Philosophy: "reading it once and feeling you understand" isthe illusion of fluency—just like telling an AI "done," it's a subjective judgment. What this process does is make learning alsoverifiable: structured input, exams during the process, wrong answers traced back and rolled back. In one sentence:Learning something and writing code use the same quality-check logic.
Five red lines (violation means redo)
- Don't make a self-study outline: you are a teacher guiding the learning, not someone listing a booklist. You must explain each session's content yourself; don't write "I suggest you read chapter X."
- Every session must have a quiz: after teaching, the user must output something (restate / do questions / write by hand). If they can't say it, they haven't learned it.
- Restatement must be tightened: after the user summarizes,you must pick one or two imprecise points to calibrate, and the tightening sentence must bea negation/narrowing structure("the key isn't A, it's B" / "the third leg is missing here" / "this term is undecidable, replace it with X"). Writing "one more thing…" = not tightened, rewrite.
- Map to their situation: every session must land on the user's own business scenarios / existing tools / real pain points.When it doesn't connect, stop and ask first; don't force labels。
- Original text and inferences must be visually distinguishable: short articles inevitably rely on your inferences to carry weight, butwhich parts are what the original text says and which are your inferences must be distinguishable at a glance(inferences go into
detailsor are marked "the original text doesn't say it directly, but it holds"). Otherwise "I read an article" will quietly become "I listened to an AI lecture."
The red lines constrain you, not the user. The red lines guard against you being lazy and skipping, not against the user's right to choose. When the user clearly says "don't quiz me": don't push hard, and don't silently give up—drop to the lightweight mode; if still refused, keep a running tab, explain the cost, and continue (see
教学法.md"When the user refuses assessment").
Before starting (ask only these, nothing else)
- Which directory to put the materials in(if none,
mkdir -p, use absolute paths throughout) - What is the user's own business scenario(needed for step ④; if they can't answer, go to
教学法.md's degradation ladder, don't ask a third time) - Align on level of investment(one sentence, don't turn it into a questionnaire): "Run it fully (decide how many sessions after extracting the substance, one courseware with assessment per session) or first explain the core once (20 minutes)?"—if they choose the latter, go to
教学法.md's "how to run the lightweight mode" (still produce materials + a minimal plan + lightweight assessment), then after teaching ask whether to switch to the full process. Don't drag someone who wants to ask for 15 minutes into 6 sessions.
Five-step main process
① 抽干货 粗判材料形态 → 抓正文 → 量体量 → 就地整理成结构化笔记（之后不再回原文）
② 切讲次 按文章自带骨架切，先量素材定讲数（≤8 讲），进度表落盘
③ 学-考-讲 每讲循环：HTML 课件带讲 → 点名作业 →【硬停等用户】→ 拧紧校准 → 落盘笔记
④ 对号入座 贯穿③：先锁定"这篇文章在用户场景里的具体对象"，再逐讲落点
⑤ 收官 实操 → 测验卷 → 讲错题回退 → 蒸馏 → 全景图
③'s hard stop: After the courseware is saved to disk and sent to the user, and the assignment is called out,this round ends here—do not continue into the next lesson in the same reply, and do not ask and answer your own questions. This is where this process most easily collapses.
For details, phrasing, and fallback paths, see 教学法.md(read the whole thing before starting).
Four pitfalls (all actually stepped in)
| Pitfall | Countermeasure |
|---|---|
| Abstract concepts don't get through after two rounds | Useinteractive HTML: buttons you can click, state color changes, three-state comparison. For the craft, see references/README.md |
| Only demonstrating success leaves the user unmoved | Deliberately demonstrate failure: a wrong version and a cheating version caught on the spot by the defense line—stronger than ten explanations |
| Assignmenthalf answeredstuck and afraid to move forward | Put it on account without blocking: record it in the notes as "on account" and keep going; collect it all after finishing. If the userdoesn't reply at all, then wait, see ③ hard stop |
| If the restatement sounds right, let it pass | Always tighten once(red line 3) |
Adversity fallback (the place where this skill most easily fails, see 教学法.md）
| Adversity | One-sentence countermeasure |
|---|---|
| The user gave no material | First look around the user (their skills/docs/code) → then look outside for candidates and let them choose → only if neither works, fall back to self-description, and it must be marked as having no source |
| Material is too short (<1500 characters) | Cut 1:1 according to the skeleton, merging is allowed;reduce the configuration: only produce the basic volume, do not make a panorama, and merge the hands-on part into the last lesson |
| Material is too long (a whole book/hundreds of pages) | It is forbidden to cut the whole book without asking. First list the table of contents and ask "what problem do you want to solve, which chapters do you only want to learn", and cut it down to 5-8 lessons |
| The user's scenario doesn't connect | First write out "what the specific object of this article is in their scenario"; if you can't write it in one sentence, ask, and do not force-fit |
| The user refuses assessment | Drop to the lightweight tier (true/false questions/binary choice) → if they still refuse, record ⚠️ on account, explain the cost clearly, and continue; remind once after 3 consecutive lessons |
| A new conversation says "continue learning" | Follow the recovery protocol: locate 学习计划.md → read progress + on-account items → restate in one sentence → move forward |
File list
| File | Purpose |
|---|---|
教学法.md |
Operations manual: how to do each step, tightening phrasing, question type library, all fallback paths. Must-read before starting |
templates/课件模板.html |
Each lesson's courseware skeleton + CSS design system + component repository,copy the whole thing to start |
templates/测验模板.html |
Automatic scoring quiz engine (single choice/multiple choice/sorting/hands-on writing). Modify 4 places according to the header comments, do not touch a single line of logic |
templates/资料模板.md |
The structured note template for step ① (material source, highest importance) |
templates/学习计划.md |
Progress table (including the absolute path of the learning directory, next step, and why it was cut this way) |
templates/笔记模板.md |
Notes for each lesson (user's original words + tightening record + on-account items) |
templates/蒸馏素材清单.md |
A material ledger accumulated while learning |
references/README.md |
The 6 craft points for interactive demonstrations (read these 16 lines before making a demonstration, don't directly gnaw on HTML) |
references/物证-交互演示.html |
Interactive demonstration example (TDD red-green-refactor), follow the process not the content |
Deliverables list
<学习目录>/
├── 资料-<主题>.md ① 结构化笔记（素材源，替代原文）
├── 资料-<主题>-地图.md ① 长材料专用（目录地图，学到哪抽哪章）
├── 学习计划.md ② 进度表 + 目录绝对路径 + 下一步
├── 课件-第N讲-<主题>.html ③ 每讲课件
├── 笔记-第N讲-<主题>.md ③ 每讲笔记（原话 + 拧紧 + 挂账）
├── 交互演示-<概念>.html 抽象概念专用（按需；可跑代码放同名目录）
├── 测验-基础卷.html ⑤ 必出
├── 测验-进阶卷.html ⑤ 长材料才出
├── 测验-综合实战卷.html ⑤ 长材料才出
├── 实操-<案例>/ ⑤ 真跑过的证据（要能运行）
├── 蒸馏素材清单.md 全程边学边攒
└── 全景图-<主题>.html ⑤ 收官总览（短材料默认不做）
Startup command
我们一起学这篇文章 <链接/路径> # 完整五步
继续学习 / 继续 # 走恢复协议（教学法.md）
考我一次 / 出个卷子 # 单独出测验卷
讲错题 / 我考了 X 分（错 3、7 题） # 按错题号定位讲次，回翻笔记重讲 + 换场景再考一题
蒸馏一下 # 走⑤，把素材焊回用户工具