The Advantage of Micro-Modular Architecture May Be Further Amplified in Vibe Coding

Abstract

We argue that, for creative projects that continuously evolve under Vibe Coding, module- level spec-like documentation (maintained in real time by the Agent and kept fresh as the code evolves), together with atomic micro-modules and module- level verification, may support an Agent in making repeated, safe changes better than existing retrieval, indexing, or static documentation. We do not claim that “modularity is beneficial in every setting”—evidence shows the granularity–performance relationship is non-monotonic, static decomposition can even be harmful, and LLMs are in fact “neutral” about modularity. We restrict our claim to the specific, overlooked setting of “creative projects + repeated Agent modification + module-level spec-like freshness.” We use the concept-proof tool Yanxi (a salt-extraction tool) as a testable prototype and give five focused research hypotheses (H1–H5). The core value lies in identifying this overlooked setting and converging the evidence through an H1–H5 testable agenda.

Keywords