Exam Cram Coach
Turn your slides, homework, and past papers into a source-aware tutor that remembers your progress.
English · 中文
Teach from your materials · Show figures before solving · Explain key questions step by step · Keep progress across chats
Give the agent your course folder, then say how soon the exam is, where you want to start, and which reply language you want. It teaches each concept beside the matching examples from lectures, homework, quizzes, and practice exams. It explains which formula applies, how the values fit, and why the answer follows.
Its most important feature is visible provenance:
- 🟢 From your materials: traceable to the original file and location;
- 🟡 AI-supplemented — may differ from what your teacher taught;
- ⚠️ AI-generated answer — not from your teacher or textbook.
When the materials do not support a conclusion, the agent should say so instead of pretending to know.
Start studying in one minute
- Install the skill.
- Put your slides, notes, homework, solutions, quizzes, and practice exams in one materials folder.
- Copy this message to the agent:
Use Exam Cram Coach to review D:\Course Materials. I am starting from zero, my exam is tomorrow, reply in English, and begin with Chapter 1. Use lightweight on-demand mode.
- The agent shows the absolute materials/workspace path, study mode, time budget, reply language, and processing choice before teaching.
You do not need to learn any commands first. Commands below are only for automation or troubleshooting.
Choose a processing mode
The skill asks once. If you are unsure, choose lightweight on-demand.
| Mode | What it does | Best for | Limitations |
|---|---|---|---|
| Lightweight on-demand (default, recommended) | Processes only the pages you are studying now, inspects their visuals, teaches in chat, and saves progress and notes | An exam tomorrow, a large folder, or anyone who wants to start quickly | Does not organize the whole course in advance; one active batch has at most 8 primary pages; does not create a complete Study Guide or printable PDF; without an unchanged standard bank that existed before initialization, a chapter is capped at covered_unverified |
| Full knowledge-base build | Organizes the whole course into a chapter wiki, standard question bank, and review queue; can later produce chapter Study Guides | Large or messy courses, systematic review, and students willing to wait longer | Initial processing is slower and uses more disk; scans, complex layouts, and question-answer pairing can still need AI or human review; a completed build does not prove every detail was recognized correctly |
Lightweight mode saves the cost of processing the whole course up front. It does not shorten the explanation shown to the student.
Processing and output are separate choices:
- Chat is the default output: teaching appears in the conversation and notebook, without automatic HTML/PDF work.
- Visual output creates a web Study Guide and can produce a printable PDF.
- In lightweight mode, a saved visual preference stays dormant. Switch explicitly to full mode before building a Study Guide.
How it teaches a chapter
Each chapter continues until its concepts and matching examples have been covered:
- Explain the concept in everyday language without assuming prior knowledge.
- Show only the relevant question text and figures, not a full page of unrelated items.
- State what the problem asks and what values can be read from the figure.
- Explain why the formula or rule applies.
- Substitute values and work through every step.
- Explain in beginner-friendly language why the answer follows.
- Link the concept, question, and answer back to their original file and location.
If a question or solution depends on an image, the image must be visibly rendered before it is used. Questions with missing required figures are not served. Tree, traversal, state-machine, and similar problems are computed deterministically before a diagram is drawn.
Quizzes come only from the workspace's standard question bank; temporary AI-written questions are never passed off as course quizzes. Wrong answers, skipped items, and “why/how” questions are saved for review.
Ordinary features and extensions
Ordinary features are already on: lightweight page-by-page teaching, visual inspection, full walkthroughs, saved progress, and notes. Extensions are optional. The score below is a practical recommendation, not a quality claim: 5/5 = use whenever the host supports it; 4/5 = enable when the use case matches; 3/5 = only for a specific problem; 1/5 = usually leave off.
| Extension | Recommendation | Default | Copyable prompt | Inherent limitations |
|---|---|---|---|---|
| Full wiki, standard bank, and source-conflict review | 4/5 | Off | Copy prompt ① below | Slower first run; review items must be resolved or reported. Skip it when the exam is tomorrow and you only need a few pages |
| One key question per teaching turn | 4/5 | Off | Copy prompt ② below | Excellent for beginners, but slower; it changes teaching cadence rather than proving mastery |
| Web Study Guide and printable PDF | 4/5 | Off | Copy prompt ③ below | Requires complete content, clean item crops, source checks, and page-by-page visual QA; costs time and disk; unavailable in lightweight mode |
| Native isolated child agent for each answer explanation | 5/5 | On automatically only when the host proves the required capability | Copy prompt ④ below | Available only for full v2 Study Guides; hosts without verified clean-context and tool/input restrictions must use the ordinary explanation path |
| Remote MinerU or Docling parsing | 3/5 | Off | Copy prompt ⑤ below | Worth considering for difficult scans or complex layouts; never downloaded or run locally by this project; output still needs review |
| Remote LangGraph orchestration | 1/5 | Off | Copy prompt ⑥ below | Usually unnecessary because the skill already has a persistent state machine; it cannot replace course truth or source receipts |
| Dense + sparse retrieval, RRF, and reranking | 1/5 | Experimental, unavailable | Copy prompt ⑦ below | Current evidence is insufficient; extra latency, size, dependencies, and false retrievals may outweigh any benefit, so BM25 remains the default |
Copyable extension prompts
① Full knowledge-base build
Switch to full knowledge-base mode and organize the whole course into a chapter wiki and standard question bank. Before starting, tell me which files will be processed. Report every unresolved review item, missing figure, missing answer, and source conflict; do not silently skip them.
② One question per turn
From now on, teach exactly one example per turn. Show the prompt and every required figure, explain what is asked, why the formula applies, each substitution or reasoning step, and why the answer follows. Save the walkthrough to the notebook before waiting for me to continue.
③ Visual Study Guide and printable PDF
In full mode, create a visual Study Guide and printable version for the current chapter. Render formulas directly, crop question and answer images to the current item, and inspect every output page for missing figures, bad crops, mojibake, raw LaTeX, and unrelated content. Do not deliver an artifact that fails inspection.
④ Native isolated child-agent explanations
Check whether this host officially supports a fresh independent child context and can restrict its input and tools to one item. If it does, enable native isolated child-agent explanations by default for every Study Guide item, passing only the current question, any official answer, target-only crops, language, and fixed explanation instruction. Do not require a separate external API. If the capability cannot be verified, keep ordinary explanations and tell me.
⑤ Remote MinerU / Docling
These materials contain difficult scans or complex layouts. Check whether this host already has a remote MinerU or Docling integration; do not download, install, or run either locally. If available, disclose the service, exact files to upload, retention period, and privacy boundary, then wait for my separate consent.
⑥ Remote LangGraph
First explain the concrete problem that the existing study state machine cannot solve and how remote LangGraph would solve it. Enable it only if this host already provides the remote service, it will not replace local course truth, and I explicitly consent after seeing the privacy boundary. Do not install LangGraph locally.
⑦ Experimental retrieval check
Check whether a frozen real multi-course retrieval benchmark proves that dense plus sparse retrieval, fusion, and reranking pass recall, false-positive, stability, latency, and size gates. If complete evidence is missing, do not enable it and keep the default retriever.
Our recommendation: start with ordinary lightweight teaching. If you later build a full Study Guide, keep the native isolated child-agent explanation enabled when the host can actually enforce it. Add the full build and visual Guide only when you have enough time; use remote parsing only for genuinely difficult pages; normally leave LangGraph and experimental retrieval off.
If an extension is unavailable, the agent continues with ordinary features and names the limitation. It must