A Claude skill that pressure-tests whether to buy at all, then finds the best real value and recommends. Advisory only: it never buys. I make every purchase myself.
It's a skill rather than a scheduled agent, so it applies anywhere in Claude the moment I'm considering a purchase. For anything over $250 it starts with a premise check: is this needed, why now, and what happens if I wait or do nothing. It revisits those questions when the conversation starts building momentum toward a specific option, and says plainly if the path no longer makes sense. Under $250 it's a one-line sanity check.
If the premise holds, it defines value as lifetime usefulness divided by total cost of ownership (price, consumables, maintenance, time, resale), researches three tiers (a value pick, a sweet spot, and a buy-it-once option) from independent long-term testing and owner reports, checks price history, refurbished and used options, model cycles and warranty, and then recommends. The recommendation leads with the verdict: buy, wait, or don't. Then the top pick and runner-up, a target price, what would change the call, and any rationalizations it caught me making. It gives me links. I click buy, or I don't.
Purchases come up ad hoc, and most of the work of a good one is deciding whether to make it at all. I wanted the research without the momentum: a second opinion that's allowed to say "wait," built into the tool I was already going to ask. And it had to be structurally incapable of spending money, because the useful version of an assistant is one I can trust with more each month, and that trust starts with the card staying in my wallet.
The premise check does most of the work. Of the briefs it has produced so far, the useful ones were less about which model and more about whether, and when. A rotating car seat before a second child: yes, and here's the one. LinkedIn Premium during a job search: here's exactly what you'd lose by cancelling, and whether that matters this month.
Rules you write for an agent are rules for yourself. Writing down where I'll happily overspend and where I refuse to turned out to be the hard part, and it's still open. The skill can't be more honest about value than I am.
In use since October 6, 2026. Open questions I owe it: where I'm happy to overspend (running, skiing, scuba, kitchen gear?), where I refuse to pay for the top end, and whether used and refurbished should be the default or only for some categories. A related question in the Concert Radar book: whether a pair of tickets over $250 should trigger the same premise check.