Prompts Aren't Real

A practicing engineer argues that obsessing over handcrafted prompts is the wrong approach for reliable agentic systems. Large language models are impressively capable yet unpredictably brittle in production: structured-output schemas usually work but sometimes fail catastrophically (a real fix was as odd as renaming a field from “title” to “heading”). Prompts behave differently when moved into new context, interact with existing agent instructions, and drift over time or across model updates, so prompt ownership by a single team and manual tuning do not scale. Safety, voice consistency, tool-calling behavior, and resistance to social-engineering-style instructions all demand repeatable guarantees that informal prompt-writing cannot provide.

The practical alternative is engineering rigorous, interlocking evaluation and optimization pipelines: generate adversarial and benign scenarios, express them as pass^k tests, and run them with and without new skills. Use LLMs to draft adversarial cases and to form graders for deterministic checks or complex judgments (brand voice requires its own labeled judge). Hook tests to an optimizer (example: genetic Pareto/GEPA) that iteratively mutates prompts with LLM guidance, then validate with holdout tests to prevent overfitting. Combine these rigs with production monitoring. The result is measurement-driven prompt artifacts that may look inscrutable but deliver robust, testable agent behavior at scale.