怎么把提示词写结构How to Structure a Prompt

分块写，别写成一整段散文：角色与背景一块、任务一块、约束一块、输出格式一块，模型对分块指令的遵循更稳。

系统提示放长期规则（语气、边界），用户消息放本次任务——两层各司其职，别把一次性要求塞进系统提示。

用标题或标签划界（如「背景：」「要求：」），长输入时把待处理原文与指令明确隔开，防止模型把材料当指令。

约束写成可检查的条目：「不超过 300 字」「面向零基础」「给出来源」，比「简短一点」「专业一点」可执行。

输出格式单独一块并给例子：一段样例输出胜过三段描述。

改提示词一次只改一处并对比效果，同时改多处就说不清是哪处起了作用。

Write in blocks, not one wall of prose: role and background in one block, the task in another, constraints in another, output format in another — models follow blocked instructions more reliably.

Put standing rules (tone, boundaries) in the system prompt and the current task in the user message — each layer does its job; don't stuff one-off requests into the system prompt.

Mark sections with headings or tags (e.g. "Background:", "Requirements:"), and on long inputs clearly separate the material to process from the instructions, so the model doesn't treat data as directives.

Write constraints as checkable items: "under 300 words", "for absolute beginners", "cite sources" — these beat "shorter" or "more professional".

Give the output format its own block plus one worked example: a single sample output beats three paragraphs describing it.

Change one thing at a time when editing a prompt and compare results; change five at once and you can't tell which one worked.