/grilling 技能

它的作用

grilling 是压力测试计划、决策或想法的审问循环，在任何人行动之前进行。它把主题映射为 设计树: every decision branches into the decisions that hang off it, and interviews you branch by branch until no assumption stays unstated.

它不是一次问一个问题，也不是一次问完所有。每个 round 问整个 frontier: every decision whose prerequisites are already settled, and nothing else. Two questions never share a round if one depends on the other; a question that hinges on an answer still open belongs to a later round. Your answers settle decisions, the frontier moves outward, and the next round asks what that unblocked. Thirteen questions typically land in about three rounds rather than thirteen.

何时使用

输入 /grilling，或这个 agent 任务合适时自行调用它。它是唯一 skill 在这个 追问审视 family that is model-invoked, so you rarely type it. Usually a skill you did type runs it for you.

输入 /grilling 直接给你纯粹的审问，仅此而已。当你想要更多时：

轮次、前沿，以及谁决定

The skill rests on three ideas.

这个 设计树 是主题的模型：挂着子决策的决策。 frontier is the set of decisions whose prerequisites are all settled: the only questions the agent can ask yet. A round 是一个前沿，完整地提问，完整地回答。

Inside a round, every question has a fixed format: numbered and titled behind a ❓，然后正文，然后智能体推荐答案单独放在一个 ➡️ line. This format lets you answer a round by number ("1 yes, 2 the second option, 3 no, here's why") instead of by quoting questions back. The format has one known problem. The recommendation sometimes argues against 按原文措辞的问题，所以同意推荐意味着对问题回答「不」。发生这种情况时，回答推荐本身并说明。

The other half of the design is the split between facts and decisions. Facts are the skill's own job. When a frontier question needs something the 环境 can settle, the agent dispatches a sub-agent to find out rather than asking you. The round does not wait for that research; only the questions that depend on it wait. Decisions are yours, and the agent must wait for them. An agent running grilling that answers its own decisions has broken the skill. The session ends when the frontier is empty. The agent then waits for you to confirm a shared understanding before it acts on what you agreed.

The frontier has one limit. The agent chooses it by judgement; it does not compute it from a graph. So it can put two questions in one round and only later find that one answer should have changed the other. The only guard is to tell it. That reopens the affected branch in the next round.

什么住在这里，什么住在包装器里

This page covers the mechanism. The wrapper skills' pages answer the questions people ask most often.

常见问题

我可以回到一次只问一个问题吗？ 是的，而且很大一部分用户确实如此。把这个加进你的全局 CLAUDE.md:

When grilling, ask one question at a time.

Not everyone prefers the round-based default. Users who read slowly, who work in a second language, or who use the sequential format to help them focus report that one question at a time works better for them. The opt-out is a supported setting, not a workaround.

原来的 /batch-grill-me 去哪了？ 进入这个技能。轮次式提问曾短暂地作为独立技能发布，然后移入 grilling itself, so everything built on the primitive (grill-me, grill-with-docs, triage, wayfinder) got it at once. There is no batch-grill-me 要安装，也没有单独的连续式技能； CLAUDE.md 上面那一行是回到一次一个的方式。

一次问整轮问题，必然会失去我先前回答会引出的新问题。不是吗？ This is the most common objection to rounds, and the frontier answers it. A round only contains questions that do not depend on each other, so no answer in a round can invalidate another question in that round. Your answers still change what comes after them, because the agent works out each next round from scratch. You do lose something, but less than "all questions at once" implies. See the frontier's limit above.

它问题问完了，开始构建。 The confirmation gate exists for this. The skill is finished when you say the understanding is shared, not when the frontier empties. Weaker and faster models still skip the gate. Users report this most often on lower-effort or non-frontier models, which collapse "interview until shared understanding" into a couple of questions and an outline. If yours does it, the reliable fix is a line in your own AGENTS.md or CLAUDE.md 告诉智能体未经许可不得实现。

它自问自答，而不是问我。 That is a bug in the run, and it is the reason the skill's text separates facts from decisions. It happens most when another skill runs grilling to resolve a ticket, and the agent takes that task as permission to keep moving. The same constraint is why there is no async mode. People have asked for a variant that reads a GitHub issue and posts one consolidated decision memo, and that is a different skill, because a grilling session that nobody answers has produced the agent's opinion rather than yours.

我能限制问题的数量吗？ No, and a cap is out of scope on purpose. Some plans need three questions and some need fifty; a fixed ceiling either truncates the hard case or feels arbitrary on the easy one. The intended control is plain language. Tell it to wrap up, or stop and accept the plan where it stands. If a session is running very long, the cause is usually that the scope was too big; break the work up and grill the pieces.

我安装了 grill-me 单独使用，什么也不会发生。 grill-me 是一个整篇正文只有「运行一个 /grilling 会话」，所以它也需要安装这个技能。 grill-with-docs，它另外需要 domain-modeling。安装整套技能可以避免这个问题；选择性安装则意味着要把这些基础技能也一起装上。

grill-with-docs 运行了，但它从未加载 grilling. This is a real bug, still unfixed, and users report it across harnesses and models. A skill that names another skill does not reliably cause that skill to load, and grill-with-docs names two. The sign is a session that asks everything at once with no recommendations. That means the model is improvising an interview rather than running this one. Asking the agent directly whether it loaded grilling and domain-modeling 通常能恢复它。

做到以下就算成功

• 一轮以编号列表形式到来，每个问题及其推荐答案分别放在 ➡️ 行，你就可以用编号回答整轮。
• 一轮之内，没有问题需要同一轮中另一个问题先被回答。
• 后面的轮次会问第一轮不可能问的事。
• It goes and looks facts up (reading files, dispatching a sub-agent) rather than asking you something it could have found out.
• 后台运行的研究不会让整轮停滞；只有依赖它的问题等待。
• 它在最后停下来，请你确认理解已达成一致，而不是开始工作。
• 问题数量保持高位，轮次数量保持低位。

在流程中的位置

grilling 是一个 primitive，不是你需要安排的一步：它是审问技术的唯一权威来源，放在一处，让每个需要审问的技能都直接调用它，而不是自创一套。 grill-me and grill-with-docs 是它的两个由用户调用的前门，而 grill-with-docs 是主构建链开始的地方，先于 to-spec. wayfinder 运行它来解析决策任务， triage 把模糊的报告审问成可用的，而 improve-codebase-architecture 在选定要深化的候选后遍历树。当你不确定哪个入口合适时， ask-matt 为你指路。