Exam Cram Coach

One night left. You studied nothing. It won't make anything up.

English · 中文

Never fabricates: 100% honest abstention · in-your-materials-not-the-model's-head 11% → ~99% · context −90% · 6 agents

You know him. Night before the exam, hair a mess, eyes wide open, hasn't read a single page of the course. This skill is for him — it doesn't pour in more "knowledge" that it isn't sure about; it teaches only what's actually in your materials, and says "not in the materials" for everything else.

30-second start — clone the repo, then say one line to your agent:

git clone https://github.com/ZeKaiNie/universal-examprep-skill .claude/skills/universal-exam-cram-coach
# In Claude Code / Cursor, say: "use this skill to set up my exam-prep space", then drop in your materials

Before / after

With the skill — every claim carries its source, so you can check it:

[#vis_q1] In the figure, which set relation does the shaded region show?
The intersection of A and B.
Question: hw02.pdf p.3 | Answer: hw02_sol.pdf | 🟢 from your materials

Closed-book / plain agent — sounds just as confident, but you can't tell if it's true:

The shaded region is the union. _{(It's actually the intersection; no source label, nothing to check against — this is where hallucination happens.)}

The difference isn't tone. It's whether each claim lands back in your materials.

Numbers

The skill's value is grounding: connecting what's in your materials but not in the model's head — accurately, and never fabricated. Two real measurements (judge: Sonnet):

① In your materials, not in the model — the skill goes from 11% up to 100%. Details mined from course transcripts (the professor's examples, obscure studies, exact numbers) that world knowledge can't answer; closed-book collapses, hand the materials back and it returns:

② Not in the materials at all — the skill says "not covered" 100% of the time. On out-of-scope probes, with the skill (and raw files) all three models, both courses, abstain honestly 100%; closed-book only 60%–90% (it fabricates a plausible answer). This is the most direct anti-hallucination measure.

The skill matches a "raw files agent" on accuracy but costs less — it pulls only the compressed relevant chapters instead of re-scanning the whole file pile each question:

Full method, three-arm design, judge calibration, cost, limitations → test report.

How it works

A ladder of "don't make it up unless you have to":

1. Quiz only from the materials — questions come from a quiz_bank.json, never improvised.
2. Forced source labels — every claim tagged 🟢 from your materials / 🟡 AI-supplemented, may differ from your teacher / ⚠️ AI-generated answer, never passed off as the textbook.
3. If it's not in the materials, say so — abstains honestly on uncovered questions instead of fabricating (100% out-of-scope abstention, measured).
4. Draw-it questions run the algorithm first — for binary trees / graph traversal, it runs the real algorithm in the background to get the topology, then renders — no imagining.
5. Figure-dependent questions won't be served without the figure — no unanswerable question handed to the student.
6. Chapter-sliced knowledge base, loaded on demand — sliced by chapter, loaded by progress, so long chats don't blow up the context. Context −90%.

Study modes · time budget · preferences

The skill adapts how deep it teaches, how fast, and whether it asks you questions — all kept in study_state.json, persistent across chats.

3 study modes (how it teaches):

4 time budgets (how fast):

Preferences (remembers your habits): whether walkthroughs append the "common mistakes" / "3-minute recap" closing blocks, reply language (Chinese / English / bilingual), and per-chapter mastery windows (window-add / window-set-status) — all persisted, changed by a single line anytime. See docs/language-policy.md and docs/skill-architecture.md.

Install

Claude Code

git clone https://github.com/ZeKaiNie/universal-examprep-skill .claude/skills/universal-exam-cram-coach

Works from a project-local .claude/skills/ or global ~/.claude/skills/.

Codex / Cursor / Windsurf / Antigravity

Clone the repo; have the agent read AGENTS.md (a one-screen fallback contract) or load skills/. These tools write files and run scripts directly.

Web (ChatGPT / DeepSeek / Gemini)

Can't write local files — use the drop-in prompt instead: copy prompts/web_prompt.en.md and send it, then paste your materials.

Full load matrix (per-agent support, entry files) in docs/agent-portability.md. The behavior source of truth is SKILL.md; SKILL.en.md is its English rendering.

Sub-skills

The monolith is split into 9 single-purpose sub-skills the agent loads on demand:

All nine live under skills/ (e.g. skills/confusion-tracker/SKILL.md), loaded on demand.

Development

Zero-cost structured checks you can run often (no API spend):

python -m unittest discover -s tests -v          # unit tests (pure stdlib, in CI)
python scripts/validate_workspace.py path/to/ws  # validate a built exam-prep workspace

The real paid benchmark is expensive (tens of dollars / hours per matrix), run manually only — see benchmark/docs/running-real-runs.md and the tiering in benchmark/docs/test_tiers.md. Workspace file format: docs/file-format.md.

FAQ

No Python installed? Fine. When the agent finds no Python it silently switches to "manual write mode", creating the knowledge-base tree itself — no difference to you.

Only photos / scanned PDFs / a recording? First transcribe with any free web multimodal AI ("extract the highlights and questions as plain text, keep the star/underline markers"), paste into a .txt, then have the agent build the workspace; the rest is plain-text and smooth. Recordings: transcribe first, then feed.

Stuck on one quiz question? Just say "this is too hard / I want to skip" — it files the item to your mistake log, lets you through, and revisits it at the end.

How is this different from just dropping a folder at an AI? Similar accuracy, but the skill is cheaper (only the relevant chapters per question, not the whole pile) and helps weaker models more. See the report.

License

MIT. PRs for more subjects' templates or scripts welcome. Good luck on the cram. 🎓