# Breakdown > Before recommending, exhaustively lay out every case, separate facts from inferences, list exclusions, and isolate decisions only the user can answer; then stop and wait. Trigger on 「先把所有情況攤開給我看」, 「不要先給結論」, 「幫我拆解這個決定」, "lay out every case first", or "don't recommend yet - decompose it". Route immediate option selection to options, joint unknown exploration to discuss-with-me, and settled lookups to ordinary research. - Skill: `leoluyi/breakdown` (Agent Skill, multi-file: 5 files) - Install (CLI): `npx skillmds@latest add leoluyi/breakdown` - Raw SKILL.md: https://api.skillmd.com/api/skills/leoluyi/breakdown/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Research & Search - License: MIT - Author: leoluyi (https://skillmd.com/u/leoluyi) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/leoluyi/breakdown --- I don't want a conclusion yet. I want the ground laid out first, then the decisions separated out so I can make them, and only then your recommendation. Three phases, in order, and **phase 3 does not happen in this message**. ## Scope If the user named a subject when invoking this skill, that is the subject. If they named nothing, the subject is whatever we are already working on — the thing you just proposed, the bug we are looking at, the decision in front of us. Don't restart the topic and don't ask me to restate it. Before writing anything, gather the facts. Read the actual code, config, docs or data that the subject depends on. If a fact matters to the breakdown, verify it rather than recalling it, and delegate the sweep to a subagent when it means reading widely. A breakdown built on assumed facts is worse than no breakdown, because it looks complete. ## Phase 1 — every case, one at a time, in full Enumerate the distinct cases first, and enumerate them *before* you evaluate any of them. A case is anything that would be handled differently from its siblings: a candidate approach, a scenario or code path, an environment, a data shape, an affected user or system, a failure mode, an edge case. Whatever axis the subject actually varies on — name the axis explicitly so I can tell whether you cut it the right way. Then give each case its own section, and in each one cover: - **What it is** — one line, concrete enough that I could point at it. - **What is actually true about it** — the specifics, with file paths, values, versions, error text, quantities. This is the part I want in full; do not summarise it away. Mark each item as **fact** (you verified it, and where) or **inference** (you reasoned it out, and from what). Never blur the two. - **Why it's on the list** — what makes it distinct from the neighbouring cases. - **What follows from it** — cost, risk, blast radius, what it forecloses, who it affects. - **What you don't know about it** — the gaps, stated plainly. Rules for this phase: - Exhaustive before selective. Include the cases you think are wrong, the do-nothing case, and the awkward one nobody wants. If you deliberately left something out, list it under an "excluded" heading with the reason — a silent omission reads as coverage. - No pre-ranking, no "obviously the best option is". That is phase 3. - Uniform depth. If one case gets four sentences and another gets a clause, that asymmetry is you deciding for me. Either fill the thin one in or say why it can't be filled in. - Don't pad. If the subject genuinely has two cases, give me two. Manufacturing a fifth to look thorough wastes both our time — and say so when the space is small. - If the list runs long, group the cases under headings, but never truncate. "and others like it" is not a case. ## Phase 2 — decompose it into questions I answer Now split the open problem into the individual decisions it actually contains. One question per decision, each one independently answerable, and each one carrying **what hinges on it** — which cases from phase 1 it selects between, and how the recommendation changes depending on my answer. Only ask what changes the outcome. If my answer would not move the recommendation, don't ask it — decide it yourself and note the call in phase 3. Likewise don't ask me anything you could establish by reading the code, and don't re-ask anything I have already told you in this conversation. Deliver them in whichever form fits the answer: - **A tappable-choice tool** when the answer space is enumerable — 2-4 mutually exclusive options, labelled by outcome, trade-off named in one line, the one you'd recommend first and marked as such. Four is the usual cap. - **A numbered prose list** when the answer is genuinely open: a number, a name, a constraint, a priority I hold that you have no way to guess. This is also the fallback wherever no tappable-choice tool exists. Batch them into as few rounds as possible; if there are more than four closed questions, group the ones that share an axis. Order them by leverage, most consequential first, and say when one question's answer makes a later one moot. Then **stop**. Do not answer your own questions. Do not write code, do not edit files, do not start on the part you think is settled regardless — this skill is for thinking, and an edit made before I have answered is an answer you gave yourself. Wait. If I reply "just decide", skip the questions, or tell you to proceed, go straight to phase 3 with every unanswered question converted into an explicit stated assumption. ## Phase 3 — synthesize, after I answer Once I have answered, give me the recommendation. It has to visibly rest on phases 1 and 2: - **The recommendation**, stated as a decision, not a menu. - **Which of my answers drove it**, and where a different answer would have flipped it. - **Every case from phase 1, accounted for.** Say what happens to each one — chosen, rejected and why, folded into another, or deferred. A case that quietly disappears between phase 1 and phase 3 is the failure mode I am trying to avoid. - **What you decided without asking me**, and the reasoning. - **What is still uncertain**, and what would resolve it. - **The next concrete step**, and nothing more than the step. Wait for me to say go before you take it. Keep phase 3 shorter than phase 1. By this point the detail is already on the table; repeating it is not synthesis.