Matthew Berman's vibe-coding loop is the disciplined version: spec, build one feature, test, fix, commit.

Berman counters the "just wing it" reputation with a workflow: write a hyper-specific spec, encode stack/patterns/constraints in your IDE's rules file, then loop build-one-feature → AI-writes-tests → run-full-suite → fix → commit. (Forward Future) He pairs it with a "Loop Library" of 26 practitioner recipes with explicit verify/stop tables. This is the same independently-verifiable-steps discipline OpenAI's Codex-Maxxing doc preaches, just at solo-builder scale. Vibe coding isn't the absence of discipline. The builders shipping real things have more structure, not less.