A few people in class our zoom meeting wanted the real walkthrough, not the summary version. So here's the whole arc: how I started doing this myself with Claude, how I asked it to turn what we learned into a system, how that system is now running itself twice a day, and — the part I think matters most — how it caught its own mistakes without me having to be in the room. Phase 1: Doing it manually, market by market I started by just working with Claude live, city by city, following the "Zompz" deal-finding process — filter Zillow for stale, low-competition, needs-work houses, screen them, compute an offer (owner-finance under $80k, cash at 20% of value above that), find the listing agent's real published email, draft the offer, log it. Cleveland and Toledo went first. Along the way I caught and fixed a stale price-band reference the process had been carrying since an earlier city. Then I moved to Memphis, and it got messy in an instructive way. Memphis had way more inventory than anything we'd worked before, but roughly 70% of the cheap listings were "Contact agent"-gated on Zillow — no name, no phone, nothing published anywhere. Claude stopped and asked me how to handle it instead of guessing: "Memphis has visible agent contact on only ~30% of its cheap listings (rest are Zillow-gated). How do you want to handle it?" I told it to try to unlock the gated ones. It confirmed some were genuine dead ends and kept chasing the ones with a visible name. A few listings later, contact-hunting was clearly costing a lot of effort per lead, so it flagged that tradeoff too and asked how I wanted to proceed. I told it to send the workable ones and top up with a new city later. Phase 2: Turning the manual process into a real system Once I saw how much repeatable judgment was going into this, I asked Claude directly: "is it possible to make a MD file, I can have for this to be done in a claude project that can pull these list and have emails ready for me to review and send and pretty much automate the whole process"