An opinion piece arguing against the trend of pushing more decision-making earlier (shift-left) in AI-assisted development through formal specs like Gherkin, Quint, or Alloy. The author contends that judgment is cheaper and more reliable when exercised on a working artifact rather than on paper, proposing a 'shift-right' approach: converse with AI to clarify intent, build quickly, decide by observing the real thing, then lock decisions into machine-readable specs afterward to prevent drift. Irreversible decisions (data deletion, billing, auth, public API/database structure) should still be decided upfront, while reversible ones (UI, wording) can be deferred using feature flags. References critiques of spec-driven development as 'the waterfall strikes back' from Thoughtworks and Marmelab writers.
Table of contents
With AI, Decide After You Build, Not BeforeJudgment Does Not Get CheaperPolicy-Driven DevelopmentDeciding on the Real Thing Is Cheaper Than Deciding on PaperBuild First, Think LaterShifting Left Puts Judgment Back Where Waterfall Had ItWrite the Formal Spec After You DecideWhat Still Gets Decided FirstClosingShare this post