LaTeX Chinese Review Materials Generator (Universal)

One-click: any course materials → 6--12 PDFs (summary, exercises, 3--5 exams + answers).

Output Files

Generate 2 + 2N files: 2 base files + one paper and one answer file per exam. Default N = 5 (12 files: A--C plus D 提高 and E 挑战); small courses (≤3 chapters) N = 2 (6 files); if the user is time-constrained or asks for fewer, N = 3 (8 files).

File	Purpose
`知识点汇总.tex`	Per-chapter detailed review notes
`章节练习.tex`	Exercises with answers inline in notebox
`模拟卷A.tex` / `_答案.tex`	Foundations — questions only / solutions with rubrics
`模拟卷B.tex` / `_答案.tex`	Core chapters
`模拟卷C.tex` / `_答案.tex`	Synthesis across chapters
`模拟卷D.tex` / `_答案.tex`	提高卷 (default in the 5-set output)
`模拟卷E.tex` / `_答案.tex`	挑战卷 (default in the 5-set output)

Chapter split for N exams (maximizes coverage, no overlap):

A: first third — foundations
B: second third — core
C: last third + cross-chapter synthesis
D/E (optional): do NOT add new chapters. They re-test the same scope at higher difficulty (see §3.6).

For small courses (≤3 chapters): fall back to 2 exams (A: Ch1-2, B: Ch2-3 with overlap allowed).

Each exam uses the format extracted in Step 0. Fallback only if nothing is known: 20 MC (×1) + 10 multi (×2) + 10 T/F (×1) + 5 essay/calc (×10) = 100 pts, 120 min. Keep question topics NON-OVERLAPPING across the first three exams: a topic in Exam A should NOT appear in B or C. D/E may revisit topics but must change the angle and raise the cognitive level.

Critical rules:

模拟卷*.tex = questions ONLY + scratch paper. NEVER put answers in exam files.
模拟卷*_答案.tex = separate answer files with scoring rubrics.
Use Chinese filenames as shown above. For English-only courses, use English filenames instead: chapter_summary.tex, chapter_exercises.tex, mock_exam_A...E.tex, mock_exam_A...E_answers.tex.

Use when

User asks to generate "复习资料", "考点总结", "模拟卷", "练习题"
User provides course materials (PDF/PPT/DOCX) and wants PDF output
Any course type, any language mix (Chinese, English, bilingual)

Skip when: only 1-2 pages of material; user needs single file only; pure English without CJK.

Step 0: Extract Exam Intelligence (BEFORE writing anything)

The biggest failure mode is generating a generic paper that does not match the real exam. Extract and write down this table first; put a one-paragraph "exam brief" on page 1 of the summary and in the delivery README.

Item	Where to find it	Example
Date / duration	syllabus, last lecture	[date], [duration]
Question types & points	recording, syllabus	[e.g.] 20 MC×1 + 4 short×10 + 2 essay×20
Scope in / out	teacher's explicit statements	Ch1--6; Ch7 excluded
Authoritative source	teacher's rule	textbook / PPT / both / past papers
Emphasis / weighting	recording, repeated topics	Ch3 = core; later chapters = basics only
Terminology variants	textbook vs slides	textbook "Term A" = slides "Term B"
Aids allowed	syllabus	closed book

Rules:

Never default to 20 MC + 10 multi + 10 T/F + 5 essay unless nothing else is available. That format is a fallback, not the default.
If the exam has no multiple-choice, do not invent MC. If it has one 30-point case, build the paper around cases.
If a recording exists, transcribe it first. The last lecture almost always contains format, scope and emphasis.
Exclusions are exam content too: "Ch7 is NOT in scope" must appear on page 1.

0.1 Intel inventory: what you can get varies by course

Different courses give you different amounts of information. Record which tier you are in, and never silently upgrade an assumption into a fact.

Available material	What you can extract	What to do about the rest
Full recording / transcript	format, scope, exclusions, emphasis, terminology	best case — quote it as the exam brief
Syllabus / 教学大纲	chapters, weights, sometimes format	structure from syllabus; emphasis still unknown
Past exams (past 3--5 years)	question types, style, repeated topics	infer format from them; say "inferred from past papers"
Slides only	topics, terminology, "重点/考点" markers	scope = slides; exam may be PPT-based — ask
Textbook only	chapter structure and depth	assume full coverage, no format info
Nothing	—	ask the user (format + scope are the two questions worth asking); if impossible, state assumptions on page 1

Authoritative-source rule: write down which of textbook / PPT / both / past papers the exam follows. If unknown, ask before generating. Do not assume the textbook wins — many courses examine the slides.

0.2 Terminology handling depends on the authoritative source

Authoritative source	Which term to use
textbook	textbook term primary, slide term in parentheses on first use
PPT	slide term primary, textbook term in parentheses
both / mixed	use the term that appears in the exam stem or past papers; list both in a "术语对照" box
unknown	use the term most frequent in the instructor's materials and flag it

Always add one line: "两种表述均有效" only if the teacher actually said so; otherwise do not invent leniency.

0.3 Ask before generating (format + personalization)

Never start writing on assumptions. If the materials don't answer the questions below, ask the user — in one batch, not one by one.

Must know (blocking):

题型与分值：选择/简答/计算/证明/画图/编程/案例各几道、多少分、多长时间。
考试范围与排除章节。
命题依据：教材 / PPT / 两者 / 往年题。

Personalization (ask once, then apply): 4. 还剩几天、每天能投入几小时（决定资料厚度与套数）。 5. 哪几章最弱（加权：多出题、多解释、多对比）。 6. 目标分数（及格线 / 中上 / 冲高分 → 决定深度与是否加挑战题）。 7. 想要几套卷（默认 5：A–C + D 提高 + E 挑战；时间紧可只要 3 套，或 2 套）。 8. 答案要多细（只给答案 / 给步骤 / 给评分点 + 易错点）。 9. 是否需要 OCR 扫描教材、是否需要双语。

Defaults if the user does not answer (print them on page 1 so the user can correct):

默认 5 套（A–C + D 提高 + E 挑战）；时间紧才降到 3 套；答案给评分点 + 关键步骤；不 OCR；中文。
范围 = 所提供材料覆盖的章节；命题依据 = 材料中出现最多的那套，并标注“未确认”。

How to ask: paste one compact card (template in references/exam-blueprint.md §1), at most 6 questions at a time. If the user says "你看着办", state the defaults and proceed.

Step 1: Read ALL Materials

PDF — ALWAYS use pdftotext first (the Read tool falsely reports "password-protected" on many unencrypted PDFs):

pdftotext -layout "file.pdf" - 2>/dev/null | head -N   # N = pages needed
# For specific pages: pdftotext -layout -f 1 -l 30 "file.pdf" -

Only if pdftotext produces garbled output: try Read as fallback. If both fail, use python -c "import PyPDF2; ..." to attempt extraction.

DOCX/PPTX — Python zipfile + xml.etree.ElementTree (see reference for template). Always use glob.glob() to avoid Windows Chinese path bugs.

Scanned textbook (no text layer) — OCR pipeline:

Check: pdftotext -f 1 -l 3 book.pdf - prints nothing → it is a scanned PDF.
Use GPU OCR (PaddleOCR models via rapidocr_onnxruntime + onnxruntime-gpu, use_cuda: true): ~1.5 s/page at 200 dpi on an RTX 4070; CPU is ~30 s/page, so GPU matters.
Get chapter page ranges from the scanned TOC (OCR the first ~10 pages) or PDF bookmarks: PDF page = printed page + offset.
OCR chapter by chapter, one chN.txt per chapter, each page prefixed with 【第 N 页】 for later citation.
Skip chapters the teacher explicitly excluded — do not waste GPU time on them.
Tell the user the OCR error rate (~3%) and that numbers/punctuation must be checked against the paper book.

Lecture recordings: transcribe first (whisper / funasr / provided transcript). The last lecture usually states exam format, scope, exclusions and emphasis — this is Step 0 input, not optional background.

After reading, extract and record:

Chapter count, structure, which chapters have/don't have materials
Course name, instructor, textbook, exam format, grading weights (from syllabus/Ch0)
Instructor's emphasis: "重点"/"考点" labels, repeated topics, past exam topic lists
Language: all-Chinese / all-English / bilingual → determines output language
Any user-provided priority list or constraints (target score, study time)

Step 2: Detect Course Type

Signals in materials:
  Formulas, derivations, "证明", Greek letters, numbers to compute   → QUANTITATIVE (math/physics/investment/finance)
  Code, architecture, circuits, protocols, "设计/实现"               → ENGINEERING
  Events, dates, concepts, policies, ideology, legal articles        → QUALITATIVE (humanities/social science)
  Official documents, "精神/意义/新时代", value judgments, 时政        → IDEOLOGICAL (思政/政治理论)
  Mixed formulas + cases (management, finance, law)                  → MIXED

Templates:

Quantitative: Background → Core Models → Assumptions → Derivations → Properties → Testing → Mistakes
Engineering: Principles → Architecture/Design → Key Mechanisms → Formulas/Calculations → Examples → Trade-offs → Mistakes
Qualitative (humanities): Background → Core Concepts → Framework → Key Arguments → Comparisons → Significance → Confusions
Ideological (思政): 背景与定位 → 核心概念/原文表述 → 逻辑框架（为什么/是什么/怎么做）→ 联系实际 → 常见表述错误
Mixed (management/law/medicine): Concepts → Frameworks → Case Application → Decision Rules → Mistakes

Pick the template that fits. For mixed courses, vary per chapter. See references/course-types.md for the per-type playbook (marking logic, answer space, fatal mistakes).

2.1 Exam design differs sharply by course type

A paper that works for management will fail for 思政 and for engineering. Set these before writing questions.

Course type	Typical question types	Answer space	Marking logic	Fatal mistakes
Engineering	选择、填空、简答、计算、画图、设计、编程	计算 8--12 cm；画图半页（给坐标/网格）；代码 20--30 行	步骤分：公式→代入→结果→单位；代码看正确性/边界/复杂度	只写结果无过程；单位/量纲错；代码只过样例
Humanities	名词解释、简答、材料分析、论述	名词 4 cm；简答 8 cm；论述 1.5 页	观点 + 论据 + 结构 + 表达；允许多元答案但要论证	只罗列不论证；无文本/史料证据；时代错置
Ideological (思政)	单选、多选、判断、简答、材料分析、论述	简答 8--10 cm；材料分析 12--15 cm；论述整页	踩点给分，关键词必须是官方表述；多选按“少选得部分分、错选不得分”	口语化、自由发挥、引用错、观点偏差；用旧表述
Quantitative	选择、填空、计算、证明	计算 8--10 cm；证明 10--12 cm	步骤分 + 结果分；有效数字/近似条件	跳步；条件写错；结论与假设不符
Mixed (mgmt/law/med)	选择、简答、案例分析、论述	每小问 3--4 cm；论述整页	框架 + 案例事实 + 结论；法学按“结论—依据—涵摄”	只背理论不结合案例；结论无依据

Type-specific rules:

Engineering: provide a formula sheet / constants table at the front; require units in answers; leave grid space for graphs; state whether calculators are allowed.
Humanities: the answer file gives "要点 + 可接受的其他观点", not a single model answer; quote primary texts where the course does.
Ideological: every answer must anchor to the textbook/official document wording; add a "表述对照" box (口语说法 → 规范说法); never invent policy quotes; check the latest 会议精神/时政 within 12 months; flag any uncertain quote as "以教材原文为准".
Quantitative: state assumptions and significance level; answers show the full derivation chain.

Step 3: Plan Content with Priority Weighting

Chapter priority logic (use the first that applies): 0. User says a chapter is weak ("我第 4 章最虚") → weight it heaviest: more questions, more explanation, more comparisons.

User explicitly says "Chapter X is most important" or provides a priority topic list → weight heavily
Instructor's materials mark topics as "重点"/"考点" → weight accordingly
Past exam questions cluster on certain chapters → weight accordingly
Default: equal weight across all chapters

Apply personalization after Step 0.3:

Time budget: ≤3 days → summary + 1–2 sets; ≥1 week → summary + exercises + 3–5 sets.
Target score: pass → key points and basics; high score → add comparisons, edge cases, challenge questions.
Answer detail: "points only" / "steps" / "rubric + common mistakes" (default: rubric + key steps).
Weak chapters: add one extra worked example and one extra question per weak chapter.

Per-chapter minimums:

Summary: 150+ lines, 5+ subsections, 1+ comparison table/timeline, [CORE]/[KEY]/[NOTE] labels
Exercises: 10+ questions (4-5 MC, 2-3 multi, 2-3 fill-blank, 2 short, 1 essay). Answers in notebox.
Exam: use the format from Step 0 and the type-specific table in §2.1 (the 20+10+10+5 / 100 pts / 120 min shape is a fallback only)
Exam answers: separate file, scoring rubrics with explicit point breakdowns per question

Missing chapters: reconstruct from other materials. If nothing available, flag in output and use standard textbook knowledge for that topic.

Language output: Match the EXAM language. If English exam → English questions + bilingual explanations. If Chinese exam → all Chinese. Let source materials determine this.

Terminology alignment: apply the table in §0.2 — the primary term depends on whether the exam follows the textbook, the PPT, both, or past papers. Put the other name in parentheses on first use and never mix them silently.

Step 3.5: Design the Paper (answer space, cases, page breaks)

3.5.1 Answer space (MANDATORY)

A mock exam without writing space is a question list, not an exam. Every question that requires writing must be followed by blank space sized to the expected answer.

Question type	Blank space	LaTeX
单选 / 判断	none (options inline)	—
填空	0.8--1.2 cm per blank	`\vspace{1cm}`
名词解释	3 cm	`\vspace{3cm}`
简答 (10 pts)	3--4 cm	`\vspace{3.2cm}`
计算	8--12 cm	`\vspace{10cm}`
证明 / 推导	10--12 cm	`\vspace{11cm}`
画图 / 设计	half a page with grid or frame	`\vspace{12cm}` + grid
编程	20--30 lines + complexity note	`\vspace{14cm}`
论述 (25 pts)	9--12 cm, or a full page	`\vspace{10cm}`
案例分析 (30 pts, 3 小问)	3--4 cm after EACH sub-question	`\vspace{3.2cm}` ×3

Rules:

Put \vspace{...} immediately after the sub-question it belongs to, never after the section heading.
If the paper would overflow, insert \newpage before the next big question instead of shrinking the space.
End every paper with a scratch area: \newpage\section*{草稿区}\mbox{}\newpage\mbox{}\newpage.
Verify: every big-question sub-question (论述/计算/编程/案例) has a following \vspace in the source (the quality checker enforces at least 3 per exam file).
Exam files must stand alone: no "见教材第 X 页" / "见课件"; every condition, number and case fact is printed in the paper.

3.5.2 Big-question templates (six types — not only cases)

A paper's big questions must match the discipline. Pick from these six; a paper may mix two or three. Do not default to essay/case for a quantitative or engineering course.

A. Calculation (investment, finance, accounting, engineering, physics)

Structure: 已知 → 求 → 公式 → 代入 → 结果 → 单位/结论.
Space: 8--12 cm per sub-question; multi-period problems may need a full page.
Rubric: formula 30%, substitution 30%, result 30%, unit/significance 10%.
Must give: discounting convention, tax rate, units, rounding rule, and whether a formula sheet is allowed.
Typical mistakes: skipping steps, mixing nominal/real rates, unit mismatch, wrong sign.

B. Proof / derivation (math, theory, algorithms)

Structure: 已知 / 求证 → 证明（关键步骤）→ 结论.
Space: 10--12 cm.
Rubric: approach 40%, key steps 40%, conclusion 20%.

C. Diagram / design (electronics, mechanical, civil, networks)

Structure: 给定条件 → 画图/设计 → 标注 → 说明.
Space: half a page with grid or a blank frame (never a 3 cm strip).
Rubric: topology 40%, parameters/labels 30%, standard compliance 30%.
Provide the axes, units and required symbols; state whether drafts are allowed.

D. Programming (CS, data science, AI)

Structure: 题目描述 + 输入/输出示例 + 数据范围与约束 + 要求（语言、复杂度、是否允许伪代码）.
Space: 20--30 lines on paper, or a code box; leave room for complexity analysis.
Rubric: correctness 50% (including edge cases), complexity 20%, structure/style 20%, clarity 10%.
Include at least one edge case in the sample; say explicitly if only pseudocode is accepted.
Typical mistakes: happy path only, ignoring constraints, no complexity statement.

E. Case analysis (management, law, medicine)

Structure: case box (breakable infobox, 500--900 字, self-contained) → "请回答：" → 3 sub-questions with explicit points.
Each sub-question maps to a course framework (five forces / value chain / strategy / entry mode / 涵摄 / 鉴别诊断), not general knowledge.
Rubric: framework 40%, case facts 30%, conclusion/recommendation 30%.
If the folder has N cases and you need 2N questions, reuse cases with DIFFERENT angles and document the mapping.
Never split a case box from its questions (see 3.5.3).

F. Material analysis / essay (humanities, ideology/politics)

Structure: 材料/引文 → 设问（是什么/为什么/怎么做）→ 论述.
Space: 12--15 cm (material analysis) or a full page (essay).
Rubric: 观点 30%, 依据/原文 40%, 结构 20%, 表达 10%; 思政按踩点给分.
For 思政: every key point must use the official formulation; add a 表述对照表 (口语 → 规范).

3.5.3 Page layout for exams

\newpage before the big-question sections (论述/计算/编程/案例/材料) so each block starts on a fresh page.
Case box + "请回答" + first sub-question should fit on one page.
Per-paper header: \fancyhead[R]{Mock Exam A} (English only).
Paper title block: course, "期末模拟试卷 X", time, full marks, closed-book, and the chapter coverage line.

3.5.4 Delivery numbering

Number the PDFs in reading order: 01_知识点汇总, 02_章节练习, 03_模拟卷A, 04_模拟卷A_答案, 05_模拟卷B … Keep the .tex sources in a _tex源文件/ subfolder and the OCR text in NN_课本OCR_考点章节/. Put the exam brief in 00_说明.md.

3.6 Difficulty ladder (D/E papers, default)

Difficulty must come from deeper thinking on complete information, never from withholding data or making the stem ambiguous. A student must always be able to solve the question from what is printed.

Three levels — defined by cognitive demand, not by missing facts:

Level	What changes	Example
基础题	one concept, one step, direct recall/application	写出波特的三种基本战略
难题	several concepts linked, multi-step, cross-chapter, requires judgment	用五力 + 价值链判断某企业该不该进入某行业
易错题	information complete, but contains a common misconception or a boundary condition	给出两项业务收入占比，判断是相关约束型还是主导业务型（阈值易混）

Ladder design (information stays complete and solvable):

Axis	A/B/C	D 提高	E 挑战
Knowledge span	single point	2--3 linked concepts	cross-chapter integration
Reasoning steps	1--2	3--4	multi-step with a stated decision rule
Task verb	describe / explain	compare / analyze	evaluate / decide / predict
Distractors	none	similar concepts in options or data (still solvable)	two defensible frameworks, student must justify the choice
Common-error density	none	1 classic pitfall	2--3 boundary conditions

Per-section design:

Section	A/B/C	D 提高	E 挑战
单选	recall	discriminate confusable concepts	scenario judgment with two plausible options
简答	restate	compare + conditions	open synthesis, no single textbook answer
论述	one framework	two frameworks + evidence	thesis + counter-thesis + rebuttal + boundary
案例	complete data, 3 sub-questions	complete data + one distractor figure	complete data + a decision under constraints + risks
计算	single-step substitution	multi-step / multi-period (all given)	requires choosing a model + discussing assumptions
编程	direct implementation	added constraints / optimization	edge cases + complexity analysis + proof

Type-specific ladder notes:

Calculation: the harder version changes the setup (multi-period, taxes, inflation, embedded options), not the amount of given data. All rates and conventions must be stated.
Programming: the harder version adds constraints (time/memory), adversarial inputs, or a required complexity bound. Sample I/O and edge cases are always provided.
Proof: D adds an intermediate lemma; E requires constructing the argument from the full set of givens.

Marking shifts:

D/E add points for argument quality and trade-off justification, not for more content.
Still provide a reference answer, but label it "参考思路，不唯一".

Hard rules:

题干必须自足：条件、数据、约束齐全；不得删条件、不得制造矛盾、不得用模糊设问。
难度靠：概念联动、推理链长度、跨章整合、评价与决策、易错点（阈值、边界、易混概念）。
D/E 仍不得超出考试范围；易错 ≠ 超纲 ≠ 缺条件。

Step 4: Dispatch Agents (up to 12 Parallel)

Always generate ALL files simultaneously as background agents:

知识点汇总.tex (or chapter_summary.tex for English courses)
章节练习.tex (or chapter_exercises.tex)
模拟卷A.tex — Exam A (1st chunk of chapters)
模拟卷A_答案.tex — Exam A solutions
模拟卷B.tex — Exam B (2nd chunk)
模拟卷B_答案.tex — Exam B solutions
模拟卷C.tex — Exam C (3rd chunk + cross-chapter synthesis)
模拟卷C_答案.tex — Exam C solutions
模拟卷D.tex — 提高卷 (same scope, higher difficulty per §3.6)
模拟卷D_答案.tex
模拟卷E.tex — 挑战卷 (highest difficulty per §3.6)
模拟卷E_答案.tex

D/E are part of the default 5-set output; drop them (N = 3) when the user is time-constrained or asks for fewer. For ≤3 chapters: use 2 exams (6 files: summary + exercises + 2 papers + 2 answers). Each exam covers DIFFERENT chapter groups to maximize topic coverage with zero overlap.

If the host agent has no parallel sub-agents, generate the files sequentially in the same order — the parallel step is an optimization, not a requirement.

## AGENT PROMPT TEMPLATE

Write LaTeX file at [ABSOLUTE PATH].

Course: [NAME], [INSTITUTION], [INSTRUCTOR]
Textbook: [TITLE, EDITION]
Exam language: [EN/CN/BILINGUAL]

### Preamble:
Use `references/latex-format-rules.md §15` verbatim, either inline or as `\input{preamble.tex}` (keep it in the same folder).

### CHAPTER-BY-CHAPTER CONTENT:

Ch N: [TITLE] — MUST cover:
- Topic 1: [specific concept/formula/event]
- Topic 2: [specific concept/formula/event]
- ... (5-8 specific topics per chapter)
- [EXAM] questions on: [specific calculation/essay types]
- Key terms to include: [list 5-10 English+Chinese terms]

[... repeat for all chapters ...]

### Mandatory Rules:
1-11 [paste the 11 rules from the "Mandatory Rules" section]

### Post-write checklist (run yourself):
- [ ] `grep -c '\\begin{enumerate}' file.tex` == `grep -c '\\end{enumerate}' file.tex` ?
- [ ] `ulem` present in the file or in the `\input` preamble?
- [ ] `grep -P '[\\x{1F300}-\\x{1F9FF}]' file.tex` returns empty?
- [ ] All tcolorbox titles are English or short Chinese, with no `[` and no ASCII comma inside?
- [ ] (exam files) every big-question sub-question (论述/计算/编程/案例) followed by `\vspace`?
- [ ] Big questions match the discipline (calculation / proof / diagram / programming / case / material)?
- [ ] No 参考答案/评分标准 in exam files?
- [ ] Format matches Step 0 (not the 20+10+10+5 fallback)?

Target: [N] lines. Write file now.

The per-chapter content list is the most important part. Don't write "summarize Ch1" — enumerate every specific concept, formula, event, and term that must appear. The quality of this list determines the quality of the output.

Step 5: Quality Check (BEFORE compiling)

python references/quality_check.py *.tex

If any check FAILS → fix with Edit, re-run check, THEN compile. The checker validates: enumerate balance, required packages (skipped when a \input preamble is used), nested [ and ASCII commas in tcolorbox titles, table floats inside tcolorbox, Chinese in headers, emoji, answer space in exam files, and answers leaked into exam files.

Step 6: Compile and Verify

# compile every .tex in the working dir (handles Chinese filenames + \input{preamble})
for f in *.tex; do
  base="${f%.tex}"
  xelatex -interaction=nonstopmode -halt-on-error "$f" >/dev/null 2>&1
  xelatex -interaction=nonstopmode -halt-on-error "$f" >/dev/null 2>&1
  if [ $? -ne 0 ]; then echo "FAILED: $f"; fi
done
# Verify: no LaTeX errors
grep -l '^!' *.log || echo "no errors"
# Check pages
for p in *.pdf; do echo "$p: $(pdfinfo "$p" | awk '/Pages/{print $2}')"; done

Notes:

If the project uses \input{preamble.tex}, keep all files in one directory and compile from there.
Compile twice for TOC and cross-references; long documents (30+ pages) may need a third pass.
Do not delete .log before running grep '^!'.

Step 7: Error Recovery

If compilation fails:

grep '^!' *.log | head -5 — find first error
Common fixes: missing \usepackage[normalem]{ulem} → add to preamble; enumerate mismatch → remove stray \end{enumerate}; undefined command → add \DeclareMathOperator or \newcommand
Fix with Edit, re-run Step 5, re-compile
If a background agent is still writing, stop it first, then fix

11 Mandatory Rules (in EVERY agent prompt)

1. \documentclass[12pt,a4paper,fontset=fandol]{ctexart}
2. ZERO emoji — use [CORE][KEY][NOTE][EXAM][TIP] labels
3. Tables: \medskip\begin{center}\begin{tabular}... — NEVER \begin{table}[h] inside tcolorbox
4. Headers: ENGLISH ONLY in fancyhdr
5. Colors: ALL \definecolor{}{RGB}{...} in preamble
6. Compile: xelatex twice; clean aux
7. tcolorbox titles: English or SHORT Chinese. NEVER put `[` or an ASCII comma inside the optional title
8. \uline{} → MUST have \usepackage[normalem]{ulem} in preamble
9. \begin{enumerate} count MUST equal \end{enumerate} count — COUNT THEM after writing
10. Exam files: leave \vspace after every big question (论述 9--12 cm; 计算 8--12 cm; 编程 20--30 行; 案例每小问 3--4 cm); never put 参考答案 inside exam files
11. Big questions must match the discipline: calculation / proof / diagram / programming / case / material analysis — never all essays

Common Errors (quick reference)

Error	Fix
`\uline` undefined	Add `\usepackage[normalem]{ulem}`
enumerate mismatch → blank PDF	Remove stray `\end{enumerate}`
`Missing $ inserted` from tcolorbox title	Use English title, no `[` inside `[...]`
`Package pgfkeys Error: I do not know the key` in a box title	Remove the ASCII comma from the title or wrap the title in braces
`\cov`/`\plim` undefined (quant. courses)	`\DeclareMathOperator{\cov}{Cov}` / `\newcommand{\plim}{\operatorname{plim}}`
`\E`/`\var` already defined	Use `\mathbb{E}` or skip (amsmath conflict)
Read tool says "password-protected"	Use `pdftotext` instead (Read tool bug)
Chinese in fancyhdr	English only

Anti-Patterns

Mistake	Fix
"summarize Ch1" as agent prompt	List every specific topic/formula/event
Forcing quantitative template on history	Step 2 type detection
Answers inside mock_exam.tex	Separate answer file always
Skipping quality check before compile	Step 5 on every file
Using Read for PDF → false encryption error	`pdftotext` first
No blank space under essay/case questions	Add `\vspace` per Step 3.5.1
Making D/E harder by obscure or out-of-scope wording	Raise integration/judgment per §3.6
D/E adding new chapters	D/E re-test the same scope at a higher cognitive level
Case box split across pages from its questions	`\newpage` before 论述/案例分析 (Step 3.5.3)
Exam requires the textbook / original case PDF	Rewrite the case self-contained (500--900 字)
Exam says "见教材第 X 页"	Print all conditions and data in the paper
Using slide terminology when the textbook differs	Use the authoritative source's term + the other in parentheses (§0.2)
Assuming "textbook is authoritative" for every course	Record the source per course; many courses examine the PPT
Assuming an exam format when no intel exists	Ask the user for format + scope; state assumptions on page 1
Using the humanities template for 思政	Use the IDEOLOGICAL template: official wording + 踩点给分 (§2.1, course-types.md)
思政 answers written in colloquial wording	Add a 表述对照表 (口语 → 规范)
Engineering paper without formula sheet / graph grid	Add them per course-types.md §2
Humanities answer file with one "model answer"	Give 要点 + acceptable alternative views
Generating 4 PDFs for a 3-exam course	2 base + 2 per exam = 8 files
Generic 20+10+10+5 format for every course	Extract real format in Step 0 first

Special Handling

Bilingual courses (English exam + Chinese understanding): All output: Chinese explanations + English technical terms. Exam questions in English. Answers bilingual.

User-provided priority list: When user gives topic list (e.g., "Topic A, Topic B, Topic C are most important"), restructure chapter weighting — priority topics get 2× the content depth, non-priority topics get minimal treatment (formula only).

Massive courses (10+ chapters, 1000+ pages): Read syllabus/Ch0 first. Read Ch1-2 fully to understand course style. For remaining chapters, read first 15 pages + last 5 pages (summary/sample questions). Time-box reading to ~20% of agent generation time.

Missing/inaccessible materials:

User-provided cheat sheets / topic lists → use as primary content outline
Standard textbook knowledge → use for well-known courses with a standard textbook
Flag uncertainty in output: "Based on standard course content; verify with your materials"

Reference Files

references/latex-format-rules.md — Full preamble, tcolorbox styles, color definitions, compilation
references/course-types.md — Per-type playbook: 思政 / 工科 / 文科 / 商科 / 投资学·金融 / 计算机 / 法学 / 医学 (question types, marking, answer space, fatal mistakes)
references/exam-blueprint.md — Exam brief table, answer-space table, case/answer templates, delivery checklist
references/quality_check.py — Automated pre-compilation checker

Ask in your favorite AI

Documentation