📄 中文摘要
Chris Laub进行了一项实验，旨在改变对模型选择的看法。他分别使用五种主要的语言模型（LLM）构建了同一个应用，并测试了五种不同的提示格式。实验结果显示，各模型的最佳得分和提示格式各不相同。其中，Claude模型获得最高得分87，最佳提示格式为XML；GPT-4得分71，最佳格式为Markdown；而Grok、Gemini和DeepSeek模型的得分较低，且没有最佳提示格式。该实验强调了提示结构在模型性能中的重要性。
模型不是瓶颈——你的提示结构才是
出处: The Model Isn't the Bottleneck — Your Prompt Structure Is
发布: 2026年3月31日
📄 English Summary
The Model Isn't the Bottleneck — Your Prompt Structure Is
An experiment conducted by Chris Laub aims to change the perception of model selection in AI. He built the same application using five major language models (LLMs) and tested five different prompt formatting styles across all of them. The results revealed varying best scores and prompt formats for each model. Claude achieved the highest score of 87 with XML as the best format, while GPT-4 scored 71 with Markdown. Grok, Gemini, and DeepSeek had lower scores without a specified best format. This experiment highlights the significance of prompt structure in determining model performance.