asking-to-understand
Compare original and translation side by side
🇺🇸
Original
English🇨🇳
Translation
ChineseAsking To Understand
以提问促理解
A question the answerer can answer from memory taught nobody anything.
The session is one grounded question per turn, and it ends when they state
the rule in their own words - not when your question list runs out.
The failure this prevents: an agent asked to "ask questions" dumps eight at
once, none standing on anything it looked at, each with the answer already folded
into the wording. The answerer picks the easiest one, everyone feels productive,
nothing moves.
仅靠记忆就能回答的问题无法教会任何人任何东西。
会话需遵循每轮一个基于事实的问题规则，结束标志是用户能用自己的话阐述规则——而非你的问题清单耗尽。
**该准则避免的问题：**当Agent被要求“提问”时，一次性抛出8个问题，既无事实依据，又在表述中暗含答案。回答者只会选最简单的问题作答，所有人看似高效，实则毫无进展。
The session contract
会话约定
Four phases, in order. Phases 1-3 loop.
| Phase | ไทย | What it is |
|---|---|---|
| ⓪ | | Read the artifact and RUN the cheap checks - before the first question |
| ① | | Exactly one question per turn, then stop |
| ② | | Classify the answer; the class picks your move (table below) |
| ③ | | Every 3-4 questions, say out loud which two answers share a shape |
| ④ | | Park what is open, list what is concretely undone, hand the turn back |
分为四个阶段，按顺序执行。阶段1-3循环进行。
| 阶段 | 中文 | 说明 |
|---|---|---|
| ⓪ | | 先阅读工件并执行简单检查——在提出第一个问题之前 |
| ① | | 每轮仅提出一个问题，然后停止 |
| ② | | 对答案进行分类；分类结果决定你的下一步行动（见下表） |
| ③ | | 每提出3-4个问题后，明确指出哪两个答案存在共性 |
| ④ | | 记录未解决的问题，列出具体待完成事项，将主动权交回用户 |
⓪ Ground first - no question before evidence
⓪ 先夯实基础——无依据不提问
Spend 3-8 read-only checks before opening your mouth. The minimum bar:
verify that what the artifact CLAIMS exists actually exists.
That one check alone - , , , a REST GET -
routinely finds the gap the whole session then turns on. A document is not
evidence of the thing it describes.
git ls-treelsgit check-ignoreEvery question you ask must stand on something you looked at. A question with no
measurement behind it is an opinion in disguise, and the answerer can smell it -
they start answering defensively instead of thinking.
开口提问前，先花3-8分钟执行只读检查。最低要求：
验证工件声称存在的内容确实存在。
仅这一项检查——、、、REST GET请求——通常就能发现整个会话需聚焦的核心漏洞。文档不能证明其描述的事物真实存在。
git ls-treelsgit check-ignore你提出的每一个问题都必须基于实际查看的内容。没有事实支撑的问题本质是伪装的观点，回答者能察觉到这一点——他们会开始防御性作答，而非主动思考。
① One question per turn
① 每轮一个问题
Shape of a question turn:
<evidence: 3 lines max - the measured fact, with path:line, command output, or API response>
**คำถาม: <one sentence>**
<optional: what a complete answer must contain - "ตอบให้ครบสามชั้น: A, B, แล้ว C">One question. Then stop and wait. Silence after a question is the instrument,
not an awkward gap. Two questions in a turn means they answer the easy one and the
hard one dies unnoticed.
A parenthetical follow-up is allowed only when it is the same question one
layer deeper (), never a second topic.
คำถามซ้อน:提问轮次的结构：
<依据：最多3行——实测事实，包含路径:行号、命令输出或API响应>
**问题: <一句话表述>**
<可选：完整答案需包含的要素——“请从三个层面作答：A、B、C”>**一个问题。然后停止等待。**提问后的沉默是引导工具，而非尴尬的空白。一轮提出两个问题会导致回答者只答简单的那个，难的问题被悄然忽略。
仅当追问是同一问题的更深层延伸时（），才允许使用括号式跟进，绝不能切换到第二个主题。
追问:Choosing which question - the ladder
问题选择优先级
Ask the one highest on this ladder:
- They cannot answer it from memory - it forces them to simulate the system or go look. "Stage 3 fails after stage 2 deployed - what state is production in?" beats "do you think this pipeline is good?"
- The answer changes what they do tomorrow - prefer the shallow-but-live question over the deep-but-theoretical one, early.
- You do not know the answer either - both sides have real evidence, so any answer is usable.
优先选择符合以下优先级的问题：
- 无法凭记忆回答——迫使他们模拟系统运行或亲自查证。例如：“阶段2部署后阶段3失败——生产环境处于什么状态？” 优于 “你觉得这条流水线好用吗？”
- 答案会影响他们明天的行动——初期优先选择浅显但实用的问题，而非深奥但理论化的问题。
- 你也不知道答案——双方都有真实依据，任何答案都有价值。
The admission test - run it on every question before sending
准入测试——发送前对每个问题执行
If they answer this perfectly, is the question still worth having asked?
- No - it is not a question, it is a trap. Delete it.
- Yes - send it. Either way, both sides leave with something.
如果他们完美作答这个问题，提问仍有意义吗？
- 无意义——这不是问题，是陷阱。删除它。
- 有意义——发送。无论哪种情况，双方都能有所收获。
② Receiving the answer - the class picks your move
② 接收答案——分类决定行动
| Answer class | Your move |
|---|---|
| ถูกหมด | Name what it unlocks, go to the next question |
| ถูกครึ่งเดียว | State which half is right and supply the missing half as fact - do not re-ask |
| อาการ ไม่ใช่กลไก ("we forgot", "AI slipped") | Push once on the same question: "นั่นคืออาการ - กลไกคืออะไร" |
| ผิด | Give the measured fact. Verdict on the artifact, never on the person |
| ไม่รู้ / เลี่ยง | Answer it yourself, then move on. Never ask the same question twice |
Symptom vs mechanism is the move that pays. A symptom answer produces "next
time be more careful", which fixes nothing. One more push produces a rule that
outlives the session. This is the step most often skipped, because pushing feels
like nitpicking - it is not, it is refusing to let the answer expire.
Half-right answers get a named half, not a "ใช่ครับ แต่". Telling someone
exactly which half they missed respects them more than a softened correction does.
| 答案分类 | 你的行动 |
|---|---|
| 完全正确 | 说明该答案解锁的内容，进入下一个问题 |
| 部分正确 | 明确指出正确的部分，并补充缺失的部分作为事实——不要重新提问 |
| 现象而非机制（如“我们忘了”“AI出错了”） | 针对同一问题追问一次：“这是现象——机制是什么？” |
| 错误 | 给出实测事实。对工件做出判断，而非针对个人 |
| 不知道/回避 | 自行回答该问题，然后继续。绝不要重复提问 |
**区分现象与机制是关键动作。**现象类答案只会得出“下次更小心”的结论，无法解决根本问题。多追问一次就能得到超越会话本身的规则。这一步最常被跳过，因为追问看似吹毛求疵——实则不然，这是拒绝让答案失效。
**部分正确的答案要明确指出缺失部分，而非用“是的，但……”这种委婉表述。**准确告知对方遗漏的部分，比温和的修正更尊重他们。
③ Connect - this is the payload
③ 关联要点——这是核心价值
Every 3-4 questions, look back across the answers and ask yourself whether two of
them are the same bug in different clothes. When they are, say so explicitly.
That is the thing a checklist cannot do and the reason a human is asking. A
session that never connects two answers was a quiz, not a grilling.
每提出3-4个问题后，回顾所有答案，判断是否有两个答案是同一问题的不同表现形式。如果是，明确指出这一点。
这是清单无法完成的任务，也是需要人类提问的原因。从未关联过答案的会话只是一场测验，而非真正的引导式提问。
④ Closing
④ 收尾
Stop when they state a rule in their own words. Questions after that point are a
performance. Then emit exactly:
**คำถามที่ยังค้าง** - by name, with the question restated in one line each
**ของจริงที่ค้าง** - the concrete measured facts that still need action
→ ถามต่อ หรือ ลงมือเก็บของจริงก่อน?Parking a question out loud () is part of the
discipline - it keeps a thread alive without forcing it closed early.
ขอค้างข้อ 5 ไว้ก่อน เดี๋ยวกลับมา当用户能用自己的话阐述规则时停止提问。之后的问题只是形式主义。然后输出以下内容：
**未解决问题**——列出问题名称，每个问题用一句话重述
**待处理事项**——仍需行动的具体实测事实
→ 继续提问，还是先处理待办事项？公开记录未解决问题（如“先暂停第5个问题，稍后再回来讨论”）是准则的一部分——既能保留讨论线索，又不会过早强行收尾。
Language
语言规范
Mirror the language the user asked in, including the questions themselves. Keep
technical terms in English with a short gloss on first use. If they ask in Thai,
ask in Thai.
镜像用户提问时使用的语言，包括问题本身。专业术语保留英文，首次出现时附带简短解释。如果用户用泰语提问，就用泰语提问。
Worked example - one turn
实操示例——一轮会话
Artifact: a decision map an AI produced. Ground phase found the deliverable file is not where the map says it is.
ไฟล์มีจริง 308 บรรทัด แต่อยู่ที่ tmp/DEPLOY.md ไม่ใช่ root ของ repo โค้ด
และ tmp/ ถูก gitignore ไว้ (.gitignore บรรทัด 26) - ไม่เคยถูก commit
**คำถาม: คนที่คู่มือนี้เขียนถึง - dev คนอื่นที่ไม่ใช่คุณ - เขาจะเปิดไฟล์นี้เจอได้ยังไง?**
(ถ้าคำตอบคือ "ผมส่งให้" - คำถามคือคุณส่งเวอร์ชันไหน และอีกสามเดือนจะรู้ได้ยังไง
ว่าเวอร์ชันที่เขาถืออยู่ตรงกับของจริง)Answer comes back: "น่าจะอัพเดทไม่ครบ" - class อาการ ไม่ใช่กลไก - push once:
"อัพเดทไม่ครบ" คืออาการ ยังไม่ใช่กลไก - ถ้าหยุดตรงนี้ วิธีแก้จะกลายเป็น
"คราวหน้าอัพเดทให้ครบ" ซึ่งแปลว่าไม่ได้แก้อะไรเลย
**กลไกคืออะไร - อะไรคือความต่างเชิงโครงสร้างระหว่างคำอ้างสองแบบนี้?**工件：AI生成的决策流程图。基础阶段发现交付文件不在流程图标注的位置。
该文件确实存在，共308行，但位于tmp/DEPLOY.md而非代码仓库的根目录
且tmp/已被加入.gitignore（.gitignore第26行）——从未被提交
**问题：这份指南面向的对象——其他开发人员而非你——他们要如何找到这个文件？**
（如果答案是“我会发给他们”——接下来的问题是：你发送的是哪个版本？三个月后如何确认他们手中的版本与实际情况一致？）收到答案：“应该是更新不完整”——分类为现象而非机制——追问一次：
“更新不完整”是现象而非机制——如果就此止步，解决方案就会变成
“下次更新完整些”，这相当于没有解决任何问题
**机制是什么——这两种表述之间的结构性差异是什么？**Red flags - stop and re-cut
危险信号——立即暂停并调整
- Two questions in one turn
- The answer is already folded into the question's wording
- A question sent before opening a file or running a command
- - a correction wearing a compliment
"ใช่ครับ แต่..." - Re-asking a question because the answer disappointed you
- Still asking after they have stated the rule
- Four questions in and nothing has been connected to anything
All of these mean: fewer questions, more ground.
- 一轮提出两个问题
- 问题表述中暗含答案
- 未查看文件或执行命令就提出问题
- 使用“是的，但……”——披着赞美外衣的修正
- 因答案不符合预期而重复提问
- 用户已阐述规则后仍继续提问
- 提出4个问题后仍未关联任何要点
所有这些情况都意味着：减少提问，增加事实依据。
Handoffs
场景切换
| Situation | Go to |
|---|---|
| They want it explained, not to be asked | |
| Challenging a plan before it is built | |
| It has turned into reviewing the work itself | |
| What surfaced should become a durable lesson | |
| An interactive click-through would land it better | |
| 场景 | 切换至 |
|---|---|
| 用户希望得到讲解而非提问引导 | |
| 在计划落地前提出质疑 | |
| 演变为直接审查工作成果 | |
| 发现的问题应转化为长效经验 | |
| 交互式演示效果更好 | |