Ask HN：您最喜欢的用于改进 LLM 输出的提示是什么？
1 分•作者: maxutility•13 天前
我经常使用 Claude Code 和 GPT 5.5，发现它们在提供极大帮助的同时，也常常陷入一些常见的性能低谷。例如，在写作方面——这也许是我对它们最大的不满——它们会表现出一些有明显标记的写作癖好（比如使用破折号，或者“不是 x 而是 y”的句式），一些不太明显的写作癖好（例如“诚实的表述”、“应得的 XYZ”），晦涩难懂的自创术语，过度缩写，以及用箭头连接的句子（“这个 -> 那个 -> 另一个东西”），它们还常常不理解受众对分析的熟悉程度，使用宏大的表述方式，以及以一种有动机的直观方式来解释事物。
另一个常见的陷阱是过度设计分析解决方案，或者在我指出之前，它们未能考虑局限性或常见的故障模式。
我很想听听你们遇到的其他故障模式，以及你们发现有用的提示（系统提示或其他）、技能和其他技巧，以获得这些工具更好的性能和更易于使用的输出。
查看原文
I use Claude Code a lot and GPT 5.5 as well, and find that they are simultaneously extremely useful and also fall into common poor-performance basins. For example, writing performance -- perhaps my biggest issue with them is writing style -- such as well documented stylistic tics (em-dash, it's not x it's y), less commented stylistic tics ("the honest framing", "the earned XYZ"), cryptic and coined jargon, overabbreviation and sentences replaces by arrow constructions (this -> that -> this other thing), poor understanding of the audience's familiarity with the analysis, grandiose framings, as well as explaining things in a motivated intuitive way.<p>Another common trap is overengineering analytical solutions, or failing to consider limitations or common failure modes unless I point them out.<p>Would love to hear what other failure modes you've navigated, and see the prompts (system or otherwise), skills, and other techniques that you all have found helpful to get better performance and more usable output out of these tools.