跳转到主要内容

跨模型、工具、服务、公司、提供商、API、套餐、文章和新闻进行搜索。

提示工程入门 8 分钟

10 个好提示与坏提示对比:2026 年 ChatGPT 到底怎么问才有效

通过前后对比示例,展示上下文、角色、约束和输出格式如何把无用的回答变成精准可用的答案。

10 个好提示与坏提示对比:2026 年 ChatGPT 到底怎么问才有效

Why short prompts fail

"Write about X" is not an instruction — it is a coin flip. Without context the model has to guess everything: who the audience is, what the goal is, what is forbidden, and what shape the output should take. Every guess is another chance at a useless answer.

The four parts of a good prompt

Every effective prompt carries the same four parts:

  • Context: the topic, audience and goal — for whom, and for what purpose?
  • Role: put the model in the right seat ("technical editor", "beginner-friendly tutor").
  • Constraints: length, tone, what to avoid, allowed sources.
  • Format: list? table? a few paragraphs? Ask for the exact output shape up front.

A before-and-after example

Before: "Write something about email marketing."

After: "Write a welcome email for new customers of a sports-equipment store; friendly tone, max 120 words, ending with one clear call to action for a first purchase."

Same model — the second version simply removes the guessing.

Common mistakes

  • Cramming dozens of instructions into one prompt; the model drops half of them. Prioritise or split.
  • Giving no example when you want a subtle output; one good sample beats ten lines of description.
  • Accepting the first answer; a short follow-up ("make the second section simpler") usually moves results from mediocre to good.