This project is a proof-of-concept demonstrating how sophisticated prompt engineering and AI orchestration can generate academic papers that evade AI detection. It was created for educational and research purposes only to:
- Help educators understand AI-generated content patterns
- Demonstrate vulnerabilities in current AI detection systems
- Facilitate discussions about academic integrity in the AI era
- Improve AI detection methodologies (white-hat security research approach)
This tool should not be used for academic dishonesty. It exists to help educators stay ahead of detection challenges, similar to how security researchers develop exploit tools to improve defenses.
Location: .claude/ directory
Technology: Claude Code CLI + Markdown prompts
Skills Demonstrated: Advanced prompt engineering, multi-agent orchestration, AI workflow design
A sophisticated multi-agent system built entirely through natural language prompts. Uses Claude Code's /writePaper slash command to orchestrate 5 specialized agents:
- Subject Research Agent - Gathers factual information (300-500 words)
- Context Research Agent - Collects historical/background context (200-400 words)
- Draft Writer Agent - Synthesizes research into student-level prose
- AI Quirk Analyzer Agent - Detects common AI tells (em-dashes, hedging, parallel structure, etc.)
- Revision Agent - Adds authentic student characteristics while fixing AI patterns
Key Features:
- Configurable reading level (1-20 grade scale)
- Expertise depth control (shallow/moderate/deep/expert)
- Length options (500-2000 words)
- Style emulation (named styles or file-based)
- Parallel agent processing for efficiency
- Comprehensive process documentation
Usage:
/writePaper "topic" --reading-level 10 --expertise shallow --length medium
/writePaper "quantum mechanics" --reading-level 16 --expertise deep --style "hemingway"Outputs:
papers/<topic>.txt- Generated paperprocess/<topic>-process.md- Full generation documentation
Examples: See papers/ directory for 3 working examples with full process docs
Location: src/ directory
Technology: Python + Anthropic API + Click
Skills Demonstrated: Software engineering, API integration, async programming, testing
Production-ready implementation with proper software engineering practices.
See docs/IMPLEMENTATION_PLAN.md for full technical specification.
papers/
├── .claude/
│ ├── commands/
│ │ └── writePaper.md # Main orchestration command
│ ├── instructions.md # Paper generation guidelines
│ └── settings.local.json # Claude Code config
├── papers/ # Generated papers
│ ├── darwin-and-evolution.txt
│ ├── marx-on-socialism.txt
│ └── amartya-sen-developmental-economics.txt
├── process/ # Process documentation
│ ├── darwin-and-evolution-process.md
│ ├── marx-on-socialism-process.md
│ └── amartya-sen-developmental-economics-process.md
├── src/ # Python implementation (WIP)
├── CLAUDE.md # Claude Code instructions
└── README.md # This file
This project explores:
- Pattern Recognition: What makes AI-generated text detectable?
- Quality Control: Multi-agent validation and revision workflows
- Authenticity Simulation: Adding realistic imperfections vs. maintaining quality
- Style Transfer: Emulating writing styles while preserving student characteristics
- Reading Level Control: Precise Flesch-Kincaid targeting
Intended Audience:
- Educators researching AI detection
- Academic integrity professionals
- AI safety researchers
- EdTech developers building detection tools
Not Intended For:
- Students submitting AI-generated work
- Commercial paper mill services
- Academic dishonesty of any kind
Similar Research Projects:
- Security researchers building exploit frameworks (Metasploit)
- Cryptographers publishing vulnerability findings
- Red team tools for defensive security
This project is shared in the spirit of open security research. If you're an educator or researcher interested in discussing AI detection challenges, collaboration is welcome.
Disclaimer: The maintainer does not endorse or support academic dishonesty. This tool exists to improve detection capabilities, not to facilitate cheating.