The rapid adoption of AI coding assistants has outpaced systematic understanding of how the collaboration is structured. Most productivity research treats model access as the independent variable; the workflow architecture surrounding that access — the configuration, memory, invocation protocols, and feedback mechanisms that determine how effectively a practitioner and model collaborate — has received almost no attention. This paper addresses that gap. We describe a four-layer workflow architecture developed over eight weeks of daily AI-assisted enterprise SQL development at DataBank Holdings, Ltd., and propose the output-to-input token ratio as a longitudinal efficiency metric for individual practitioners. The architecture's layers — project configuration, cross-session memory, domain skill modules, and structured invocation protocols — function collectively as correction loop suppressors, increasing the leverage of each human input token. A fifth layer introduces a paradigm not previously documented in the literature: the AI system is directed to analyze, critique, and improve the human-AI interaction itself, turning the model's analytical capacity back on the collaboration in real time. The result is a system that compounds rather than plateaus. We offer the token efficiency ratios produced by this architecture as existence proof: a specific, documented operating point achieved under precisely identified conditions, offered not as a universal benchmark but as evidence that the conditions exist and are worth replicating. The measurement framework we propose — longitudinal ratio tracking by individual practitioners — is offered as a contribution independent of our specific implementation. The ideas, methodology, data, analysis, and editorial judgment are solely those of the author; the writing craft was substantially aided by Claude (Anthropic), an AI system. The author assumes full responsibility for every scientific assertion, explicit or implicit, in this paper.
Stephanie Anne Livingston (Wed,) studied this question.