Lecture Builder

Lecture Builder（讲义生成器）

Turn a teacher's raw materials into a polished, print-ready revision lecture in two flavours from the same content:

• 学生版 (student) — knowledge + worked-example statements, each followed by a blank 答案/答题区 to fill in.
• 教师版 (teacher) — identical layout but each example carries the full 【答案】 and 【详解】 pasted from the solutions file.

Both are real .docx that open in WPS or Microsoft Word.

Why this skill exists (read this first)

Source teaching docs are full of figures and equations that cannot be retyped faithfully — physics problems are unsolvable without their diagrams, and option lists are often images/formulas. So the engine copies the original paragraph XML verbatim (images, OMML equations, tables) and only adds new structure (cover, headings, formula boxes, answer lines) around it. Never transcribe a problem or a formula by hand — let the engine carry the original bytes across. This is the whole point and the source of the quality.

The only thing that needs human-level judgment each time is mapping content to structure: which knowledge paragraphs form each section, where each problem begins/ends, and which knowledge point each problem belongs to. You produce that judgment as a small config.json; a deterministic engine does the heavy lifting.

Inputs

Ask the user for (or locate) up to three .docx files:

1. knowledge_file — the 知识点 doc (required).
2. question_file — the 题目 doc with problem statements (required).
3. solution_file — the 解析版 / answers + 详解 doc (optional but needed for a meaningful 教师版). If the user only has two files, build just the student version and tell them a solutions doc is required for the teacher version.

PowerPoint (.pptx) sources

The engine only reads .docx (a slide deck is not a linear document). If any source — most commonly the knowledge file — is a .pptx, convert it first:

python <skill>/scripts/pptx_to_docx.py "知识点.pptx" "知识点_转换.docx"

This renders every slide at high resolution (tables, figures, equations stay pixel-perfect) into a one-slide-per-page .docx, then you use that .docx as the knowledge_file. Tell the user the converted knowledge is faithful slide images (not editable text) — the right trade-off for dense slides. Needs Microsoft PowerPoint (Windows) or LibreOffice installed; the script auto-detects which.

Workflow

1 — Outline every source file

Run the extractor on each file so you can see paragraph indices, figures, formulas, and where problems start:

python <skill>/scripts/extract_outline.py "知识点.docx"          > kp_outline.txt
python <skill>/scripts/extract_outline.py "题目.docx"            > q_outline.txt
python <skill>/scripts/extract_outline.py "题目（解析版）.docx"   > sol_outline.txt

Read all three. Each line is pt_index TAG | text… flags, where ★START marks a problem's first paragraph and 〖IMG〗/〖MATH〗/〖TBL〗 flag figures/equations/tables. Those pt_index numbers are exactly what go into the config.

2 — Design the knowledge sections

Follow the knowledge file's own logical structure (its 模型/类型/章节 order). Pick 4–6 sections that match the subject. For each, record the [start,end] pt ranges of the knowledge paragraphs to include (skip pure decorative/watermark images at the very top). Keep the original wording — you are only choosing ranges, not rewriting.

3 — Segment the problems

For the question file (and solution file), find the contiguous blocks of problems and write them as segments {kao,start,end,type}:

• kao groups problems so identical numbers in different parts stay distinct (e.g. 考点01 → kao 1, 考点02 → kao 2).
• type is choice (选择题 → short 答案 line) or solve (计算/论述 → ruled 答题区).
• start,end bound the block; the engine finds each problem inside it automatically.

The solution file usually has different indices than the question file (solutions add paragraphs), so outline it separately and write solution_segments from its own ★START lines.

4 — Map each problem to a knowledge section (the judgment call)

For every problem, decide which knowledge point it best tests, by physics/subject theme — not just by source order. A problem may combine topics; pick its dominant model. Put the chosen {kao,num} into that section's problems list, in a sensible order. Aim to place all problems somewhere; if a few don't fit any of the chosen sections, widen the closest section's scope (and its title) rather than dropping them. It's fine for some sections to hold many and others few — reflect reality.

5 — Write config.json

Fill in the schema. Read references/config_schema.md for every field, the segment/section shapes, and the formula-token mini-language. Add 2–4 core formulas to the most foundational section(s) only; the equations already inside the copied knowledge are preserved automatically, so don't re-author the rest.

6 — Build

python <skill>/scripts/lecture_engine.py --config config.json --open

This writes the student file and (if solution_file is set) the teacher file, then opens them. --open uses the OS default app (WPS/Word) on Windows.

7 — Finalize the table of contents & page numbers

The engine sets the doc to auto-update fields, so the TOC and page numbers fill in the first time it opens. To bake them in (so a printed/handed-off file is correct without a manual F9), run the finalizer if Microsoft Word is installed:

powershell -ExecutionPolicy Bypass -File <skill>/scripts/finalize.ps1 -Path "lecture.docx"
powershell -ExecutionPolicy Bypass -File <skill>/scripts/finalize.ps1 -Path "lecture_教师版.docx"

If only WPS is installed (no Word COM), skip this — the TOC still updates on open; just tell the user to accept “更新域” / press F9 once.

8 — Verify before declaring done

Spot-check the result (e.g. export a few pages to images with pymupdf/fitz or open it) to confirm: figures render, equations look right, the TOC lists the sections, examples are renumbered 例1/例2…, and student answer regions vs teacher solutions are correct. Fix the config and rebuild if anything is off — rebuilding is cheap.

Common pitfalls

• File locked: if a build fails with a permission error, the target .docx is open in WPS/Word — close it (or Stop-Process -Name wps,WINWORD) and rebuild.
• Wrong segment bounds: if a problem's figure or options go missing or bleed into the next example, your start/end clipped it — widen the range using the outline.
• TOC shows placeholder text: fields haven't updated yet; open once and update, or run the finalizer.
• Don't strip or rewrite the copied content; the engine already removes only bookmarks/numbering/heading-styles that would conflict, and neutralizes the green source-citation color for clean printing.

Layout produced (fixed, professional, print-clean)

A4 portrait · margins 上下2cm 左右2.5cm · 1.5 line spacing · 黑体 headings (三号 H1) · 宋体 body (小四) · centered page-number footer · cover with 班级/姓名/日期 fields · auto TOC · each section = ◆核心知识 → ◆重点公式 → ★笔记区 → ◆对应例题训练. Decoration uses ■ ★ ◆ and rule lines only — no loud colors, suitable for printing.