Before Blind AI Adoption Destroys Engineering Organizations: Four Real Bullets and the Four-Dimensional Matching Framework

Brooks (1987) argued that no single technique would deliver an order-of-magnitude im-provement in software productivity. The community accepted this thesis but overlookedthe other half of Brooks’s argument: that cumulative, context-sensitive advances couldcollectively achieve what no single technique could. This paper develops that overlookedhalf. It identifies four genuine advances that emerged after 1986 — Extreme Programming(XP), Scrum, DevOps, and Human-Centered Design (HCD) — each addressing a distinctdimension of software development’s essential difficulty: XP provides technical practicesfor making code change safe, built on the xUnit testing foundation that did not existin Brooks’s era; Scrum operationalizes Nonaka and Takeuchi’s SECI model as the mostwidely practiced knowledge creation framework in any domain; DevOps minimizes thefriction of releasing for perpetually evolving SaaS systems; and HCD distills requirementsengineering into value verification. Each is a real bullet, but each works only againsta specific beast. The paper defines a four-dimensional matching space — environment,novelty, risk tolerance, and capability — that determines which practices are appropriate,at what intensity, and whether the organization can execute them. Capability is decom-posed into organizational and individual layers, with individual capability further dividedinto skill, execution experience, and knowledge-transfer aptitude. The analysis draws onover fifteen years of engineering leadership practice in which all four practices were im-plemented across multiple organizations. Original theoretical contributions include: thecharacterization of TDD as cognitive externalization grounded in Ericsson’s deliberatepractice theory, with boundary analysis as a precondition; a categorical boundary be-tween precise refactoring (zero failure probability, requiring static typing and tooling) andrewriting (nonzero probability, compounding multiplicatively), which reinterprets Ernstet al.’s (2015) finding on architectural technical debt as a structural capability deficit;a three-layer testing taxonomy (construction scaffolding, functional verification, qualityengineering); and the identification of label collapse — the systematic conflation of qual-itatively incommensurable activities under identical practice labels — as the mechanismthat renders self-report-based research structurally unable to measure what it purportsto measure. The urgency of this articulation is temporal: as AI coding tools enter thefashion cycle, the knowledge structure that the field built over four decades risks beinglost before it has been explicitly formulated.