Exam Cram Coach · Flash is the lightweight edition of this project: one command prepares a whole course, the tutor shows the figures it cut out of your lecture notes and printed problems, quizzes you only with your own homework and past papers, and works with small models. The v4.3 full edition (HTML/PDF study guides) now lives side by side in full/.
中文说明在下方 / Chinese version below
English
What it is
- One
setupcommand reads PDF / PPTX / DOCX / Markdown / TXT / HTML, splits chapters, pairs homework with solution files (hw2 (4)(1).pdf↔homework2solutions.pdf,Problem 1.3.10↔ its solution), crops figures and builds the retrieval index — a thousand pages in a few seconds. - Figures in the chat: vector drawings and embedded images are cropped from lecture pages; the printed question and solution regions that contain a picture are cropped separately.
nextlists the figures of each slice,quizshows the question figure,checkthe answer figure. Scanned or handwritten pages (your own submissions) are detected and never shown as answers. - Seven commands run a whole session:
next / ask / quiz / check / answer / note / done. Every command ends with a 📍 progress line naming the next command, so the agent never has to remember the CLI. - Honesty labels are unchanged: 🟢 from your materials / 🟡 AI supplement /
⚠️ AI-generated answer. When only a textbook number is known, the quiz restates the givens from the reference solution instead of inventing a problem. - Small models verified: Gemini flash-lite, Gemini flash (Antigravity) and Claude Haiku 4.5 followed the skill end to end on a real 1000-page course — setup → next → quiz → check, page citations, labels, cropped figures embedded, no invented commands.
New in 5.1.1
export [--to DIR] [--qid …] [--chapter N] [paths…]copies the figures listed by the last command into a folder the chat UI can render (VS Code / Electron panels blockfile://images outside the open workspace) and prints relative paths; the skill now describes a three-level display strategy.- 📍 progress line after every command +
plan [--days N](day-by-day chapter plan, last day reserved for mistakes and the cheat sheet) +statusshows today's target — long sessions and new chats stay on track. - The v4.3 full edition is restored verbatim under
full/with its own README, SKILL and tests; the original contributing guides are back; CI runs the v4.3 suite whenfull/changes.
Install
Download exam-cram-coach-flash.zip below and unzip it into your agent's skills folder (~/.claude/skills/, ~/.codex/skills/, ~/.cursor/skills/, ~/.codeium/windsurf/skills/, ~/.gemini/antigravity/skills/ …) so that exam-cram-coach/SKILL.md and exam-cram-coach/coach.py exist, then run pip install pypdfium2. Or paste the install prompt from the README into your agent. Then say:
Use the Exam Cram Coach skill. My materials are in D:\Courses\Probability. The exam is in 3 days, I am starting from zero, teach me in English from chapter 1.
Measured
| Course | Files | setup |
Chapters | Questions paired | Figures |
|---|---|---|---|---|---|
| MIT 6.006 (OCW) | 8 PDFs | 1.4 s | 6/6 | 9/9 | 34 + 10 question/answer crops |
| Open Yale PSYC 110 | 4 Markdown | 0.1 s | 4/4 | — | — |
| EEC 160 Applied Probability (private, with scanned handwritten homework) | 27 PDFs, 1000 pages | 9 s | 9/9 | 89/89 | 235 + 16; 140 scanned pages skipped |
Versus v4.3: skill text 140 KB → 6.5 KB; script calls before teaching 5 → 1; tests 12 min → 1 s; code 144k → 2.5k lines.
Relation to v4.3
Flash removes the full/lightweight dual mode, .ingest/ ledgers, HTML/PDF study guides and page QA, crop receipts, LangGraph/OpenAI/MinerU/Docling adapters, the benchmark matrix, three-layer language packs, the 3×4 mode matrix and knowledge windows. Users who want printable study guides can keep using the full edition from full/ or the v4.3 release. The two editions do not share workspaces; run setup again after switching.
Docs: docs/feature-audit.md (feature by feature), docs/v5-refactor.md (before/after), docs/weak-model-test.md (small-model sessions).
中文
这是什么
- 一条
setup读取 PDF / PPTX / DOCX / Markdown / TXT / HTML，切章节、配对作业与解答（hw2 (4)(1).pdf↔homework2solutions.pdf，Problem 1.3.10↔ 它的解答）、裁图、建索引——一千页也只要几秒。 - 图进对话：讲义里的矢量图和嵌入图自动裁出；题目和解答所在的印刷区域含图时单独裁出。
next列出每段的配图，quiz给题面图，check才给答案图。扫描/手写页（你自己交的作业）自动识别，绝不当答案展示。 - 七条命令跑完整个学习过程：
next / ask / quiz / check / answer / note / done。每条命令末尾都有一行 📍 进度并写明下一条命令，智能体不用记 CLI。 - 诚实标注不变：🟢 来自资料 / 🟡 AI 补充 /
⚠️ AI 生成答案。只知道教材题号时，quiz复述参考解答开头的已知条件，而不是自己编题。 - 小模型实测：Gemini flash-lite、Gemini flash（Antigravity）和 Claude Haiku 4.5 在一门 1000 页的真实课程上只凭 SKILL.md 就完成 setup → next → quiz → check，引用页码、保留标签、嵌入裁好的图，没有编造命令。
5.1.1 新增
export [--to 目录] [--qid …] [--chapter N] [路径…]：把上一条命令列出的图复制到聊天界面能显示的目录（VS Code / Electron 面板会拦截工作区外的file://图片）并打印相对路径；SKILL 写明三级展示策略。- 每条命令末尾的 📍 进度行 +
plan [--days N]（按剩余天数排章节、最后一天留给错题和小抄）+status显示今天目标——长会话和新对话都不会丢线。 - v4.3 完整版原样恢复到
full/（自带 README、SKILL、测试）；原版贡献指南恢复；CI 在full/变更时运行 v4.3 测试。
安装
下载下方的 exam-cram-coach-flash.zip，解压到智能体的技能目录（~/.claude/skills/、~/.codex/skills/、~/.cursor/skills/、~/.codeium/windsurf/skills/、~/.gemini/antigravity/skills/ 等），得到 exam-cram-coach/SKILL.md 和 exam-cram-coach/coach.py，再运行 pip install pypdfium2。也可以把 README 里的安装提示词直接发给智能体。然后说：
用期末极速备考教练。资料在 D:\课程\概率论，三天后考试，我零基础，用中文从第 1 章开始讲。
实测
| 课程 | 文件 | setup |
章节 | 题答配对 | 裁图 |
|---|---|---|---|---|---|
| MIT 6.006（OCW） | 8 个 PDF | 1.4 秒 | 6/6 | 9/9 | 34 张 + 10 张题面/答案图 |
| 耶鲁 PSYC 110 | 4 个 Markdown | 0.1 秒 | 4/4 | — | — |
| EEC 160 应用概率（私有，含手写扫描作业） | 27 个 PDF，1000 页 | 9 秒 | 9/9 | 89/89 | 235 + 16 张；140 页扫描件跳过 |
与 v4.3 相比：技能文本 140 KB → 6.5 KB；开讲前脚本调用 5 → 1；测试 12 分钟 → 1 秒；代码 14.4 万行 → 2500 行。
与 v4.3 的关系
Flash 版删除了完整建库/轻量按需双模式、.ingest/ 代际账本、HTML/PDF 讲义与逐页 QA、crop 回执、LangGraph/OpenAI/MinerU/Docling 适配、benchmark 矩阵、三层语言包、3×4 模式矩阵和知识点窗口。需要可打印讲义的用户可以继续使用 full/ 目录里的完整版或 v4.3 发布。两版工作区互不兼容，切换后重新 setup。
文档：docs/feature-audit.md（逐项功能核查）、docs/v5-refactor.md（前后对比）、docs/weak-model-test.md（小模型实测）。