ZeKaiNie/universal-examprep-skill

Exam Cram Coach · 跨会话记忆与讲义溯源防幻觉的极速备考教练 | AI exam-prep tutor for Claude Code, Cursor, Codex, Antigravity: teaches from slides with page citations, crops figures, quizzes with real homework, runs smoothly on small cheap models. 期末复习·考前突击·刷题·错题本·小抄

What it is

Exam Cram Coach is an AI exam-prep tutor that installs as a "skill" into coding agents like Claude Code, Codex, Cursor, Windsurf, or Gemini CLI. You point it at a folder of course materials — lecture slides, notes, homework, past exams — and it teaches you the course chapter by chapter, shows the actual figures from your own notes, quizzes you with real homework/exam questions, tracks your mistakes, and compiles a cheat sheet.

How it works

The project is an Agent Skill (a set of instructions the agent reads) plus one small Python tool, coach.py. The agent runs commands like coach.py setup to parse your materials: it splits the course into chapters, crops figures out of PDF pages, matches homework files with their solution files (e.g. hw2.pdf with hw2solutions.pdf), and extracts questions into a question bank. During study sessions the agent teaches with every claim citing a specific file and page, shows cropped figures inline, and quizzes you from that bank. All state lives in a local study_state.json, so you can close the chat and resume later.

Key technical points

  • Anti-hallucination by design: every sentence is tagged with provenance — 🟢 (from your materials, with file p.N citation), 🟡 (AI supplement), or ⚠️ (AI-generated answer where materials have none). Out-of-scope queries exit with a non-zero status rather than fabricating an answer.
  • Persistent memory without a database: progress, quiz results, and mistakes are stored in a plain JSON file; mistakes are prioritized in spaced re-quizzing. No vector DB, no token bloat.
  • Works on weak/small models: benchmarked on Gemini Flash-Lite and Claude Haiku over 1,000+ pages of real probability notes — zero command fabrication and zero phantom citations reported.
  • Automatic figure extraction: uses pypdfium2 to crop vector figures from PDF pages, so the tutor can show the exact diagram from your lecture or solution.

Scope and limitations

It's not a standalone app — it requires a coding agent with a terminal. PDF support needs pypdfium2 (optional; other formats like PPTX/DOCX/Markdown work without it). Scanned or handwritten pages are detected and skipped, and Excel/audio files are not read. Two editions exist: Flash (lightweight, recommended, works on small models) and Full (v4.3, adds printable HTML/PDF study guides and a verification pipeline but needs a stronger model).

Related

  • Project
    Stability-AI/stable-audio-metrics

    A collection of evaluation metrics for music and audio generative models, providing tools to calculate Fréchet Distance, KL divergence, and CLAP scores for long-form stereo audio.

  • Project
    Signet-AI/signetai

    Signet AI is a local daemon that ingests documents, code, chats and web pages, builds a continuously‑updated knowledge graph, and supplies relevant memories to LLM agents via plugins (“harnesses”). It offers a web/desktop dashboard, provenance‑rich recall, secret storage, and optional telemetry.

  • Project
    intent-hq/intent

    Intent is a desktop app that coordinates AI coding agents, notes, tasks, git and terminals via a Rust backend daemon (`intentd`). All state lives locally in SQLite, accessed through JSON‑RPC over a Unix socket or secure WebSocket. The packaged app bundles the daemon; advanced users can install the daemon separately. It supports multiple AI providers, optional GitHub/Linear/Sentry integrations, and an upcoming iOS companion. No telemetry is sent; updates are pulled from GitHub releases only.

  • Project
    xiaYuTian11/maskit

    Maskit is a local privacy‑gateway that automatically masks sensitive data in prompts sent to large‑model APIs and restores the original values in the responses, offering desktop, Docker, and browser‑extension deployments.