What is evaluation-driven development for AI applications?

It's applying test-driven discipline to AI: define how you'll measure success before and while you build, not after. You start by writing an evaluation set and metrics that encode what 'good' means for the task, then iterate on prompts, retrieval, and models with that eval as the feedback loop — every change is scored, so you keep what improves the number and revert what doesn't. It turns prompt engineering from vibes-based tweaking into measurable progress, catches regressions when you upgrade a model, and gives you evidence to ship. In practice: build a golden dataset early, wire evals into CI, watch the metric distribution as you develop, and expand the set as you discover failure cases. The eval becomes the spec.