📄 中文摘要
在销售、支持和金融科技的工作流程中，团队依赖提示来分类对话、提取信号和做出决策。尽管熟练的提示工程师能够使100个示例看起来完美，但这实际上是一个问题。更高的技能水平使得过程变得更加危险，因为直觉在小样本上有效，但无法推广到更大规模的输入、多个失败模式以及未测量的成本约束。专家的直觉虽然能产生看似合理的提示，但这些提示无法可靠地重现、版本控制或用指标进行辩护。解决方案并不是提升直觉，而是用客观函数来替代直觉。
优化提示，而非猜测提示
出处: Prompt Optimization, Not Prompt Guessing
发布: 2026年2月12日
📄 English Summary
Prompt Optimization, Not Prompt Guessing
In sales, support, and fintech workflows, teams depend on prompts to classify conversations, extract signals, and make decisions. While a skilled prompt engineer can make 100 examples appear perfect, this presents a significant problem. The more skilled one is at writing prompts, the more perilous the process becomes, as intuition works well with small samples but fails to generalize to larger inputs, multiple failure modes, and unmeasured cost constraints. Expert intuition produces prompts that seem right but cannot be reliably reproduced, versioned, or defended with metrics. The solution is not to enhance intuition but to replace it with an objective function.