This project provides an AI agent skill designed to act as a last-night exam cram coach. It addresses the critical issue of AI hallucination in educational contexts by strictly adhering to user-provided materials, transforming slides, notes, and past papers into a structured knowledge base and quiz bank. This enables students to study effectively using only their own content, ensuring accuracy and preventing fabricated information, while also offering a more cost-efficient solution compared to general-purpose AI agents processing raw files.

How It Works

The skill operates by constructing a chapter-sliced knowledge base and a quiz bank from user-supplied documents. Its core methodology employs a "ladder of don't make it up" approach, ensuring that all generated content is directly traceable to the source materials. Key features include quizzes drawn exclusively from a quiz_bank.json, forced source labels (🟢, 🟡, ⚠️) for claim verification, and a measured 100% abstention rate on out-of-scope queries. For complex questions like algorithm visualizations, it runs the actual algorithm first. The knowledge base is loaded on demand per chapter, significantly reducing context window usage by up to 90%. Study state, including modes, time budgets, and preferences, is persisted in study_state.json for a tailored learning experience.

Quick Start & Requirements

Installation involves cloning the repository into a specific agent skill directory:

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

Integration with Claude Code, Cursor, or similar agents requires initiating the skill via a natural language command (e.g., "use this skill to set up my exam-prep space") followed by dropping in study materials. For web-based agents (ChatGPT, DeepSeek, Gemini) that cannot write local files, users should copy and send the content of prompts/web_prompt.en.md and then paste their materials. No specific Python installation is required for the agent's core function, though handling scanned documents or recordings necessitates prior transcription.

Highlighted Details

• Anti-Hallucination: Rigorously tested to achieve 100% out-of-scope abstention and significantly boost the utilization of in-material information (from ~11% to ~99%).
• Cost Efficiency: Demonstrates lower cost per question compared to "Raw files agents" by processing only relevant chapters (e.g., PSYC 110: $0.117 vs. $0.102).
• Source Grounding: Every AI-generated claim is tagged with a source indicator, allowing users to verify information against their original materials.
• Modular Architecture: Comprises nine single-purpose sub-skills loaded on demand, enhancing efficiency and maintainability.

Maintenance & Community

The repository shows active development through its CI pipeline. However, the README does not detail specific community channels (like Discord or Slack), notable contributors beyond the owner, sponsorships, or an explicit roadmap. Contributions via Pull Requests for additional subject templates or scripts are welcomed.

Licensing & Compatibility

The project is released under the MIT License, which is highly permissive and allows for commercial use, modification, and distribution, including integration within closed-source applications.

Limitations & Caveats

For web-based AI agents, the inability to write local files necessitates a workaround using prompt injection. Inputting non-textual materials like photos, scanned PDFs, or audio recordings requires a preliminary transcription step using external tools before they can be processed by the skill.