Breakdown全案攤開

Lay every case out in full before evaluating any of them, split the problem into decisions only you can make, then wait — recommendation comes last先把每個情況完整攤開再評估,把問題拆成只有你能拍板的決策,然後停住等你回答,建議放到最後

EN繁中Model-invoked模型叫用
Read SKILL.md on GitHub在 GitHub 看 SKILL.md

The failure this skill exists to prevent is a recommendation that arrives before the ground has been laid — where three options were considered, two were dismissed in a clause each, and you never got to see the one that was quietly dropped. Breakdown inverts the order: every case in full first, then the decisions split out for you to make, and only then a recommendation that has to account for every case it started with.

Install

npx skills add https://github.com/leoluyi/skills -g -a breakdown -y

To update later:

npx skills update breakdown

Source

What it does

Three phases, and phase 3 deliberately does not happen in the same message.

Phase 1 enumerates the distinct cases before evaluating any of them, and names the axis they vary on so you can judge whether it was cut the right way. Each case gets its own section covering what it is, what is actually true about it (with paths, values, versions, error text — each item marked fact, verified and where, or inference, reasoned and from what), why it's distinct from its neighbours, what follows from it, and what remains unknown. The do-nothing case is on the list. So is the awkward one nobody wants. Anything deliberately left out appears under an "excluded" heading with a reason, because a silent omission reads as coverage.

Phase 2 splits the problem into the individual decisions it contains, one question per decision, each carrying what hinges on it — which phase 1 cases it selects between, and how the recommendation moves with your answer. It only asks what changes the outcome; anything it could establish by reading the code, or that you already told it, doesn't get asked. Then it stops. It doesn't answer its own questions and doesn't start editing the part it thinks is settled anyway.

Phase 3 arrives after you answer: the recommendation as a decision rather than a menu, which of your answers drove it, where a different answer would have flipped it, what happened to every case from phase 1, what it decided without asking, what's still uncertain, and the single next step.

When to use

When a conclusion showed up too fast and you want the ground under it. Also when the space is genuinely wide — several credible approaches, several code paths, several failure modes — and you'd rather see them all laid out at uniform depth than trust a summary of them.

When not to

Not when there's one pending decision you'd like made clickable — that's Options, and it's a fraction of the work. Not when nobody knows the answer yet and the value is in exploring it together. And not for a question with a settled answer you could simply look up.

How it works

Two rules do most of the work. The first is that facts and inferences are never blurred: every claim is tagged as one or the other, with its source or its reasoning. The second is uniform depth — if one case gets four sentences and another gets a clause, that asymmetry is a decision made on your behalf, so the thin one either gets filled in or gets an explicit note saying why it can't be.

It also refuses to pad. If the subject genuinely has two cases, you get two, and it says so rather than manufacturing a fifth to look thorough.

If you'd rather not answer, saying "just decide" skips phase 2 — every unanswered question becomes an explicitly stated assumption in phase 3 instead of a silent one.

Related skills

options handles the narrow version: one pending decision, re-asked as tappable choices, without the full enumeration. discuss-with-me is for the case where the answer is unknown to both sides and the point is joint exploration rather than laying out a space you already suspect the shape of.

這個技能要防的失敗是:結論在底盤還沒鋪好之前就出現了——考慮過三個選項,其中兩個各用半句話帶過,而那個被悄悄丟掉的你從頭到尾沒看見。Breakdown 把順序反過來:先把每個案完整攤開,再把決策拆出來交給你拍板,最後才給建議,而且建議必須交代它一開始列出的每一個案。

安裝

npx skills add https://github.com/leoluyi/skills -g -a breakdown -y

之後更新:

npx skills update breakdown

原始碼

它做什麼

三個階段,而且第三階段刻意不會出現在同一則訊息裡。

第一階段在評估任何案之前先把它們列舉出來,並且明講它們變化的軸線是什麼,讓你判斷這條軸切得對不對。每個案自成一節,涵蓋:它是什麼、關於它實際成立的事實(附路徑、數值、版本、錯誤訊息,每一條都標明是事實——查證過、來源在哪,還是推論——怎麼推的、根據什麼)、它跟鄰近的案差在哪、由它會導出什麼、以及還有什麼不知道。不做的那個案在名單上,沒人想選的那個案也在。刻意排除的會列在「排除」標題下並附理由,因為沉默的省略讀起來就像已經涵蓋了。

第二階段把問題拆成它實際包含的個別決策,一題一決策,每題都帶著「這題影響什麼」——它在第一階段的哪些案之間做選擇,以及你的答案會讓建議往哪邊移動。它只問會改變結論的問題;讀程式碼就能查到的、或你已經講過的,都不會問。然後它停住。它不會自問自答,也不會先去改那些它認為反正已經確定的部分。

第三階段在你回答之後才來:建議寫成一個決定而不是一份選單、是你哪個答案主導了它、換一個答案會在哪裡翻盤、第一階段每一個案的下場、它沒問你就自己決定了什麼、還有什麼不確定、以及下一個具體步驟。

何時使用

當結論來得太快、你想看清楚它底下的東西時。或是空間真的很寬——好幾種可行作法、好幾條程式路徑、好幾種失敗模式——而你寧可看到它們以一致的深度全部攤開,也不想只拿到一份摘要。

何時不要

只有一個待決事項想改成可點選時不要用——那是 Options,工作量只有這個的一小部分。答案雙方都還不知道、價值在於一起探索時也不要用。已經有定論、查一下就有答案的問題同樣不要用。

運作方式

兩條規則承擔了大部分的作用。第一條是事實與推論絕不混淆:每一條主張都要標明是哪一種,並附上來源或推理過程。第二條是深度一致——如果一個案寫了四句、另一個只有半句,那個不對稱就是替你做了決定,所以薄的那個要嘛補齊,要嘛明確說明為什麼補不了。

它也拒絕灌水。如果題目真的只有兩個案,你就會拿到兩個,而且它會直說,不會為了看起來周全硬生出第五個。

如果你懶得回答,說一句「你決定就好」就會跳過第二階段——每一個沒回答的問題會在第三階段變成明確寫出來的假設,而不是一個沉默的假設。

相關技能

options 處理窄版的情況:一個待決事項,改成可點選的選項重問一次,不做完整列舉。discuss-with-me 則是給答案雙方都不知道、重點在於一起探索的情況,而不是攤開一個你多少已經猜到形狀的空間。