A free playbook by The Aigent Lab


Underwriting & Analysis · Playbook 01

Underwrite any listing in under a minute.

Anthropic's Financial Services agents can run a full ten-year model on a live listing in about sixty seconds. Below: the install, the exact prompt to use with a client's target property, and four things the output still needs a real agent to catch.

i.Why It Matters ii.What You're Building 01.Install 02.Run The Prompt 03.Read The Output ★.A Real Run 04.What It Gets Wrong

Why this matters for your business.

A serious underwrite takes a competent analyst four to eight hours per deal. Most agents skip it, and most first-time investor clients never ask for one. They read the listing sheet, drive by the block, and make an offer on gut feel.

That worked when rates sat near zero and prices only moved one direction. It stopped working the day the Fed started hiking, and it has not started working again since. The agents who win investor clients now are the ones who can hand over a real model, not just a listing link.

Before this generation of AI agents, the alternative was a half day building the model yourself, or paying an analyst a few hundred dollars per deal. Now one prompt produces the model, the debt schedule, the ten-year cash flow, and the sensitivities in about sixty seconds, with your name on the work.

Decent-looking deals nobody stress-tested are the ones that damage your reputation with an investor client.

What you're building.

The End Result

A sixty-second deal screen

You paste any listing link into Claude. The Model Builder agent returns a full underwriting pack: every assumption, the debt schedule, the ten-year cash flow, the sensitivities, and a buy-or-pass read you can hand straight to a client.

What You Need

Fifteen minutes, once

Claude Pro at minimum, the desktop app, Co-Work enabled in settings, and one real listing URL. Setup only happens once; every future chat inherits the agent after that.


Install once. Use forever.

01

Open the Financial Services marketplace page

Copy the marketplace link from the browser.

02

Open Co-Work in your Claude desktop app

Click Customise, then Create Plugin, then Add Marketplace.

03

Paste the link and sign in

All ten finance agents become available in every future Co-Work chat: Model Builder, Valuation Reviewer, Market Researcher, Pitch Builder, and more.


Pick a real listing. Paste. Send.

Open a new Co-Work chat and use the prompt below on the next listing you or your client are seriously considering. The Model Builder agent asks four or five clarifying questions before it starts. Answer them; the ones you can't answer yet are the ones worth researching before an offer goes in.
The PromptCopy
Use the Financial Services plugin to underwrite the following property: [paste listing link]

Two files come back. Both editable.

.xlsx

Underwriting model

Every input lives on the Assumptions tab. Change the LTV, rent, or appreciation rate and the whole workbook recalculates: assumptions, debt schedule, ten-year cash flow, returns summary, and sensitivities.

.md

Client-ready summary

A one-page written read: the headline returns, what drives the outcome, where the deal is fragile, and a buy, pass, or push-the-seller call. Paste it straight into an email to your client.


The moment it earned trust.

The first listing run through this stack was a four-unit multifamily asking $1.85M, roughly nine percent under the block's recent comps. Standard conventional financing, seventy-five percent LTV, five-year hold.

The first model came back fast: an 18% Year 1 cash-on-cash return. That number looked too good to hand to a client. Before it could be flagged, Claude caught the inconsistency itself, mid-output.

Claude · self-audit

The Year 1 cash-on-cash figure is inconsistent with the sensitivity table. Debt service isn't being subtracted from the levered cash flow row. Let me trace the formula.

Claude · after the fix

Found it: the debt schedule referenced the wrong row, so debt service was reading as zero. Corrected. Base case now ties out at 14.1% levered IRR, 1.9x equity multiple over five years.

That sequence, an agent catching its own bug and re-verifying before handing off the answer, is the difference between a tool and something you can actually put your name behind in front of a client.

What it still gets wrong.

i.

Hyperlocal comps

Claude doesn't know one block is on the wrong side of a zoning line. Feed it real comps if you want them used.

ii.

New-construction risk

Strong on resale. For new builds, tell it which builder and which past projects ran late, and by how long.

iii.

Local tax and fee nuance

Transfer tax, HOA, closing cost norms. Sanity-check the Assumptions tab before you trust the return.

iv.

It will agree with you if you let it

Don't ask for a positive model. Ask for the bear case explicitly, and read that version twice before your client does.


Next Playbook

Build a deal-sourcing agent that watches the market for you.

Read Next ↗