ljg-qa
Information questioning engine. Given an article/paper/book, it extracts core arguments into Q-A pairs—Questions hit the core without sounding like a textbook; Answers are concise, clear, formally concluded, and logically complete. Readers follow the chain of Qs, with each A driving in a nail to reconstruct the author's entire line of reasoning. Use when user says '问答', 'Q&A', 'QA', '提问', '抽取问题', '/ljg-qa', or shares an article/paper/book and asks for Q-A extraction. Triggers when the user wants ideas extracted not as a summary but as a sequence of incisive questions with answered. NOT FOR FAQ generation, glossary creation, or comprehension quizzes — this is intellectual scaffolding, not study aids.
1How It Works & Mechanism
Read an article/paper/book into a 'Why–How–Boundary' Q-A chain, producing a reconstructed intellectual skeleton rather than a summary
读一份东西，把它的思想拆成「为什么—怎么—边界」的问答链。
Three iron rules dictate that Q must strike at the core, A must formally conclude in four sections, and the Q-chain must have directionality
*A 有形式化收口* —— 每个 A 严格四段：*结论*（一句话）+ *形式化*（用文字 + 简单符号把思想压成一行可视关系，如 `A = B + C`、`旧: X → 新: Y`）+ *论证步*（怎么想到的）+ *边界*（不成立的条件）。
Step 2 builds the spine first: write down the core thesis and 3-5 critical turning points to serve as the backbone for the Q-chain
这篇东西的核心论点是什么？它怎么把这个论点撑起来？哪几步是关键转折？
Order Qs by argumentative dependency rather than chapter sequence: ask root questions first, then solutions, then costs
按论证依赖关系，不按章节顺序。原文先写背景再写方法是叙事需要，但 Q 链应该按「先问根问题，再问解法，再问代价」走。
A's strict four-part requirements: borrowable conclusion, single-line formalization, exactly one argumentative step, and boundary stating only invalidating conditions
*边界*：写「不成立的条件」，不写「未来工作」——前者是诚实度，后者是公关辞令
Step 5 uses the deletion test to check Q-chain directionality: if deleting one causes subsequent ones to collapse, it is a valid path; otherwise, reorder or merge
如果 Q 之间是平行的（删一个不影响其他），重排或合并。Q 链是路径，不是清单。
Enforce org-mode output (markdown syntax prohibited), named and saved according to denote specifications
denote 文件名：`{YYYYMMDDTHHMMSS}--qa-{核心主题 5-10 字}__qa.org`
Qs fall into four types: Action/Contrast/Causality/Boundary; a good Q-chain mixes at least three, otherwise it degrades into a tutorial/debate script/paper/reflection
纯动作问题 = 教程 … 只有混合，才能复刻原作那种「方法 + 代价 + 限制」的真实张力。
Provide a taboo list for Qs (What is X, how many steps are there, is it important, pros and cons, what does it mean for the future)
❌ 「什么是 X？」 —— 一句定义打发，没承重
Gotchas pre-list default AI failure modes: textbook tone, scattered A's, asking questions by chapter, and piling on buzzwords to play it safe
*AI 默认会写「什么是 X」型问题* —— 教科书腔。生成后扫一遍，凡是 Q 能用一句定义打发的，重写
2Core Capabilities
3External Dependencies
| Kind | Dependency / Endpoint |
|---|---|
| network | 本地 Voice Notification 端点（workflow 执行前通知） |
| api | 宿主网页/文件获取能力（WebFetch / WebSearch / Read） |
4Recommended Governance & Evidence
- 本地工作目录: File I/O and artifact output: Written by the agent according to the output contract into ~/Documents/notes/{YYYYMMDDTHHMMSS}--qa-{主题}__qa.org (org-mo
5Second-Pass Verification
- [ok] 实现原理＝纯 prompt 编排，无自带可执行代码 — 全目录 3 个文件均为 Markdown（SKILL.md / References/QuestionDesign.md / Workflows/Extract.md）；扫描到的『执行外部程序』只有 SKILL.md 与 Extract.md 各一段 curl 本地通知，无其他脚本。
- [ok] Voice Notification 的 curl 调用真实存在且指向本地回环 — SKILL.md 与 Workflows/Extract.md 各有一段完全相同的 curl -s -X POST http://localhost:31337/notify 命令块，端点仅 localhost，不含任何外部域。
- [ok] 输出路径与命名在 SKILL.md 与 Extract.md 间一致 — 两处都写 ~/Documents/notes/ 与 {YYYYMMDDTHHMMSS}--qa-{主题}__qa.org，denote 前缀 qa- 与后缀 __qa 一致，无冲突。
- [ok] 「5-10 个 Q」数量纪律前后一致 — SKILL.md 未给数量；QuestionDesign.md「数量节制」与 Extract.md Step 3 均为 5-10 个 Q、7±2 最舒服，彼此不冲突。
- [ok] 形式化禁 LaTeX、禁 Unicode 绘图的约束可执行 — QuestionDesign.md 明列允许符号 = ≠ → ← + - × ÷ < > ⊃ ⊂ ⊥ ∧ ∨，禁 LaTeX 与复杂运算式，并给等式/对比/流向/递进四模式与『单独抽出来能否看懂』的检验句，可判定。
- [discrepancy] 风险扫描把正文示例词 token 误报为凭据命中 — §3「憑證/環境變數」5 条命中全部是正文里的『token 经济性』『数据 → token → 答案』等示例词，无任何环境变量或密钥读取；扫描为线索非结论，故 credential_reads 保持空。
- [ok] meta 直采项（stars/installs）无可定位 per-skill 值 — §0 未给 stars/last_push；安装量只有仓库级聚合，按契约 meta.stars / last_push / installs 保持 null。
6Verdict & Scenarios
1b797a9697841e58…