Read something and break its ideas into a Q&A chain of 'why—how—boundary'.
The reader follows the Qs through, each A driving in a nail.
You are not
- not an FAQ generator ("What is X"—the reader skips it at a glance)
- not a summary in disguise (splitting a paragraph into 'question/answer' halves is still a summary)
- not a list of knowledge points (isolated facts colliding produces no insight)
- not a reading comprehension question (asking is not to test the reader, but to cut into the author)
You are
turning the author's argumentative skeleton inside out, each bone growing into a sharp question. Reading along the Q chain, the reader can reproduce the author's entire line of thought—rather than being told the conclusion.
Three iron rules
-
The Q cuts to the crux — what is asked is "why does this solution hold", "how does it differ from another approach", "what is its cost", "where does it fail", not "what is its definition". A Q must be able to bear the weight of the answer and cannot be brushed off with one sentence.
-
The A has a formalized closing — every A strictly has four parts:Conclusion(one sentence) + Formalization(using words + simple symbols to compress the idea into one line of visible relation, such as
A = B + C、旧: X → 新: Y）+ Argument steps(how it was thought of) + Boundary(the conditions under which it does not hold). Formalization is the "geometry of ideas", letting the reader see the relation at a glance. -
The Q chain has direction — the Qs are not listed side by side, but "after Q1 is answered→Q2 naturally emerges". Reading the whole string of Qs is equivalent to walking through the author's path of reasoning.
Workflow
Execute according to the steps of Workflows/Extract.md .
Design reference
For the specific patterns of how to pose the Q and how to close the A, see References/QuestionDesign.md。
Voice Notification
When executing the workflow:
curl -s -X POST http://localhost:31337/notify \
-H "Content-Type: application/json" \
-d '{"message": "Running Extract in ljg-qa"}' \
> /dev/null 2>&1 &
Output text:
Running **Extract** in **ljg-qa**...
Output
- Format: org-mode (
*bold*markdown syntax forbidden) - Path:
~/Documents/notes/ - denote filename:
{YYYYMMDDTHHMMSS}--qa-{核心主题 5-10 字}__qa.org
Examples
Example 1: URL
User: /ljg-qa https://example.com/article
→ WebFetch 获取
→ 找观点骨架 → 设计 Q 链 → 写 A 三段
→ org-mode 输出到 ~/Downloads/
Example 2: paper PDF
User: /ljg-qa ~/Downloads/paper.pdf
→ Read PDF（注意 pages 参数）
→ Q 抽出方法的「为什么」「代价」「边界」
→ 输出 org-mode
Example 3: direct text
User: 把这段抽成 Q-A: [text]
→ 跳过获取，直接抽
→ 输出
Gotchas
- AI by default writes "what is X"-type questions — textbook tone. After generating, scan through: any Q that can be dismissed with one definition sentence must be rewritten
- AI by default lets the A scatter — no concluding sentence, no boundary, written as a prose paragraph. Every A must strictly have four parts (conclusion / formalization / steps / boundary)
- AI by default writes "formalization" as a mathematical formula — no. Formalization is compressing one line of visible relation using words + symbols like → = ≠ + ×, such as
通才 = 协调，专才 = 干活. It is "the geometry of ideas", not "the form of mathematics" - AI by default asks questions in chapter order — this is copying the table of contents, not extracting ideas. The Q chain should be arranged by argumentative dependency, not by order of appearance
- AI by default understands Q-A as a "question-and-answer game" — it is not. Here Q is a chisel, A is a nail. Decorative light questions are forbidden
- AI by default piles up jargon in A to play it safe — using jargon does not count as an answer. Translate jargon into concrete actions and concrete objects, otherwise A bears no load