"Be concise" is the same underspecified trap as "act as a senior developer," just smaller
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
OpenAI 相关模型动态已经出现，适合跟踪能力变化、生态影响和后续可用性。
“简洁”这一指令常常不够明确，导致模型输出的内容虽然简短，但却可能遗漏关键细节或保留冗余信息。一位用户发现，最初尝试让模型“简洁”或“保持在100字以内”时，模型要么生成了难以理解的密集文本，要么删除了重要信息。最终有效的解决方案是提供一个具体的规则：“先给出结论，然后只包含如果删除会改变结论的推理。”这表明，像“简洁”或“像高级开发人员一样行事”这类模糊的指令，需要清晰、可操作的规则才能达到预期的效果。
Kept telling a model to "be concise" and kept getting output that was short but still somehow missed the point, cutting detail that mattered while keeping filler that didn't. Took three rounds of adjusting the instruction to actually get what I wanted.
First attempt: "be concise." Model's idea of concise turned out to mean shorter sentences, not less information, so the output was dense and harder to parse, not actually more useful. Second attempt: "keep it under 100 words." Better, but now it was cutting the wrong things to hit the number, sometimes losing the one detail that mattered to stay under the limit. Third attempt: "state the conclusion first, then only include reasoning that would change the conclusion if removed." That one actually worked, because it stopped being a length instruction and became a rule for what qualifies as necessary.
"Concise" is a persona-shaped word the same way "senior developer" is. It names an outcome without specifying the actual decision rule that produces it, so the model has to guess what "concise" means for this specific case, and it'll guess differently depending on the request. The fix in both cases ends up being the same move: stop describing the vibe you want and start describing the rule that would actually produce it.