AI is powerful but how much does it really know about you? Every new chat, you re-explain who you are, what you sell, and which deals matter. This week we fixed that permanently. We built a memory layer that lives on our own laptop, that any AI can read, and that gets smarter every time we drop a document into it. We ingested two real offering memorandums and then asked it questions no third-party tool could answer. Stick with us all the way down. The full build is sitting at the bottom, waiting for you.

Brdg is the development management platform built for the messiest part of a project: construction financing. It's for the developers, lenders, and cost consultants who live in draw schedules and disconnected spreadsheets. Founded by Ness Cabessa, Samuel Brand, and Daniel Bensoussan, who paired civil engineering, finance, and software backgrounds to rebuild how capital moves through a build. The platform has already processed 30,000+ construction documents and manages $300M+ in active development.

What makes it different is that it treats construction finance as infrastructure, not another dashboard. Brdg ingests project documents, flags risk with an intelligence agent, and runs the financial control layer from pre-financing all the way through draw disbursement. One customer cut 5.5 days off their average draw cycle. On a $130M project, that is real money and real headaches gone.

We built a local knowledge base in Obsidian that any AI can read, then taught Claude Code to file new deals into it and pull answers back out. Two real offering memorandums went in. A proprietary comp set came out.

Here is the whole thing, from empty folder to answering questions on our own data.

What you'll need

Step 1: Create an empty vault

In Obsidian, we opened Manage Vaults, hit Create, named it "CRE Broker Demo," and saved it in our Documents folder. That is it. A blank vault is just a folder your AI can see. Everything we add from here lands in that folder as a markdown file, which is the format AI reads most cleanly.

Why this matters: the vault is local. Your deals never leave your machine, and no single AI vendor owns them.

Step 2: Point Claude Code at the vault and let it build the structure

This is the move most people get wrong. They start making folders by hand. Instead, we handed Claude one setup prompt and told it to build the whole skeleton. The prompt, verbatim from the start:

Then we told it to also create a README, an index, blank templates for every note type, structured front matter so notes are searchable, duplicate-checking rules, and wiki-link rules to connect contacts, listings, comps, and notes.

The "do not create another vault" line is not optional. Without it, the AI happily builds a second folder somewhere else and you lose half your work. Be specific or it will make assumptions, and assumptions are where hallucinations come from.

Claude found the folder and built it out: a full structure plus a README we started calling the Constitution. That README tells any AI that opens the vault how to use it. Folder structure, naming conventions, front-matter rules, and one line that does a lot of heavy lifting: never invent missing information.

Step 3: Load your own business context

We used voice dictation for this part. We said, roughly: "My name is [broker], I work at [firm], license number is X, I sell apartment buildings in Los Angeles. Update the company profile in the vault." Claude checked for existing notes, found none, and created clean company and personal profile files.

The more context you give it here, the better every future answer gets. This is the difference between an AI that guesses and one that actually knows your business.

Step 4: Build the ingest skill

A skill is just a saved standard operating procedure the AI follows every time you trigger it. We built one called "broker vault ingest." The instruction: whenever we attach a document, paste info, forward an email, or drop in notes, read the vault's README first, figure out what the information is, and file it in the right place.

Claude loaded its own skill-creator skill to build ours. You do not need to know how skills are made. You describe what you want and it writes the skill for you. One tip we would give past-us: add "ask me clarifying questions before building" to the prompt. Fewer assumptions, better skill.

Step 5: Build the query skill

Same idea, other direction. We built "broker vault query" so that whenever we ask a question about the business, it reads the Constitution, then pulls the answer from the right notes. Ingest puts information in. Query takes it out. That is the entire system in two skills.

Step 6: Feed it a real deal

We dragged a real OM into a fresh chat, typed the ingest command, and hit send. Claude read the whole offering memorandum and came back with a filing plan before touching anything: add a new listing, add the two listing brokers as contacts, add the comps from the OM, add market notes, save the original attachment.

It asked for approval first. We approved. Then it filed everything, checked for duplicates, and connected the pieces in the knowledge graph. Six comps from one OM, each broker with license number, email, and phone. Then we ran a second OM the same way. No duplicates found, all new data, filed clean.

Step 7: Ask your own data a question

Here is the payoff. We opened the query skill and asked: "Using only the sales comps stored in our vault, what is the average sale price per unit?"

Claude read all eleven comp notes, noticed there was no price-per-unit field in the template, and calculated it itself by dividing price by units, matching the exact method the source OMs used. The answer came back at $240,798 across all eleven comps.

Then we pushed it: "Find me a broker in my database who focuses on South Los Angeles." It surfaced the right listing agent, with a caveat it flagged on its own, plus full contact info. That is a query no outside tool can run, because it is running on data only you own.

The full setup prompt, both skill files (ingest and query), and the vault templates we used to build this are posted inside the AI for CRE Collective, ready to copy. We have hundreds of step-by-step workflows with demo videos, copy-and-paste prompts, and downloadable skills inside the Collective, and we post two new workflows every week.

What this is: a portable, compounding memory layer for your business. Because it lives locally as plain markdown, it is not chained to one AI. Feed it two deals and it is useful. Feed it two hundred and you have a proprietary comp set and vendor database that no competitor can buy off the shelf.

What this isn't: frictionless for teams. Everything sits on one computer, so a colleague on a different machine cannot see it out of the box. There are real workarounds (storing the vault in Dropbox, Box, Google Drive, a GitHub repo, or Obsidian Sync), but that is another setup step. It also takes upfront discipline to build, and the AI will still stop to ask clarifying questions when a deal does not match your profile. That is a feature, not a bug, but it is not magic.

The part that surprised us most was the model independence. Everything is stored locally, so this vault is not really a "Claude" thing. If Claude went down tomorrow, we could open ChatGPT or any tool with file access and it would read the exact same vault with the exact same context. You are building infrastructure, not renting it. And watching the knowledge graph fill in as each deal connects to its brokers and comps is genuinely satisfying.

Download Obsidian, create one empty vault, and hand Claude Code the setup prompt. Start with a single deal. Get Obsidian here.

Using AI to Find & Screen Land for Development Workshop
Wednesday, August 5
9:00 AM PT / 12:00 PM ET
$199 (free for Collective members)
Use AI and public parcel data to find, screen, and pursue development sites across residential, industrial, and retail. You leave with two ready-to-use take-home resources.

Jake Heller & Quinn Edwards.

We run the AI for CRE Collective: where 775+ CRE professionals learn to implement AI, with hundreds of video walkthroughs, copy-and-paste prompts, downloadable skills, 700+ AI tools in the database, and so much more.

If you'd like more hands-on help, we work with enterprise and institutional clients helping them learn, utilize, and implement AI into their business.

Jake & Quinn | AI for CRE Collective