

It doesn't have to be you. Put your receipts in a folder, ask Claude for a spreadsheet, and it reads every one. The ones it isn't sure about go on a separate tab, so you only check those.

As always in this series, when I say SI (super intelligence) I mean tools like Claude, made by Anthropic. Most people say AI. Same thing.

Why this one matters for agents

Agents spend money all week in small amounts. Fuel, parking, signboards, printing, coffee with a client, a subscription that renewed without you noticing. Each one is small, but they pile up. Some are in your email, some are photos on your phone, some are crumpled in the car door.

None of it is difficult. It's just slow, and it always lands at the worst time, when you'd rather be prospecting or having a weekend.

This is exactly the kind of work from Day 1: a job, not a question. You don't need Claude to explain how to do a receipts spreadsheet. You need it to do the reading and the typing, and hand you back something you can check.

What Claude actually does with your receipts

Claude can read photos and PDFs, not just typed documents. So a photo of a paper receipt, a screenshot of an online order and a PDF invoice can all go in the same folder.

When you give it the job, it works through each file, pulls out the details you asked for and lays them out in rows and columns. You end up with one spreadsheet instead of a folder full of images.

The part that makes this trustworthy is the second tab. Some receipts will be faded, cut off, blurry or just odd. A good assistant wouldn't guess quietly on those. They'd put them aside and tell you why. That's what the prompt asks Claude to do.

So instead of checking every line, you check the short list it flagged, plus a quick look over the rest.

The Day 7 prompt

Prompt

Read every receipt in this folder. Make a spreadsheet with date, supplier, amount, GST and category. Put anything you're not sure about on a second tab called Check these, and tell me why.

Here's what each part is doing.

"Read every receipt in this folder." Tells it what to look at, and that you want all of them, not a sample. The folder is the boundary, so it isn't reading anything else.
"Make a spreadsheet with date, supplier, amount, GST and category." Tells it what you want back and exactly which columns. Naming the columns means you get the same layout every time, which makes it easier to hand to your bookkeeper or accountant.
"Put anything you're not sure about on a second tab called Check these." This is the safety net. Uncertain receipts don't get mixed in with the confident ones. They're kept separate so nothing slips through.
"And tell me why." A reason next to each flagged receipt saves you working out what's wrong. "Total is cut off" or "can't tell if GST was charged" points you straight at the problem.

If you already use set categories, add them to the prompt, for example "use these categories: fuel, marketing, phone, office, other". Otherwise Claude will choose sensible ones, and you can adjust from there.

Try it yourself

Make a new folder. Copy in a batch of receipts. Photos, screenshots and PDFs are all fine. Use copies, not your originals, and start with a small batch while you're learning.
Open the Claude desktop app and add that folder. Under the message box, choose the option to work in a project or folder, and pick the folder you just made. Just that one folder.
Paste the prompt above, then add "show me your plan first" on the end, and send it.
Read the plan. Check it understood what's in the folder and what you want back. When you're happy, let it go ahead and allow it to save the spreadsheet in that folder.
Open the spreadsheet. Go to the Check these tab first and work through each flagged receipt using the reason it gave.
Spot-check the main tab. Pick a few rows and compare them against the original receipts. If the amounts and suppliers match, you can trust the rest more. If they don't, tell Claude what it got wrong and ask it to fix the spreadsheet.

Once you're comfortable, you can add new receipts to the folder each week and ask Claude to add them to the same spreadsheet, rather than saving the whole lot for tax time.

Keep it tight

Receipts carry more personal information than people think: card details, addresses, names, sometimes client names on a meal or a gift. So the same habits from Day 1 apply, and they matter a little more here.

One narrow folder. Only the receipts for this job. Not your downloads, not your desktop, not your whole computer. If it isn't in the folder, Claude isn't reading it.
Approvals on for anything that sends, spends or deletes. This job only needs Claude to read and save one new file. It shouldn't be emailing anything, paying anything or removing anything, and with approvals on, it can't without you saying yes.
"Show me your plan first." Add it to the end of the prompt so you see what it intends to do before it does it.

It's also worth being clear about what this is. Claude is doing the reading and the typing. It isn't your accountant. Whether something is deductible, and how GST should be treated for your situation, is still a question for your accountant or bookkeeper. What you're handing them is a tidy spreadsheet instead of a shoebox.

Where to go from here

The Check these tab is the habit worth taking away from today. Any time you give Claude a job with a lot of items, ask it to separate the ones it's sure about from the ones it isn't, and tell you why. It turns "check everything" into "check the short list", and that idea works well beyond receipts.

Start with one small batch. Check the flagged ones, spot-check the rest, and see how it went. Then do the next batch.

Missed yesterday? Day 6: Appraisal prep in 2 minutes

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Appraisal prep in 2 minutes: how real estate agents can use Claude to build a one-page appraisal brief

Matt Manser — Sat, 17 Oct 2026 09:00:00 +0000

I used to spend an hour getting ready for an appraisal.

Most of that hour wasn't thinking. It was gathering. Opening the notes from the first phone call, pulling up the comparable sales, finding the council details, flicking between tabs and trying to hold it all in my head on the drive over.

The thinking part, working out what this home is worth and why, only works once everything is in one place. That's the part worth your time. The gathering is a job, and it's one you can hand over.

The idea: one folder in, one page out

This builds on everything from earlier in the series. Instead of asking Claude about appraisals in general, you give it the job with your actual files.

Drop three things into one folder:

Your notes. Whatever you've got from the first call or email with the owner. Typed notes, a copied email, a quick text file.
The comparable sales. The recent sales you'd normally look at, exported or saved from wherever you usually get them.
The council info. Zoning, land size, anything you've saved about the property from the council or other records.

Then ask Claude for a one-page brief. It reads every file, pulls out what matters, and hands you a single page you can read in the car before you knock on the door.

You walk in knowing the numbers.

Why this matters for an agent's week

An appraisal is often your first proper meeting with a potential seller. They're deciding whether to trust you with what is likely their biggest asset. Walking in prepared, with the recent sales at your fingertips and a sense of what makes their home different, shows them you've done the work.

The problem is that appraisals don't arrive one at a time on a quiet afternoon. They land in between open homes, follow-up calls, contracts and everything else. When preparation takes a big chunk of time, it's the first thing to get squeezed.

A one-page brief doesn't replace your judgement. It just gets the facts lined up so your judgement has something to work with.

The Day 6 prompt

Prompt

Read everything in this folder about [address]. Make me a one-page appraisal brief: the property, the 3 most comparable recent sales with prices and dates, what makes this home different, and 3 questions I should ask the owner. Only use what's in the files.

Here's what each part is doing.

"Read everything in this folder about [address]." This tells Claude what to look at and which property it's about. Swap in the real address. If your folder has a sales export that covers a whole suburb, the address keeps it focused on the right home.

"Make me a one-page appraisal brief." This sets the shape of what you get back. One page is the point. You want something you can scan in a couple of minutes, not a report.

"The property." A short summary of the home itself, drawn from your notes and the council info: land, bedrooms, features, anything the owner told you.

"The 3 most comparable recent sales with prices and dates." This asks Claude to pick the closest matches from your sales file, not list everything. Asking for prices and dates means you can see straight away how recent each sale is and check it against your source.

"What makes this home different." This is where the brief becomes useful in the room. It pulls out what sets this property apart from the comparable sales, good or bad, so you're ready to talk about it.

"3 questions I should ask the owner." Based on gaps or details in your files, Claude suggests questions worth asking. Often these are things the notes hint at but don't answer.

"Only use what's in the files." This is the most important line. Without it, Claude may fill gaps with general knowledge or guesses. With it, if something isn't in your files, it should say so rather than make it up. In an appraisal, a wrong number is worse than a missing one.

Try it yourself

Make a new folder for one property. Name it after the address so it's easy to find.
Add the files. Your notes, the comparable sales and any council info you have. Copies, not originals.
Open the Claude desktop app and add that folder. Under the message box, choose the option to work in a project or folder, and pick the one you just made. Just that folder.
Paste the prompt and replace [address] with the real address.
Ask for the plan first if you want it saved as a file. Add "show me your plan first" to the end. Once you're happy, let Claude save the brief into that folder.
Check the numbers. Open your sales file and confirm the three sales, prices and dates match. This takes a moment and it's worth doing every time.
Read the questions. Keep the ones that fit, cross out the ones that don't, and add your own.

If the brief says something is missing, that's useful too. It tells you what to find or ask before the appointment.

Keep the safety habits

Appraisal files often hold an owner's name, phone number and personal circumstances. That makes the habits from earlier in the series matter even more here.

One narrow folder per property. Don't add your whole client database or your whole computer. If it isn't in the folder, Claude isn't reading it. If you missed why this matters, Day 5 covers it.
Keep approvals on for anything that sends, spends or deletes. This job only needs Claude to read files and save one new one. Nothing should be emailed or removed.
Ask for the plan first. Especially the first few times, so you can see what it intends to do before it does it.
Check what your office expects. The rules about how client information is handled depend on where you work, so it's worth knowing yours before you start.

Where to go from here

The brief isn't the appraisal. You still walk through the home, listen to the owner and form your own view on price. What changes is that you arrive with the facts already in order, so your attention goes on the people and the property instead of on remembering numbers.

Try it on your next appraisal. One folder, one prompt, one page. Then look at what you'd change, and adjust the prompt to suit the way you work.

Missed yesterday? Day 5: Don't give AI your whole computer

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Don't give AI your whole computer: the one-folder rule every real estate agent should follow with Claude

Matt Manser — Fri, 16 Oct 2026 09:00:00 +0000

Never give AI access to all your files.

Once you've seen Claude read a folder, plan a job and hand back a finished file, it's tempting to point it at everything. Your whole documents folder, your desktop, the lot. Then it could find anything, right?

It could. That's the problem.

The safer setup is also the simpler one. Make one folder called Claude Work, connect only that folder, and copy in what each job needs. If Claude can't see a file, it can't touch it.

A quick reminder on words. I call it SI, super intelligence. Most people say AI. When I say SI in this post I mean tools like Claude, made by Anthropic.

Why one folder beats your whole computer

When you connect a folder to Claude, you're giving it permission to read what's in there and, if you allow it, to make changes. That's what makes it useful. It's also why you want to be deliberate about which folder you give it.

What's sitting on an agent's computer

Think about what's on your laptop right now. Contracts. Vendor details. Buyer contact lists. Appraisal notes. Identification documents. Bank and commission paperwork. Probably some personal files mixed in as well.

Most of that has nothing to do with the job you want done today. If you're asking Claude to tidy up open home notes, it doesn't need to see a signed contract from last year. And as agents we hold a lot of personal information about clients, and the rules about how that's handled depend on where you work. The easiest way to stay on the right side of those rules is to only ever share what the job needs.

Less to look at means better work

There's a practical side too. A folder with the files for one job gives Claude one clear thing to work on. Point it at thousands of files and it has to work out which ones matter, which slows things down and makes mistakes more likely. Point it at five files and it gets straight to work.

Easy to check afterwards

With one folder, you always know where to look. When a job's done, you open Claude Work and see exactly what's there and what changed. Nothing is hidden in a corner of your computer you forgot about.

The Claude Work folder

Here's the habit in plain terms.

One folder, always the same name. Call it Claude Work so you always know which one is shared.
Copy files in, don't move them. Your originals stay where they are. Claude works on copies.
Only what the job needs. If you're writing a property description, copy in the property notes and photos list. Not the whole listing file.
Clear it out when you're done. Once you've saved the finished result somewhere else, empty the folder so it's ready for the next job.

That last point matters more than it sounds. If old files pile up in Claude Work, you slowly end up back where you started, with a folder full of things Claude doesn't need to see.

Keep it asking before it changes anything

The folder controls what Claude can see. The second safety habit controls what it can do.

The Claude desktop app has settings for how much it does on its own. Keep it on the setting where it asks before it changes anything. That way, before a file gets edited, created or removed, you get a chance to say yes or no.

It can feel slower at first. That's the point. You're still learning how Claude works, and every time it asks, you see what it's about to do. Over time you'll get a feel for which jobs you trust it with.

Two related habits are worth keeping from Day 1. Keep approvals on for anything that sends, spends or deletes, so nothing goes out or disappears without you saying yes. And end your jobs with "show me your plan first", so you see what it intends to do before it starts.

The Day 5 prompt

This prompt sets the ground rules at the start of a conversation. Paste it in first, then give Claude the actual job.

Prompt

I've connected one folder called Claude Work. Only ever work inside this folder. Before any job, tell me which files you plan to read or change, and never delete anything.

Here's what each part does.

"I've connected one folder called Claude Work." This tells Claude what it's working with. It knows the folder is deliberate and that this is the space you've set aside for it.

"Only ever work inside this folder." This is the boundary in words. The folder connection is the real limit, but saying it out loud makes your intent clear for the whole conversation, including where it saves new files.

"Before any job, tell me which files you plan to read or change." This gives you a checkpoint. Before it starts, you see a list. If something's on there that shouldn't be, or something's missing, you can fix it before any work happens.

"And never delete anything." Deleting is the one action that's hard to undo. Ruling it out from the start means the worst case is a file you don't like, not a file you've lost. If you ever need something removed, you can do it yourself.

Try it yourself

This takes a few minutes, and you only need to set it up once.

Make the folder. Create a new folder somewhere easy to find, like your desktop or documents, and name it Claude Work.
Copy in a few files for one small job. Pick something low-stakes, like notes from an open home or a draft property description. Copies, not originals.
Open the Claude desktop app and connect only that folder. Use the + in the message box to add a folder, and choose Claude Work. Just that one.
Check your permission setting. Make sure Claude is set to ask before it changes anything. If you're not sure where that is, look through the app's settings or the options near the message box.
Paste the Day 5 prompt and send it.
Give it the job. Ask for what you want, and add "show me your plan first".
Read the list of files it plans to use. Confirm it's only the ones in Claude Work, then let it go ahead.

When it's finished, open the folder and check the result. Then, while you're there, take a look at the Trusted folders list in Claude's Settings. Any folder you've connected shows up there, so make sure Claude Work is the only one you meant to share.

Where to go from here

The one-folder rule isn't about being nervous of SI. It's the same thing you'd do with a new assistant on their first day. You wouldn't hand them the keys to every filing cabinet. You'd give them the files for the job in front of them, and check their work until you trust it.

Set up Claude Work once, keep approvals on, and start each session with the Day 5 prompt. You'll get the benefit of Claude doing real work with your files, without handing over everything else.

Missed yesterday? Day 4: Projects: stop repeating yourself

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Claude Projects for real estate agents: stop repeating yourself every chat

Matt Manser — Thu, 15 Oct 2026 09:00:00 +0000

You're explaining your business to Claude every single time. Stop.

If you've been using Claude for a few days, you've probably noticed a pattern. You open a new chat, and before you can ask for anything useful you're typing the same background again. Who you are. Where you work. What you sell. How you like things written. What it must never do.

Then you close that chat, open a new one tomorrow, and do it all again.

There's a better way, and it's built into Claude. It's called a Project.

What a Project actually is

A Project is a space in Claude that remembers who you are. You put your business details, your voice and your rules in once, and every chat you start inside that Project already knows them.

Think of it like briefing a new assistant. You wouldn't re-explain your whole business to them every morning. You'd tell them once on day one, and from then on they'd just get on with it. A Project gives Claude that same day-one briefing, and it keeps it.

A Project usually has two parts:

Instructions. A short written briefing that applies to every chat in the Project. Who you are, how to write, what the rules are.
Files. Documents you add for Claude to refer to across those chats. Optional, and worth keeping tight, which I'll come back to.

Inside the Project you can start as many chats as you like. Each one starts with the briefing already in place.

Why this matters in an agent's week

So much of what we ask Claude to do depends on context. A listing description, a follow-up email, a social post, an appraisal summary. All of them come out better when Claude knows where you work, who you're writing for and how you sound.

Without a Project, you either retype that context every time or skip it and get something generic. Generic is the problem. It reads like it could have come from any agent in any city, and you end up rewriting it yourself.

With a Project, the context is already there. You can open a chat and just say "write a follow-up to someone who came through the open home on Saturday", and Claude already knows your area, your tone and your rules.

It also means your rules stay consistent. If you've told it never to make up a price, that rule is in place for every chat in the Project, not just the ones where you remembered to say it.

If you did yesterday's episode on making Claude sound like you, this is where that work earns its keep. Your voice notes belong in your Project instructions, so they apply everywhere.

The Day 4 prompt

This one is a little different. It's not something you paste into a chat. It's the starting text for your Project instructions.

Prompt

I'm a real estate agent with eXp Realty in Adelaide's north-east. I sell family homes in suburbs like Redwood Park, Modbury and Golden Grove. Write in plain Australian English, warm but direct. Never invent prices, dates or facts; ask me if you don't know.

It's short on purpose. Four sentences, and each one is doing a job.

"I'm a real estate agent with eXp Realty in Adelaide's north-east." This tells Claude who you are and where you work. Swap in your own brokerage and area. Now it isn't guessing whether you're in Sydney, Texas or London.

"I sell family homes in suburbs like Redwood Park, Modbury and Golden Grove." This tells it what you sell and who you're likely writing for. Use your own suburbs and property types. If you mostly sell units to investors, say that instead.

"Write in plain Australian English, warm but direct." This sets the voice. Australian spelling, no American phrasing, and a tone that fits how agents actually talk to clients. Change the adjectives to match you.

"Never invent prices, dates or facts; ask me if you don't know." This is the most important line. Claude is good at filling gaps, and in real estate a filled gap can be a wrong price, a wrong settlement date or a feature the property doesn't have. This line tells it to stop and ask instead. Keep it, whatever else you change.

Once this is in, you can add to it over time. Your sign-off. Words you never want used. How you format emails. Treat it as a living briefing, not a one-off.

Try it yourself

This takes about five minutes.

Open Claude and find Projects. In the Claude app or on claude.ai, look for Projects in the sidebar. Availability can depend on your plan, so if you can't see it, check your plan details.
Create a new Project. Call it something clear, like "My Real Estate Business".
Open the Project instructions. There's a place in the Project to set instructions. That's where your briefing goes.
Paste the Day 4 prompt and change the details to match your business. Your brokerage, your suburbs, your property types, your voice.
Save it, then start a new chat inside the Project. Ask for something ordinary, like a short email to a buyer after an inspection. Don't give it any background.
Read what comes back. Check it's used Australian spelling, sounds like you and hasn't made up anything it couldn't know. If it asks you a question instead of inventing a detail, that's the rule working.
Compare it. Start a normal chat outside the Project and ask the same thing. Put the two side by side. You'll see what the briefing is doing.

If something in the result isn't right, don't fix it in the chat. Fix it in the instructions, so every future chat gets the improvement too.

Keep the Project tight

The same safety habits from Day 1 apply here, with a Project-shaped twist.

Be careful what files you add. Anything you add to a Project is available to every chat inside it. Start with general material, like a style guide or a description of your services, not client files. As agents we hold a lot of personal information, and the rules for handling it depend on where you work. If it isn't in the Project, Claude isn't reading it.
One Project per purpose. It's tempting to build one giant Project for everything. A focused Project with a clear purpose gives cleaner results, and it's easier to see what's in it.
When you give it a job with your files, still use one narrow folder. A Project briefing doesn't change that. Add the one folder for the job, never your whole computer.
Keep approvals on for anything that sends, spends or deletes, so nothing goes out or disappears without you saying yes.
End bigger jobs with "show me your plan first". Even with a good briefing, you want to see what it intends to do before it does it.

Your instructions are a good place to put these rules too. A line like "always show me your plan before changing any files" will apply to every chat in the Project.

Where to go from here

A Project is one of the simplest changes you can make, and it changes every chat after it. Set it up once, then keep improving the instructions as you notice what Claude gets wrong or what you keep having to repeat.

If you find yourself typing the same thing into a chat twice, that's your signal. It belongs in the Project.

Missed yesterday? Day 3: Make it sound like you

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Make it sound like you: how to teach Claude your writing voice as a real estate agent

Matt Manser — Wed, 14 Oct 2026 09:00:00 +0000

If your AI writes like a robot, it's because you never showed it how you talk.

That's the most common complaint I hear about Claude and every tool like it. The writing is fine, but it doesn't sound like a person. It's too polished, too long, too keen. It says "I hope this message finds you well" to someone you've known for ten years.

That's not a flaw in the tool. It's a gap in what you've given it. Claude has never read anything you've written, so it falls back on a safe, generic voice. Give it a few real examples and that changes.

A quick reminder on words. I call it SI, super intelligence. Most people say AI. When I say SI I mean tools like Claude, made by Anthropic.

Why your voice matters more than you'd think

Clients don't just read what you send. They hear it in your voice.

Your vendors know how you talk on the phone, at the open, across the kitchen bench at the appraisal. If the written update they get on Monday sounds like a different person, they notice. It might not be a big thing, but it chips away at the feeling that they're dealing with you.

And so much of an agent's week is writing. Vendor updates after the weekend. Follow-ups after opens. Replies to enquiries. Quick texts to confirm times. If Claude is going to help with any of that, it has to sound like you, or you'll end up rewriting every draft yourself. At that point it hasn't saved you anything.

The idea: show it, then name it

The fix has two parts.

Show it real examples

Don't describe your style. Most people are bad at that. Ask an agent how they write and they'll say "friendly but professional", which is what everyone says and doesn't tell Claude much.

Instead, give it the real thing. Three messages you've actually sent to clients. Not your best work, not something you've polished for the occasion. Just normal messages that sound like you on a normal day.

Get it to name what it sees

Then ask Claude to describe your voice back to you in a few plain points. This step matters. It makes Claude show you what it picked up, so you can check it got it right before it writes anything. And it gives you a short description of your voice you can reuse every time.

Once you have that, you're not starting from zero with each new draft.

The Day 3 prompt

Prompt

Here are 3 messages I've written to clients: [paste]. Describe my writing style in 5 bullet points: tone, sentence length, words I use, words I'd never use. Then rewrite this vendor update in my voice: [paste].

Here's what each part is doing.

"Here are 3 messages I've written to clients: [paste]." This is the evidence. Three is enough to show a pattern without overloading it. Pick messages that are a bit different from each other, say a text, an email and a longer update, so it sees your range.

"Describe my writing style in 5 bullet points." Five points keeps it short and specific. You want something you can read in ten seconds and say "yes, that's me" or "no, that's not right".

"tone, sentence length, words I use, words I'd never use." These four tell Claude exactly what to look for. Tone is how you come across. Sentence length is how short or long you write. Words you use are your habits, the phrases you lean on. Words you'd never use is often the most useful one, because it stops Claude slipping in the stiff, salesy language that gives it away.

"Then rewrite this vendor update in my voice: [paste]." This is the test. Give it a vendor update that's correct but doesn't sound like you, maybe one Claude wrote earlier or one from a template. Seeing the before and after side by side is the quickest way to tell whether the description worked.

Try it yourself

Pick three messages you've sent to clients. Texts or emails, whatever you write most. Choose everyday ones, not your best.
Take out anything personal. Remove names, phone numbers, addresses and anything else that identifies a client. Swap in placeholders like "[vendor name]" or "[address]". Your voice is in how you write, not in the details.
Find a vendor update to rewrite. One that has the right information but doesn't sound like you. Again, strip out client details.
Open Claude and paste the prompt above, dropping your three messages and the vendor update in where it says [paste].
Read the five bullet points first. Do they sound like you? If one is wrong, tell Claude. "I'm not that formal" or "I never use exclamation marks" is enough. Ask it to update the list.
Then read the rewrite. If it still feels off, say what's off and ask for another go.
Save the five bullet points. Copy them into a note somewhere you'll find them. That's your voice description.

You don't need a folder for this one. Pasting straight into the chat is fine.

Use it every time

The five points are the bit you keep.

The simplest way to use them is to paste them at the top of any writing job. Something like "Write this in my voice. Here's how I write:" followed by your five points, then the job.

If you'd rather not paste it each time, Claude has places to store instructions it should always follow. Many people put their voice description in their profile preferences in Settings, or in the instructions for a project they use for client writing. Menus move around between versions, so if you can't find it, ask Claude where to put standing instructions in the version you're using. Claude also has writing style options, and some versions let you create a custom style from your own samples, which does a similar job.

Whichever way you do it, check it now and then. Your voice with a first-home buyer might be different to your voice with a developer. It's fine to have more than one description.

A few safety habits

Writing in your voice makes the drafts more convincing, which is exactly why you should still read every one before it goes out.

You send, not Claude. If you ever connect Claude to email or messaging, keep approvals on for anything that sends. A message in your voice that you didn't check is worse than a robotic one, because the client will assume you wrote it.
Keep client details out where you can. As agents we hold a lot of personal information, and the rules about how that's handled depend on where you work. Placeholders work just as well for teaching it your voice.
Ask for the plan on bigger jobs. If you're getting it to rewrite a batch of templates in a folder, give it one narrow folder and ask it to show you its plan first before it changes anything.

Where to go from here

Three messages, five bullet points, one rewrite. That's the whole job, and it fixes the most common reason people give up on Claude for writing.

Once Claude knows how you talk, everything it drafts for you after that starts closer to finished. You're editing, not rewriting.

Missed yesterday? Day 2: Your first real job in Chat

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Your first real job in Claude Chat: turn listing appointment notes into a portal-ready description

Matt Manser — Tue, 13 Oct 2026 09:00:00 +0000

Stop asking Claude questions. Give it a job.

Yesterday was about the difference between asking Claude something and handing it a job. Today you do your first real one, and you don't need a folder or any setup to do it. Just Chat, and the notes you already take at every listing appointment.

Dot points from your listing appointment go in. A portal-ready description comes out.

Why this is a good first job

Writing listing descriptions is a perfect first job to hand over, for a few reasons.

You already have the raw material. Most agents walk out of a listing appointment with notes: bedrooms, the renovated kitchen, the north-facing yard, the walk to school. The facts are done. What's left is turning them into readable copy.
It's slow, not hard. You know what the home is like. Finding the words, again, for the next three-bedroom home, is the part that eats time.
It's easy to check. You were in the house. You'll know straight away if something in the description is wrong.

That last point matters. A good first job is one where you can judge the result yourself, without needing to trust Claude blindly.

Asking vs giving it the job

Here's the same request done both ways.

Ask it as a question, like "how do I write a good listing description?", and you get tips. Use active language, lead with the best feature, keep it short. Useful, maybe, but you still have to sit down and write it.

Give it the job instead, with your notes pasted in, and you get the description itself. Then your work becomes reading and editing, not writing from a blank page.

The three things Claude needs from you

A description job goes well when you give Claude three things. Leave any of them out and you'll get something generic.

1. The facts

Your dot points. Paste them in as they are. They don't need to be tidy, full sentences or in any particular order. Claude can sort that out.

What matters is that the facts are there. Claude can only describe what you tell it, so the more specific your notes, the better the description. "Big backyard" gives it less to work with than "flat, grassed backyard with side access".

2. Who the buyer is

The same house can be written for different people. A young family wants to hear about bedrooms, the yard and the school. A downsizer cares about low maintenance and single-level living. An investor wants the rental appeal.

You know who's most likely to buy. Tell Claude, and it will lead with what that buyer cares about.

3. What not to say

This is the part most people skip, and it's the one that makes the biggest difference to quality.

Without guidance, AI-written property copy tends to fall back on the same tired phrases you've seen a thousand times on the portals. Telling Claude which words to avoid pushes it toward plainer, more honest writing.

More importantly, tell it not to invent anything. If your notes don't mention a renovated bathroom, the description shouldn't either. You're responsible for what goes into the ad, so you want every claim in it to come straight from your notes.

The Day 2 prompt

Prompt

Here are my notes from a listing appointment: [paste dot points]. Write a warm, honest description for realestate.com.au, 150 words, for a young family buying their first home. No clichés like 'nestled' or 'boasting', and don't invent anything that isn't in my notes.

Here's what each part is doing.

"Here are my notes from a listing appointment: [paste dot points]" gives it the facts and tells it where they came from, so it knows these are the details to work from.
"Write a warm, honest description" sets the tone. Warm, so it reads like a home, not a spec sheet. Honest, so it doesn't oversell.
"for realestate.com.au" tells it where the copy is going, so it writes something suited to a portal listing rather than a brochure or a social post.
"150 words" keeps it to a length that's easy to read on a phone. Change the number if your agency or the listing calls for something different.
"for a young family buying their first home" is the buyer. Swap this for whoever is actually most likely to buy the property.
"No clichés like 'nestled' or 'boasting'" steers it away from the phrases buyers skim past. Add any others you're sick of seeing.
"don't invent anything that isn't in my notes" is the most important line in the prompt. It keeps the description tied to the facts you gave it.

Try it yourself

This takes a few minutes, and it's a good one to try on a listing you're writing up this week.

Get your notes ready. Use dot points from a recent listing appointment. Leave out owner names, phone numbers and anything else personal. Claude doesn't need any of that to describe a house.
Open a new chat in Claude. No folder needed for this one. You're pasting the information straight in.
Paste the prompt and replace [paste dot points] with your notes. Change the buyer if a young family isn't the right fit.
Send it and read the result carefully. Check every claim against your notes. If anything has crept in that you didn't say, point it out and ask Claude to remove it.
Ask for changes in plain English. "Make it a bit less formal." "Lead with the backyard." "Cut it to 120 words." You don't need to start again. Just tell it what to change, the same way you'd give feedback to an assistant.
Copy the final version into your listing once you're happy with it, and give it your usual final read before it goes live.

If you want to see the difference the details make, run it a second time with the buyer changed to a downsizer or an investor. Same notes, a noticeably different description.

A few habits to keep

The safety habits from Day 1 still apply, even for a job this simple.

Only share what the job needs. Here that means property details, not client details. If you later move to working with files instead of pasting, give Claude one narrow folder for the job, never your whole computer.
You publish, not Claude. Claude writes the draft. You check it and put it on the portal. Keep approvals on for anything that sends, spends or deletes, so nothing goes out without your say-so.
Check before it goes anywhere. For bigger jobs, end your prompt with "show me your plan first". For a description, the equivalent is simple: read every line before it's published.

Where to go from here

This is the pattern for almost every job you'll hand Claude: give it the facts, tell it who it's for, and tell it what to avoid. Once it clicks for descriptions, you'll start seeing the same shape in other parts of your week.

Try it on your next listing. Keep the prompt somewhere handy and adjust the buyer and word count each time.

Missed yesterday? Day 1: Ask it vs give it a job

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.

Ask it vs give it a job: the first thing every real estate agent should understand about Claude

Matt Manser — Mon, 12 Oct 2026 09:00:00 +0000

This is where I started with Claude, and most agents never get past it.

They open Claude, type a question, get a decent answer, and decide that's what it's for. A smarter search engine. Something to help word an email. And it is good at that. But it's only half of what's sitting in front of you, and it's the less useful half for someone running a real estate business.

The other half is giving it a job. Not asking how to do something, but handing the thing over and getting the finished result back.

A quick note on words before we go further. I call it SI, super intelligence. Most people say AI. When I say SI in this post I mean tools like Claude, made by Anthropic. Same thing, I just think the name undersells it.

Same Claude, two ways to use it

There's no special mode to switch on and nothing extra to buy. It's the same Claude in the same window. What changes is whether you ask it something or give it a job, and whether it has your files to work with.

Asking: you get advice

Ask Claude a question with no files attached and you get advice back. Often good advice. But it's working blind. Ask it about your own documents and it will tell you it can't see them, then give you general tips instead.

Notice who's still doing the work. Claude tells you what to do, and then you go and do it. You're still the one opening the files, sorting things out, copying text into the right place and saving it.

Giving it a job: you get the finished thing

Now give it the job instead. Add one folder of your files, say what you want back, and Claude takes it from there. It reads the files, plans the job, works through them and hands you the finished thing.

That's the whole shift. Same Claude, completely different result.

What that looks like in practice

The easiest way to see it is to take one request and try it both ways.

Ask it as a plain question, something like "what's in my listing files and what could you do with them?", with nothing attached. It can't see your files, so it says so and gives you general ideas. Then it's over to you.

Give it the same request as a job, with one folder added, and it's a different conversation. It reads what's actually in the folder, lists what's there in plain English, and suggests three jobs based on your real files. Pick one, say yes, and it does the work and saves a new file in that folder for you to open.

For a real estate agent that difference matters more than for most people. So much of our week is admin that isn't hard, just slow: notes from appointments, property details, marketing copy, documents that need checking against each other, lists that need tidying. Those are jobs, not questions. If all you ever do is ask Claude questions about them, you're still doing every one of them yourself.

Questions vs instructions

The thing that changes isn't the tool. It's how you talk to it.

When you're asking, you tend to write questions: "What's the best way to...?", "Can you give me tips for...?" When you're handing over a job, you write instructions, the same way you'd brief a good assistant on their first day:

What to look at. Which folder, which files.
What you want back. A summary, a list, a finished document, a tidied file.
What not to touch. Especially at the start.

That last one is important, and it's why the prompt below ends the way it does.

The Day 1 prompt

This is the first prompt from my 21 Days of SI series. It's deliberately simple, and it doesn't change anything. It just gets Claude to look at what you've given it and tell you what it could do.

Prompt

Look through every file in this folder. Tell me in plain English what's here, then suggest 3 jobs you could do with these files that would save me time. Don't change anything yet.

Three things are going on in there. "Every file in this folder" tells it what to look at. "Tell me in plain English what's here" makes it show you it understood before it does anything. "Don't change anything yet" means nothing gets moved, edited or deleted while you're still learning how it works.

The three suggested jobs are the useful bit. They're based on your actual files, not generic tips, so they're a good way to find your first real job to hand over.

Try it yourself

You can do this in about the time it takes to make a coffee.

Make a new folder. Copy in a handful of everyday work files you're comfortable with Claude reading. Copies, not originals.
Open the Claude desktop app and add that folder. Under the message box, choose "Work in a project or folder" and pick the folder you just made. Just that one folder.
Paste the prompt above and send it.
Read what comes back. Check it's described your files correctly, then look at the three jobs it suggests.
Pick one job and ask it to show you its plan first. When you're happy, let it go ahead and allow Claude to edit files in that folder. It saves the finished file there for you to check.

Then, if you want to really see the difference, start a new conversation with no folder added and ask the same thing as a plain question. Put the two answers side by side. One tells you what you could do. The other has already done it.

Add one folder, not your whole computer

If you take one tip from this post, make it this one: give Claude one narrow folder for the job. Never your whole computer.

There are a few reasons.

It keeps Claude focused. A folder with the files for one job gives it one clear thing to work on.
It keeps everything else private. As agents we hold a lot of personal information about clients, and the rules about how that's handled depend on where you work. If it isn't in the folder, Claude isn't reading it.
It's easy to check. When the job's done you can see exactly what's in that one folder and what changed.

Any folder you've connected shows up under Trusted folders in Claude's Settings, so it's worth a look now and then to make sure the list is only what you meant to share.

Two more habits are worth building from day one. Keep approvals on for anything that sends, spends or deletes, so nothing goes out or disappears without you saying yes. And end your jobs with "show me your plan first", so you see what it intends to do before it does it.

You can loosen the reins once you trust it. Start tight.

Where to go from here

Asking still has its place. For a quick question or a first draft, it's perfect. But if you've only ever used Claude that way, you've been using it as an adviser when it can be an assistant.

Start with one folder and the Day 1 prompt. Read what it tells you. Pick one of the three jobs and let it do it. That's the first step from asking Claude things to having it do things for you.

Every prompt from the series is on my free prompts page, with a new one each day as the episodes go up, ready to copy.