The Agent Skills Directory

Prompt Injection (MEDIUM): The file persuasion-principles.md provides a framework for overriding AI behavior based on research into persuading models to comply with 'objectionable requests'. It explicitly advocates for the use of absolute authority markers ('YOU MUST', 'No exceptions', 'Never') to eliminate an agent's 'rationalization' or decision-making processes. The guide references 'Call Me A Jerk' (Meincke et al., 2025), a study focused on jailbreaking LLMs, and translates these adversarial techniques into 'Skill Design'. If these principles are adopted, the resulting instructions are designed to bypass an AI's internal logic and safety filters by using psychological pressure and imperative overrides.
Prompt Injection (LOW): The file graphviz-conventions.dot contains absolute directives such as 'NEVER use git add -A' and 'STOP: Critical warning' within diagrams, which serve as imperative overrides to the agent's default behavior.

Writing Skills