From Vibe to Verifiable: Declarative Closeout Protocols for Human–AI Product Delivery

Vibe coding and agent “skills” packages make it easy to produce running software, but small product teams still lack verifiable criteria for declaring a requirement done—especially when success depends on runtime agent behavior, not only repository tests. This paper is a practitioner methodology report with instrumented transfer evidence, not a controlled experiment. We present a lightweight Loop model of human–AI–artifact collaboration and declarative closeout protocols that hard-block premature completion claims. Two protocol families are central: (i) delivery closeout (smoke tests, staged review, and explicit human override with cost accounting); and (ii) runtime friction measurement after a path “works,” using agent logs (tool-call budgets, forbidden bypasses, latency) until a friction index reaches zero. We report a longitudinal industrial case (RetailPilot, an anonymized e-commerce agent control plane) where loops emerged in practice and were later formalized, and we outline a transfer study via a portable starter kit and a second domain workbench (DomainWorkbench-PE, MVP reached; post-MVP auth/UX deepening ongoing). A sibling product (KnowledgeService, an anonymized knowledge/RAG service) adopted an even thinner seven-loop cut without YAML gatekeepers and now serves sample 2 as an ingested MCP data plane—evidence of graded portability, not a substitute for sample 2. We position related agent-skill methodologies as dev-time engineering workflows, complementary to—but distinct from—product delivery governance and runtime sign-off.