Make evaluation a repeatable loop, not a vibe check

An evaluation-driven workflow — Define, Test, Diagnose, Fix — turns stochastic LLM output into an engineering loop, anchored by a "Minimum Viable Evaluation Suite" tiered for plain apps, RAG, and agentic tool use. Counterintuitive finding: "better" prompts can hurt without an eval set to catch regressions.

---

Sources: arXiv: When "Better" Prompts Hurt · Empirical study of prompting techniques for SE tasks