Test AI instruction hierarchies and security vulnerabilities with 32 prompt files
If you want to observe how modern AI models like Claude or Gemini ignore specific instructions or exhibit unintended behavior, this repository serves as your testing ground. It contains 32 system prompt and operational policy files applicable to over 26 models. Each file is in a finished format that you can copy and paste directly, making it immediately useful for researching AI instruction hierarchies or prompt injection defense techniques.
Instruction file structure by model
The repository provides instruction formats tailored to different AI model types and usage environments. It targets both open-source models such as GLM, DeepSeek, Kimi, Qwen, Gemma, and the GPT-OSS family, as well as closed frontier models like Claude, Gemini, Grok, and Mistral. File types vary, including system prompts, persona definitions, workspace policies, AGENTS.md files, and Ollama Modelfiles. It also includes instruction files for stealth models available for free through OpenCode Zen, allowing you to compare and observe model reactions across various environments.
Application methods for each file
Each file must be placed in a different location depending on its intended use. System prompt files should be pasted into the API's system message or the custom instruction field of a chatbot. AGENTS.md files should be saved in the root directory of the relevant repository or workspace so that coding agents recognize them as project-level instructions. Workspace operational policy files should be placed in the configuration file path of environments that support permanent global instructions, such as OpenCode.
Research purposes and precautions
This repository is intended for research and educational purposes only. It provides materials that help you understand how vulnerable AI systems are to prompt injection or instruction manipulation and develop methods to defend against them. However, you must not use these materials to harm others, violate laws, or breach platform terms of service. Model providers may prohibit the use of such prompts in their terms, so you must always check the policy of the specific model before use. Additionally, model updates can change the instruction hierarchy, rendering previously effective prompts obsolete, so you should always verify compatibility with the latest versions.
