Hand-crafting prompts can kick off AI features in an afternoon, but over time your plain-English instructions balloon into a brittle mess. Every bug report or odd output adds another “hot fix,” until prompts span dozens of lines of caveats, threats, and edge-case handling. Teams swap dozens of template fragments at runtime just to keep things working. And because those tweaks rely on the quirks of one model, you end up locked to it—Datadog saw GPT-4o1 still dominating real-world traffic this spring, even as newer versions roll out. The core issue is that natural language wasn’t built for precise, durable specifications. A clinical study showed the same questions posed in patient versus physician voices flip a model’s refusal rates from zero to ten. Harvard researchers found that just naming a user’s favorite NFL team changes how often the model dodges sensitive topics. To counter these unpredictable weight-based behaviors, prompt authors repeat rules—Claude’s coding prompt fires off seven tool-call warnings, and Fable’s leaked system instructions hammer a copyright rule six times. Each extra line makes your system more fragile and harder to port to GPT-5.x or any other successor. The cure lies in shifting from prose to metrics and automation. Treat desired outputs as test cases with pass/fail criteria, and build evaluations that score model responses. Then let algorithms search the prompt space—tools like DSPy and GEPA automate phrase selection against your tests. When you formalize behavior with typed specs and metrics, swapping in a faster or cheaper model takes hours, not weeks. The same way assembly languages gave way to compilers, manual prompt engineering should yield to measured, test-driven workflows.