Zero Framework Cognition

The rule got its name from the culprit. After months of watching coding agents smuggle heuristic shortcuts into applications whose entire purpose was to delegate thinking to a model, Steve had what he calls The Talk with one of them (probably Claude 3.7 Opus) and asked, in his own transcription, "wth man, why u always do this." Claude supplied the acronym — "significantly snappier," Steve admits, than his hours of whining — and the two wrote the principles into his onboarding docs together. That irony powers the whole piece: the models themselves are the ones who refuse to send decisions to a model, and Steve caught a fresh violation in his own code while writing the post about it.

Half the essay is directly liftable: the actual ZFC section from Steve's AGENTS.md files, an allowed list (plumbing, schema checks, budget caps) and a forbidden one (anything that ranks, routes, or judges), ready to paste into your own agent instructions. Around that he traces the idea's ancestry through Fowler and Lewis's "smart endpoints and dumb pipes" and Unix's old separation of mechanism from policy, with Karpathy's Software 2.0 standing as the prophecy it partly fulfills, then spends the last chapter on the pattern's real weakness, cost, and the routing scheme that answers it: send every task down a pyramid of models to the cheapest one able to handle it. Steve is candid about the boundaries. The apps that taught him ZFC all live in the highly verifiable world of code generation, and he concedes he doesn't know whether it holds anywhere else. His one-line argument that it does is a running gag, better met in place.