On the workbench,
right now.

Looking for a house in intown Atlanta with a tight set of criteria: 8+ rated public schools for elementary through high, 4+ bedrooms, 2.5+ baths, ≥2,250 sqft, lot ≥8,000 sqft, not on a busy street, costing $650K–$1.3M. The internet's home-search tools are built for browsing supply that exists today. They aren't built to answer how rare these homes are historically. So I built a database for that (with the help of Claude Code). Lessons so far:

• Define the universe before filtering - Rarity is a fraction of matching homes / total homes. If you never pinned down the denominator (the full universe), you have a numerator floating with nothing underneath it
• Check your estimates against real data - the property-tax estimate was 57% high until checked against one real bill
• Consider data biases — MLS-tied sources can't see the rarest-traded homes, which are exactly the ones I'm looking for

I wanted to know whether I was leaving money on the table by buying the same staples ad hoc across four stores instead of subscribing where it's cheapest. So I pulled three years of order history and put it all in one place. The honest finding: there's no big-ticket move to make. My spending is already reasonable across stores. Real savings land in the $10–35/year range. The decision isn't per-item price optimization; it's habit and convenience. Which store is already in my routine, what delivery cadence actually fits my week, what's reliably stocked when I need it? That's a more useful answer than I expected, even if it's less dramatic.

• Only Amazon offers a clean CSV export. For the rest I used Claude in Chrome to read the order history straight off the page.
• The scrape meant working around Walmart's press-and-hold CAPTCHAs and Claude Code Pro's context limits on heavy DOM pages (where order histories are deep and repetitive markup eats a context window fast.)
• Lesson: the analysis was the easy part. Getting clean, comparable data out of four stores that don't want you exporting it was the whole job.
• Lesson: 'Should I optimize this?' is worth asking even when the answer is no. Confirming the savings are small is itself the result and it frees me to choose on convenience without second-guessing.