The Specification Crisis

Fred Brooks argued in 1986 that no single technological development would produce an order-of-magnitude improvement in software productivity because the essential complexity of software—understanding the problem, specifying requirements, designing the architecture—would resist any tool that addressed only the accidental difficulty of implementation. AI has now produced the order-of-magnitude implementation improvement and confirmed, with unprecedented clarity, that Brooks was right about everything else. The specification crisis is the condition this confirmation produces: the accidental difficulty of writing code has collapsed toward zero, exposing the specification problem in its full, formidable, entirely untouched dimensions. Requirements are not sitting somewhere in users’ heads waiting to be extracted; they are negotiated across contradictory stakeholder interests, incomplete domain understanding, and the gap between what users say they want and what they will actually use. The hospital administrator needs a system that handles patient records, medication schedules, insurance rules, and clinical protocols—but she does not know what is possible, the