What I’m thinking about: How a custom AI skill reads the soil reports, wetland studies, and engineering files I can barely parse myself…and surfaces the questions that decide whether a deal lives or dies.
A 15-page soil report lands in the inbox, some of it handwritten, half of it scanned at a resolution that would make your eyes water.
I barely read these things.
Unless you spent years as a civil engineer, you probably don’t either. A handful of operators can read borings and percolation tables in their sleep. Most of us cannot…myself included.
For a while now, AI has been paired into basically everything we do at SLC. On the deal side, we’ve talked before about SLC Chat (our underwriting brain, trained on our own deal data) and the Cowork plugin we’re building to tactically pull (accurate) comps, and then feed it into SLC Chat’s underwriting functions (recorded a pod about the progress last week, you can read about it next newsletter).
But comping is only one slice of due diligence. The bigger, higher-value deals coming our way routinely carry a mountain of other complexities and associated documentation. Earlier this year I plugged AI into the stack of paperwork on a complex entitlement deal out west, and it helped me suss out the follow-on questions we actually needed to ask (and build the correct legal structure).
This latest ‘deal stress test’ Claude skill (detailed below) is the evolution of that process, and is getting woven into the SLC Chat Cowork plugin.
The skill I built to stress-test deals I don’t fully understand
The move is simple: when a back-and-forth with Claude gets me a genuinely useful result, I don’t want to run that whole conversation from scratch every time. Instead, I have Claude turn it into a repeatable skill (validating that useful result is critical, otherwise you’ll just compound poor outputs).
First time it clicked: someone sent us a triple-net retail coffee shop land deal (triple-net = the tenant covers taxes, insurance, and maintenance, so the owner mostly just collects rent). Well outside our usual buy-box…the old me tosses it on sight (or subconsciously creates an excuse why it’s not worth pursuing).
Instead, I handed Claude everything the operator sent and asked a slew of underwriting-related questions.
Within ~15min, I had a read on whether it was even worth a deeper look (unfortunately the price and anticipated return did not match the risk required, but I now have a MUCH deeper understanding of what to look for on these deals).
So I had Claude build that into a repeatable Skill I named the deal stress test. The setup:
- A local folder with all the diligence (e.g. engineering reports, the proforma, CMAs, PDFs of the email threads)
- Our conservative underwriting standards baked in
- The AI instructed to play whatever expert the deal calls for (e.g. triple-net, infill, subdivisions, you name it), and to poke as many holes as humanly possible
(I’m simplifying the details, the language and logic backing the Skill are far more…you guessed it…nuanced).
Out comes a memo of red flags, follow-up questions, and exactly what to request next. And it adapts, like sometimes I just want to know what a soil test means, so I don’t even load the purchase price (telling the AI to ignore a full analysis and focus on the specific request).
Once the stress test builds the memo, I still have to send the follow-up to the operator. If you just tell the AI “write the follow-up email,” it will CERTAINLY come back sounding like a robot (a wall of em dashes that reads nothing like me). So I stacked a second skill on top: a Chris Duff Voice skill, trained on my own emails and newsletters, to turn that memo into a send-ready note in my voice.
By no means perfect (AI writing tends to run long, so I tighten it), but much faster than scrubbing generic AI prose into something human.
A $250K infill deal, three parcels, and a seller in a hurry
To further demonstrate the power of this Claude skill (backed by highly trained human validation…which is the most important ingredient, don’t forget), consider this live deal:
Three contiguous parcels, roughly an acre each, in the Midwest, and specifically one of the hottest markets in the country right now. $250K all-in. Off-market, marketed with a yard sign, sourced by an operator who lives locally. There’s a luxury condo going up across the street that apparently had its own eye on this same dirt (but lost the seller’s trust and took itself out of the running).
As a reminder, infill lots have burned us (badly) before, so our threshold to fund them runs higher than usual. The comps point to near-zero chance of losing money, which is always our first test (remember, you profit on the BUY, not the sell, and downside protection comes before upside).
The 2X gross margin we like to target probably isn’t here. Base case, we gross somewhere around $390-400K on the $250K buy across all three parcels. Typically, we would balk at that margin, but the downside protection and speed of the market influenced the equation.
And perhaps the biggest wrench was that the seller was playing hardball, originally wanting to close in about 9 days after we received the deal.
What the AI caught that I would have skimmed past
On infill lots, deal viability often comes down to utilities. Get it wrong and the value of the lot, or whether you can build on it at all, can swing dramatically. The operator had soil reports and a wetland delineation study (the aforementioned 15 pages of blurry, hand-written docs).
Instead of groaning in exasperation, I built a folder with the docs, dropped it into the deal stress test skill, and gave it the situation: fast timeline, tell me the utility flags, what else should we be asking, and what could REALLY go wrong if we’re missing something. Then I let the most powerful publicly available Claude model (Opus 4.8 on max effort mode) go to work. The depth was remarkable.
Some of what it surfaced:
- Two of the three lots reportedly had city sewer, per the operator, who had spoken with the county. But the soil borings showed the water table at roughly 2 to 4 feet below the surface. That makes those same two lots a potentially expensive problem for an end buyer (requiring an engineered septic system if sewer doesn’t pan out, or just flat out unbuildable). Plus, in a Midwest market where most buyers expect basements, 2 to 4 feet to water likely nixes that option.
- Claude’s flag: get written confirmation of sewer for those lots. It ran its own web search, found a sewer main running down the road, and named the open question…are lateral connections actually available, at what cost, and can you even tie in along that stretch? No written confirmation, probably no deal.
- The third lot, the one we assumed would need septic, actually had the best soil and the deepest water table. The operator quoted around $10-15K for septic there.
- Claude’s flag: regionally, that number could run 2 to 3 times higher. Verify it before you trust it.
- Eight soil borings were supposed to have been done. Six showed up…a gap I would almost certainly not have caught trying to read all 15 pages line by line, and Claude flagged it instantly.
- The wetland delineation was dated 2019. The AI worked out that locally, anything older than five years is stale, so we’d need a fresh one regardless. That’s a cost and a timeline we now have to price in, as end buyers would expect it, given the shallow water table levels.
- Confirming setback and buffer limits for building near wetlands. We hit this on a NJ deal once where you couldn’t build within roughly 100-plus feet of the line. On infill especially, you’d better be certain you have room to build with the correct setbacks.
- Minimum lot size for septic (rule of thumb: over a half acre, sometimes over a full acre, depending on the area).
This is the point where a lot of people slip, when they take the AI’s word faster than they should (as I detailed last week).
The work Claude did on the above bullet points all checked out, but it also took the liberty of sketching out where the wetlands sat per the 2019 delineation report.
However, when I eyeballed the original reports sent to us to see how AI had arrived at its conclusions, I could barely make heads or tails of the wetlands report given poor resolution and difficulty orienting where the parcels actually sat (LandID also did not note any wetlands, so I couldn’t cross-reference).
So I pushed back on Claude: “Show me exactly how you sketched out the wetlands map, what evidence did you use?” It walked back through the reports…then admitted it wasn’t actually certain (Iit still refined its output to maintain confidence in the rough location of the wetlands, but deferred on precise buildability).
That exchange is the whole game. AI is a tireless analyst that will surface things a human skims right past. The output is a stack of leads to verify (what’s still uniquely human is the judgment to Decide and to Check…beware offloading those tasks to AI, especially high impact, irreversible items).
=====
The message back to the operator was straight. There’s no way we move a quarter million dollars in 9 days without answers to the above questions, and any serious buyer is going to ask the exact same ones. Push the seller for an extension. On an off-market deal being marketed with a yard sign, that’s a reasonable ask…finding another all-cash buyer at $250K, in this economy, could take them a lot longer.
I half-expected the operator to walk the moment they saw our list. Instead, they secured a 17-day extension (better than nothing, given a seller who is apparently a nightmare to deal with) and they’re still running down the utility answers right now.
Where it lands is still TBD. We might fund it, we might pass. Either way, Claude and the deal stress test skill did their job (augmented by my own review), which is to surface every flag before a single dollar goes out the door.
=====
This deal stress test skill is one piece of a bigger system. The Cowork plugin primarily handles comping and pricing, the deal stress test handles the messy diligence around it, my voice skill preps comm’s, and more workflows are coming on top of all of those.
The whole point is to make complex, high-cost deals simpler to assess and harder to get wrong (of critical importance as we navigate the most difficult real estate market in 100 years, by most measures), utilizing the foundation of the national underwriting we’ve already built across thousands of reviewed deals, and the $6.5M we’ve funded at 41% operating margins.
The forcing function behind all of it isn’t glamorous. I’ve got two young kids at home, including a one-month-old, and my bandwidth is at an all-time premium. When you’re forced to make the business simple, you finally build the systems you should have built all along (backed by tech that gets more capable by the day).
If you’re an experienced operator sitting on a deal that needs capital, including the messy, complicated ones that live outside the usual box, send it our way. We write checks from $50K+, we close 100% of the deals we commit to, and our underwriting is sharpened by every AI tool we’re building internally.
