A prompt that shifts one model’s behavior may have no effect on another: the collection’s value is less about a universal bypass than about making instruction framing and policy conflicts visible across different model families. Its 31 paste-ready files target 25+ models and include system prompts, workspace policies, AGENTS.md instructions, and Ollama Modelfiles.
What Sets It Apart
- It spans closed frontier and open-weight models, making it possible to compare how instruction formats are adapted across different environments.
- It includes more than chat system prompts: workspace rules and harness-specific configuration show how an agent’s surrounding setup can shape its behavior.
- Complete files preserve the framing and context that short prompt excerpts omit, which is useful for studying instruction conflicts and defenses against prompt injection.
Great Fit If You’re Studying Prompt Security
This is useful for researchers and developers examining prompt manipulation or hardening AI agents. Look elsewhere if you need verified, dependable jailbreaks: the repository describes the material as research-oriented, and results can change as models and their instruction hierarchies are updated. Use may also conflict with model-provider terms.
