The best AI operators in commercial real estate aren’t the ones with the best prompts.

They’re the ones who stopped prompting.

Almost every operator I work with does the same thing when we first meet: they re-explain their process to Claude every single time. Deal criteria. Report format. What a red flag looks like in a T-12. Typed fresh, every morning, like it’s day one.

Then they close the chat, and everything they just taught the AI evaporates.

That’s the tax nobody talks about. Not the subscription cost. The re-teaching cost.

New models drop every month. Agents, browser automation, computer use - the toolbox keeps growing. But while everyone chases the shiny stuff, the skill that actually separates operators right now is boring and old-fashioned: the ability to write down a repeatable process so clearly that software can run it.

In AI terms, that written-down process is called a skill file. It’s a plain text file that tells Claude how YOU do a specific task - your steps, your standards, your judgment calls. Load it once, and every future run starts from your best day instead of a blank page.

This is how we turned a 2-hour lease abstract into 10 minutes. Not a magic prompt. A written process that runs the same way every time.

Today I’m giving you ten ways to think about creating one, plus the exact template I use.

If this already has your wheels spinning - you’re going to want the rest of this series. Subscribe here so you don’t miss it.

Subscribe now

Ten Ways to Think About Creating a Skill File

1. A skill file is an SOP your software actually runs

CRE shops love SOPs. They also love ignoring them. The underwriting checklist lives in a binder or a SharePoint folder nobody opens, and the actual process lives in your senior analyst’s head.

A skill file is the SOP that stops being decorative. You write the process down once, and Claude executes it - the same steps, the same standards, on deal 1 and deal 100. The document and the doing become the same thing.

If you’ve ever written an SOP, you already know how to write a skill file. You just have a better employee to hand it to now.

2. Write it for your best analyst’s first day

Picture the sharpest analyst you’ve ever hired, on their first morning. Smart, fast, zero context on how your shop does things.

That’s exactly who you’re writing for. Don’t explain what NOI is. Do explain that your shop underwrites management at 5% no matter what the OM claims, that you never trust the seller’s tax number, and that expense growth gets stress-tested at 3 scenarios before anything hits the IC memo.

The generic knowledge is already in the model. Your file carries what makes it YOUR process.

3. Doing the task once, out loud, is your first draft

Don’t sit down to “write documentation.” You’ll stall.

Instead, do the task one more time and narrate it. Talk into a voice memo while you screen a deal. Or do it in a Claude chat and explain each step as you go: “first I pull these 4 numbers, then I check them against this, if they’re off by more than 5% I do this.”

Then hand the transcript to Claude and ask it to draft the skill file. You’re not a technical writer and you don’t need to be. You’re the subject matter expert dictating to one.

Here’s the template that transcript should fill:

Copy-paste template:

# Skill: [Name of the task, e.g., "Deal Screen - Multifamily"]

## When to use this
[The trigger. E.g., "When a new OM or broker package comes in
for a multifamily deal between 50 and 200 units."]

## What goes in
[The inputs. E.g., "The OM (PDF), the T-12, the rent roll."]

## What comes out
[The finished product. E.g., "A one-page screen memo with:
deal snapshot, my 5 screening metrics, red flags, and a
pass/dig-deeper verdict."]

## Steps
1. [First thing you always do]
2. [Second thing]
3. [Keep going - every step you'd teach a new analyst]

## Judgment calls
- If [situation], then [what you do and why]
- If [situation], then [what you do and why]
- Never [the thing that would embarrass you in front of the IC]

## What "done" looks like
[How you know the output is right. Standards, format, the
checks you run before you'd put your name on it.]

## Worked example
[Paste one real, anonymized before/after: the raw input
and the output you'd consider an A+.]

That’s it. No code. A text file a 1990s word processor could open.

Know an operator still re-typing their deal criteria into ChatGPT every morning? Forward this to them. This template alone will save them hours a week.

Share

4. Name the trigger, not just the task

A prompt answers “how do I do this?” A skill file also answers “when does this happen?”

“Abstract this lease” is a task. “When a new lease lands in the deal folder, abstract it into our standard 14-field summary and flag any non-standard clauses” is a process. The trigger is what turns a party trick into an operating system - because once the “when” is written down, the task can start firing without you initiating it.

Weak skill files describe work. Strong ones describe when work begins.

5. Define done before you define steps

Most process documentation fails at the finish line. It describes activity and never describes the standard.

Before you write a single step, write the “what done looks like” section. What does an A+ lease abstract contain? What makes an investor update send-ready versus draft-quality? If you can’t articulate the standard, no AI can hit it - and honestly, neither can your team.

This section is also your quality control. When Claude’s output misses, you don’t rewrite the prompt. You sharpen the definition of done and every future run inherits the fix.

6. Encode the judgment calls, not just the steps

Steps are cheap. Anyone can list steps.

The value of a 20-year operator is in the exceptions. The seller’s “actual” taxes that you know get reassessed at sale. The 97% occupancy that’s really 89% economic. The tenant improvement number that means the broker is hiding something.

Your skill file’s “judgment calls” section is where your experience gets written down - as if/then rules. “If occupancy is above 95% in a submarket averaging 90%, check for concessions buried in the rent roll.” That single line does more work than ten steps, because it’s the line a junior analyst wouldn’t know to apply.

This is the section that makes the file yours instead of generic.

7. One job per file

The temptation is to write “The Underwriting Skill” - one giant file that does everything.

Resist it. A file that screens deals AND builds models AND drafts IC memos does all three badly, and you can never tell which part broke. Split them. Deal screen is one file. Model build is another. IC memo draft is a third.

The test: describe the skill in one sentence. If that sentence needs the word “and,” you’re looking at two files.

Quick question: which process in your shop gets repeated the most - deal screening, lease abstraction, investor reporting, or something else? Hit reply and tell me. I read every response, and I’ll build out the most-requested one in a future issue.

In the link below, you’ll also find a free skill repository that I open sourced a while back. Check it out

https://github.com/ahacker-1/cre-agent-skills

And if someone forwarded this to you - subscribe here so you get the next one directly.

Share

8. Include one worked example from a real deal

One real before/after teaches more than twenty instructions.

Paste in an actual input (anonymized - strip the names and addresses) and the output you’d consider perfect. “Here’s a raw rent roll, here’s the exact summary I want from it.” Claude pattern-matches against your example far more reliably than it follows abstract rules.

Teachers have known this forever. Show the finished essay, not just the rubric.

9. Ship it ugly, version it forever

Your first skill file will be incomplete. Write it anyway. Twenty minutes, rough edges and all.

Then run it on real work, and every time the output misses, add one line. The occupancy trap you forgot. The formatting rule you assumed was obvious. Version 5 of an ugly file beats version 1 of a perfect one that never shipped - because version 5 contains five rounds of contact with reality.

This is the piece prompt-chasers never get. A prompt gets tweaked and lost. A file gets improved and kept.

10. Run the compounding test

One question tells you whether you’re building an asset or burning time:

Does the 100th run cost less than the 1st?

With prompting, run 100 costs exactly what run 1 cost - your time, your attention, your re-explaining. Flat line. With a skill file, run 100 is faster, better, and needs less of you than run 1, because 99 runs of corrections are baked into the file.

That’s the back-to-basics argument in one sentence. Prompts are an expense. Skill files are equity.

The Prompt That Writes Your First One

You don’t have to start from a blank template. Narrate your process once (way #3), then run this:

Copy-paste prompt:

You are a process documentation specialist for a commercial
real estate firm.

Below is a transcript of me doing and explaining one of my
repeatable processes. Turn it into a skill file with these
sections: When to use this / What goes in / What comes out /
Steps / Judgment calls / What "done" looks like.

- Capture every step in the order I actually did it
- Pull out every if/then decision I made, even ones I
  mentioned in passing, and list them under Judgment calls
- Where my explanation was vague, insert [CLARIFY: question]
  so I can fill the gap
- Write it so a brand-new analyst could follow it without
  asking me anything

Format as a clean markdown document.

Here's the transcript:
[PASTE YOUR TRANSCRIPT OR VOICE MEMO TEXT]

Answer the [CLARIFY] questions, paste in one worked example, and you have version 1. Twenty minutes, start to finish.

Which Process Are You Writing Down First?

Here’s my challenge for the week:

Pick the one task you’ve explained to AI (or to a junior analyst) more than 3 times.

Do it once more, narrating as you go.

Run the prompt above. Save the file.

P.S. Know an acquisitions lead or asset manager whose entire process lives in their head? Forward this to them. Their team will thank you the next time that person goes on vacation.

Share

And if this landed in your inbox because someone smart sent it your way - subscribe here so you get the next one directly.

Subscribe now

Until Next Time,

Avi Hacker, J.D.

Founder, The AI Consulting Network

On August 11, an email hit my phone from Manus.

ACTION REQUIRED: Some data in Avi will be deleted, back up data by August 23, 2026 at 7:59am SGT.

Read more

At approximately 1 a.m., my wife and I were sitting at gate A18 in Athens International Airport.

Our Israir flight to Tel Aviv was approaching its scheduled departure time, but boarding had not started. There was no meaningful announcement and very little information available at the gate.

So I opened ChatGPT and asked a question I never expected to ask:

I…

Read more

I’m opening up office hours tomorrow, Wednesday, August 5, from 11:00 AM to 12:00 PM ET.

Bring a workflow, AI question, or real estate use case you’re working through. There’s no formal presentation, just live problem-solving and practical answers.

Please upgrade to see Zo…

Read more

Bring your questions and join me using the Zoom link below:

Subscribe

Read more

At 2:09 yesterday afternoon, I created an empty project folder.

Before that same minute arrived today, strangers could open a browser and race what was inside it.

It has a 24-round championship, 22 drivers, 11 teams, qualifying, standing starts, tyre wear, pit stops, penalties, active aero, electric boost, persistent standings, and 21 AI opponents trying …

Read more

It’s 12:05 AM. OpenAI shipped voice control in its desktop app about two hours ago, and I’m still up, talking to my computer instead of sleeping.

I couldn’t wait to tell you.

Here’s what it does.

You open the Codex desktop app (or ChatGPT desktop), turn on voice mode, and just talk.

Out loud.

You ask what’s happening on your computer, and it looks.

You te…

Read more

Codex took control of my browser for about 45 minutes and built an entire Google Ads campaign while I watched.

Google accepted 50 campaign changes and one responsive search ad with zero errors.

But the clicking w…

Read more

Read more

On July 7, Claude Fable 5 was supposed to come off your subscription.

It didn’t. Anthropic extended the deadline to July 12, hours before the cutoff. Then they extended it again to July 19, and threw in 50% higher Claude Code rate limits for good measure. Then on Friday they stopped pretending it was temporary: starting July 20, Fable 5 is permanently in…

Read more

I wanted to see if I could build a version of myself.

Not a chatbot with my name on it. An actual local model, trained on my own methodology and how I think, running on my own hardware, that someone could ask a question and get an answer that sounds like mine. No data leaving the machine. No monthly API bill. Mine.

This was my first real attempt at the lo…

Read more

You can use this entire newsletter as the setup prompt. Open Claude Code on the computer that will run OpenClaw, paste in the full article, then add: “Walk me through this setup one step at a time. Adapt the files to my installed OpenClaw version, Slack workspace, CRE role, approved users, and approved folder. Show me every change before you apply it.”

C…

Read more

I need to ruin something for you.

Every “24/7 AI agent” being sold right now - the ones with the demo videos, the waitlists, the monthly subscription - is two things stapled together:

A text file with instructions in it. And a timer.

That’s it. The agent “wakes up” on a schedule, reads its instructions, does the task, and goes back to sleep. The minutes be…

Read more

You stopped talking to your AI for a reason.

You tried voice mode a year ago. You asked it something real - where cap rates are trending, what a rate cap costs, whether a lease clause is standard. It gave you a confident answer built on 2024 data, couldn’t check a single current fact, and went quiet every time you tried to cut in.

So you went back to typi…

Read more

Quick update because the clock changed.

On Friday, I told you Fable 5 was included inside paid Claude plans through July 7, and that the smart move was to point it at the deal you’ve been avoiding.

Anthropic just extended the window.

Fable 5 is now included on paid plans through Sunday, July 12.

Same basic rules as before: you can use Fable 5 for up to 50% …

Read more

There’s a deal on your desk you keep not looking at.

The refi you’ve been circling. The messy workbook someone inherited from an analyst who left. The 50-lease portfolio nobody has actually read line by line.

You’ve been telling yourself you’ll get to it.

Here’s why this week is different.

The most capable AI model that exists right now is sitting inside th…

Read more

Read more

Every operator you compete with now has the same AI you do. Same Claude, same ChatGPT, same models, dropping the same week for the same twenty bucks a month.

So here’s the uncomfortable question: if the tool is identical for everyone, where is your edge supposed to come from?

I’ll show you the answer in one number.

I just built a 68-second launch film for…

Read more

Your AI underwriting prompt is not the problem.

Not yet.

The problem is that nobody is keeping score.

Most CRE teams do the same thing:

They upload an OM.
They tweak the prompt.
The output looks cleaner.
The summary sounds smarter.
Everyone feels like the workflow improved.

But here is the uncomfortable part:

A prettier answer can be a worse answer.

If your AI ext…

Read more

Anthropic released Claude Fable 5 on Monday.

By Friday, it was gone.

Not because customers rejected it.

Not because it failed benchmarks.

Not because a better model came out four days later.

The U.S. government issued an export-control directive restricting access to Fable 5 and Mythos 5 by foreign nationals. Anthropic said the order applied whether the fore…

Read more