2026年7月16日 · Grilling · Decision · Prompt

别急着让 AI 干活，先让它把你问清楚

AI 很容易把"我还没想清楚"误判成"我需要一份完整方案"。你给它一个模糊计划，它马上开始写标题、列步骤、排优先级，看起来很有效率。问题是，如果上游判断错了，后面越完整，返工越贵。

我用 Grilling / grill-me 解决的就是这个问题：先不让 AI 执行，而是让它像一个严厉的面试官一样追问。

核心规则

规则很简单，一次只问一个问题，每个问题都必须给推荐答案，并且按依赖顺序往下问。

阶段	先问	再问
定位层	这个东西要服务谁？	受众是谁？
目标层	成功长什么样？	形式是什么？
风险层	最容易错在哪？	怎么执行？

为什么不用"帮我提建议"

这和普通"帮我提建议"不一样。普通建议经常会同时扔出十几个方向，最后你还是不知道先改哪个。

Grilling 的强约束是，一轮只解决一个分叉。

比如我拿一段实习经历问它，它不会先帮我润色，而是先问：这段经历到底要服务 AI 产品、Agent Eval、测试开发，还是售前解决方案？这个问题不回答，所有措辞都只是装饰。

我最常用的开场 prompt

请围绕这个计划持续追问，直到我们达成共同理解；一次只问一个问题；每个问题都给你的推荐答案；能从文件里查到的事实不要问我；在我确认之前不要执行。

这个 prompt 的关键不在"严厉"，而在"暂停执行"。

判断标准

如果某个问题的答案会改变后面 50% 以上的内容，它就应该先被问。

比如简历里先问岗位定位，博客里先问读者是谁，产品方案里先问成功标准，项目取舍里先问约束是什么。AI 如果跳过这些问题直接写正文，通常是在制造看起来完整的废稿。

适用边界

场景	是否适合 Grilling
方向没定、取舍很多、做错代价高	✅ 最适合
需求已经很明确	⚠️ 浪费时间，直接执行
改一个拼写错误 / 生成一条 SQL / 翻译	❌ 机械任务，不需要追问

换句话说，它不是提速工具，而是防止你高速跑偏的工具。

Before AI Does the Work, Make It Ask Better Questions

AI easily mistakes "I haven't figured this out yet" for "I need a complete plan." You give it a vague idea, and it immediately starts writing headlines, listing steps, and prioritizing—it looks productive. The problem is: if the upstream decision is wrong, the more complete the output, the more expensive the rework.

Grilling / grill-me is my answer to this: instead of letting AI execute, make it interrogate you like a tough interviewer first.

Core Rules

The rules are simple: ask one question at a time, every question must come with a recommended answer, and questions follow dependency order.

Stage	Ask First	Ask Next
Positioning	Who is this for?	What's the audience?
Goal	What does success look like?	What format should it take?
Risk	What's most likely to go wrong?	How will you execute?

Why Not "Give Me Suggestions"

This is different from generic "give me suggestions." Regular suggestions throw a dozen directions at you at once, and you still don't know which to address first.

Grilling's key constraint: one fork at a time.

For example, when I ask it about an internship experience, it doesn't start polishing the wording. Instead it asks: is this for AI products, Agent Evaluation, QA engineering, or pre-sales solutions? Until that question is answered, all wording is just decoration.

My Go-To Opening Prompt

Please keep asking questions about this plan until we reach shared understanding. One question at a time. Include your recommended answer for each question. Don't ask me about facts you can look up from files. Don't execute anything until I confirm.

The key to this prompt isn't the "tough" tone—it's "pause execution".

Judgment Criteria

If the answer to a question would change 50% or more of what comes after, it should be asked first.

In a resume: ask about target role first. In a blog post: ask about the reader first. In a product plan: ask about success criteria first. In a project trade-off: ask about constraints first. When AI skips these questions and jumps to writing, it's usually producing output that looks complete but is worthless.

Scope of Use

Scenario	Is Grilling suitable?
Direction unclear, many trade-offs, high cost of mistakes	✅ Ideal
Requirements are clear	⚠️ Wasted time—just execute
Fix a typo / generate a SQL query / translate	❌ Mechanical task, no questioning needed

In other words, it's not a speed tool—it's a tool to keep you from speeding in the wrong direction.