How I Stopped Retraining AI on Every Deal

There's a lot of AI slop out there right now. I'm going to try to make this the opposite.

This is how I built a framework for running acquisitions inside Claude Cowork. Nothing about the property gets lost. Every task I teach it, it keeps, and it gets sharper every week. When I open the next property, it already knows everything I taught it on the last one. No reteaching.

This isn't technical, I promise. It's just a different way to work with AI on long projects like an acquisition or a due diligence, so the training compounds instead of starting from scratch.

Back to round one, every time

Here's what used to happen. I'd ask it to abstract a lease, for example. It'd hand me back some AI crap. I'd give feedback. Wrong again. Ten rounds in, it finally did it my way. Then the chat would get too long and forget what we agreed on at the top. And none of that training came over to the next property. Back to round one.

One folder per property

So here's how I solved it. I stopped working in the chat app and moved to Claude Cowork, which runs off my own Mac, out of my Documents folder. Every property gets a folder. Inside is an index file that tells Claude where everything is, plus three subfolders: Reference, Skills and Outputs.

Index

The index tells it what's here, what to trust, and which job to run when. It's the memo you'd hand a new analyst on day one. Here the analyst is Claude, and the index is the "master SOP" pointing to every skill (every SOP) underneath it.

Reference

Reference is everything about the deal. Purchase agreement, title work, rent roll, inspection report. And every conversation: tenant interviews, Zoom calls, the planner, the fire marshal, the sprinkler guy. I drop the transcripts in, and a task runs every day that pulls any email about the property into the same place. The context stays current.

That one is quietly the biggest win. A lot of the time, deals stop moving smoothly because you're waiting on information that already exists somewhere. A document, a call from two months ago, an email nobody remembers. Now I ask the AI and have it in thirty seconds instead of waiting two days for someone to confirm.

Skills

Skills is where the knowledge compounds. Every week I teach it my way, and that knowledge moves out of my head and onto my Mac, where it gets a little better each time. The first time I taught it to abstract a lease took an afternoon of corrections. Now it already knows my way, instantly, on every lease since. Proforma, zoning research, critical dates, the weekly update: same thing.

Due diligence is 40 to 70 tasks. Each one gets taught once instead of every time, and none of them start from zero anymore. Some of them Claude now runs on its own without me asking.

Outputs

Outputs are the files Claude produced in that project. Pretty much anything digital: the PSA, the LOI, redline comments, proformas, decks, debt memos, emails. This compounds too. Every week a few of them get a little better, and that better version is where the next property starts.

Starting the next deal

Now when I get a new deal, I make a folder for it on my computer, start a new project in Cowork, and tell Claude to wire in the skills from the last one. Forty-five seconds later, that deal has an analyst who already knows how I underwrite and how I run an acquisition due diligence.

And when a new model comes out, ChatGPT or Gemini or whoever is faster and better, I point it at the same folder. It's up to speed instantly. Nothing to retrain. This analyst also doesn't quit after five years and walk out the door with everything you taught it. That knowledge is on your computer.

Try it

Index, Reference, Skills, Outputs. It works for me, and it'll work for you.