Student Wrong-Question Diagnosis（学生错题诊断卡）

Overview

Turn a student's scanned or photographed wrong-question material into a self-contained offline HTML diagnosis card deck (KaTeX formulas, no internet needed) and a 7-step Markdown diagnosis report. The workflow is proven across multiple students and is fully offline-capable. The question-selection method is optional and never mandatory—students may circle with a highlighter, draw boxes/frames around the problem, or simply photograph the wrong-question pages; any of these tells the system "these are the questions I want diagnosed". Key features: flexible question identification (circle / box / photo), 7-step error attribution (知识没掌握 / 方法没想到 / 思维不完整 / 计算失误 / 审题问题), KaTeX offline rendering with bundled fonts, a triage rule for invalid selections, and a portable generator with zero third-party dependencies.

触发方式（Trigger）

当用户表达以下任一意图时即触发本技能（表述不必逐字一致，意思对即可）：

• 核心触发语句（标准说法，推荐统一使用）
  1. 处理学生错题扫描件 / 照片
  2. 生成错题诊断卡
  3. 整理荧光笔圈出的错题
  4. 把我框出来的错题做成卡
  5. 把拍的错题整理成诊断卡
  6. 离线打包错题卡

• 等价说法（同义即可触发，无需照抄）

• "把荧光笔圈的题目整理成诊断卡" / "扫描件里圈出来的题做归因"
• "把我框出来 / 画圈的题做成卡" / "我拍的这几页错题诊断一下"
• "生成可离线看的错题卡" / "把错题扫描件 / 照片做成带公式的卡片"

• 泛指"处理我的错题扫描件""诊断一下这份作业""这几张错题帮我归因"等，只要输入是学生标出或拍出的错题（荧光笔 / 框选 / 直接拍摄均可）即可。

• 不适用 / 不触发

• 纯文本错题（无任何扫描件 / 照片、也未标出任何题）建议走通用问答，不必走本流水线。
• 老师蓝笔标记的例题 / 变式 / 已订正，归入"巩固"卡，不视为诊断对象（与选取方式无关）。

> 触发后固定走 7 步流程（见 references/workflow.md）；资源与模板见下方「Bundled resources」。

When to use

• Input: a scanned PDF or a photograph of the student's wrong questions, where the target questions are indicated by the student—highlighter circles, hand-drawn boxes/frames, or simply the photographed problem pages. Blue-pen-titled 例题 / 变式 / 已订正 become "巩固" cards, not diagnosis targets.
• Output requested: 错题诊断卡 HTML + 诊断报告 .md, or any intermediate step toward them.
• 触发条件：见上方「触发方式（Trigger）」小节——用户只要说出其中任一核心触发语句，即按本技能处理。

Workflow (end-to-end)

Follow references/workflow.md for the full 7-step procedure. Key points:

1. Receive materials: PDF / photo + student info (姓名 / 科目 / 日期 / 文件名). If the student supplies a "第几页第几题" list, use it directly. Note the selection method the student used (highlighter circle / box / photo / list)—it is optional and never required.
2. PDF → PNG at 150 dpi (PyMuPDF/fitz), one image per page, page-numbered. For photos, use the image(s) directly.
3. Identify the target questions (selection method is flexible, see conventions):
   - Preferred: read the images, locate the student-indicated questions (highlighter ring, drawn box/frame, or the photographed problem pages).
   - Triage invalid selections: if a marked/boxed area clearly contains no real wrong question (blank space, page header/footer, teacher's remark, an already-corrected example, unrelated content), skip it—do not generate a card. Record it as an "无效选取" and confirm with the student when in doubt; never guess.
   - Fallback: if the environment cannot view images, do not guess—ask the student for the page/question list and proceed from that.
4. Author one card per wrong question using the schema in references/schema.md. Each card has 4 segments: correct approach (one), what went wrong (mistake), how to avoid it (action), plus ability / checked (error attribution). Consolidation cards use kind="巩固" with empty mistake.
5. Generate the HTML deck with scripts/generate.py (exact command in references/workflow.md). Default to offline mode with the bundled assets/katex-dist so the result is a zero-dependency single file.
6. Validate formulas: extract each \(...\) and render with node katex renderToString({throwOnError:true}); fix any LaTeX error (see references/formula.md) and regenerate.
7. Write the diagnosis report per references/report.md, then register in index / 台账 if a central repo exists.

Bundled resources

• scripts/generate.py — portable offline generator (Python 3 only, no third-party deps). CDN by default; --offline <katex-dist> inlines 60 KaTeX fonts for a fully offline single file.
• assets/katex-dist/ — KaTeX dist (css / js / contrib / fonts) for offline builds. 完整性证明见 assets/MANIFEST.md（含 60 字体 SHA-256 校验和）。
• assets/cards_data_template.py — fill-in skeleton for card data (Chinese-field comments, 错题 / 巩固 examples).
• references/workflow.md — full 7-step procedure + generate.py usage + deliverables + offline self-check.
• references/schema.md — card fields, CHECKS, examples.
• references/formula.md — <bs> placeholder, wrap_ce, LaTeX pitfall table, validation.
• references/report.md — 7-section report structure + 台账 format.

Conventions (must hold)

• 题目选取方式非必须（核心改动）：荧光笔圈、手绘框/画圈、直接拍错题页，都可作为「待诊断错题」的来源；核心约定是「学生自己标出或拍出的题 = 他想诊断的错题」，不限定必须用荧光笔。若学生未做任何标记、仅拍了错题页，整页即目标。
• 无效框选处置：框/圈/拍出的区域若不含真实错题（空白、页眉页脚、老师批注、已订正例题、与题无关内容），不生成卡片，记为「无效选取」并在报告/台账说明；存疑时向学生确认，不臆测。
• 老师蓝笔标题的例题 / 变式 / 已订正 = 巩固卡（kind="巩固"），与选取方式无关，永远不作为诊断对象。
• In card text, represent backslash with <bs> (not a literal \), restored by tex() at build time.
• Offline output must be zero external CDN references (verify: grep for cdn.jsdelivr → 0; inline data:font/woff2;base64 → 60).