Seven rules for moving cheap decisions off the big model
Seven practical rules: send state as JSON, ask many questions in one call, write the levels yourself, keep exact math in code, use thresholds for the final call, set a confidence floor, and offer only valid actions.
This is f*cking insane. Feed Jev like a chatbot and you waste it. These seven rules move the cheap decisions off your big model: → send state as JSON. it takes JSON directly, no need to flatten your data into prose → ask many questions in one call. DiffJury asks nine, killmyidea asks ten → write the levels yourself. killmyidea gives every question five levels, each one a concrete situation → keep exact math in your code. typesafe-mario computes the timing itself and Jev interprets the result → let code make the final call. killmyidea turns scores into kill, fix or ship with plain thresholds → set a confidence floor. anything below it goes back to the big model → offer only valid actions. ultrafast shows Jev just the permitted operations, so it can't answer with a command you never allowed if every fork in your agent still goes to a frontier model, you're paying frontier prices for multiple choice before you move a decision over, ask three things: → can every valid answer be listed in advance? → would a person settle it at a glance? → does it happen often enough to add up? three yeses go to Jev. one no stays on the big model tonight: take one fork from your last agent session, write its options as a list, send the state as JSON in one call, and compare the answer with what the big model gave you
This post quotes another post: x.com/rvaniaaaa/status/2101709732314558639