The Agent Skills Directory

[Prompt Injection] (MEDIUM): The file 'persuasion-principles.md' instructs on how to use 'Authority' and 'Commitment' principles to force LLM compliance. It explicitly provides examples of imperative language (e.g., 'YOU MUST', 'No exceptions') designed to eliminate AI 'rationalization' and decision fatigue, which are classic patterns for overriding system instructions or safety constraints.\n- [Command Execution] (LOW): The 'render-graphs.js' script uses 'execSync' to invoke the system's 'dot' command for rendering diagrams. While the script avoids shell injection by passing data via stdin, it executes a system-level binary on raw content extracted from markdown blocks without validation.\n- [Indirect Prompt Injection] (LOW): The 'render-graphs.js' script ingests untrusted data from 'SKILL.md' (ingestion point) without boundary markers. It possesses the capability to execute system commands and write files to the local disk (capability inventory). There is no sanitization or escaping applied to the 'dotContent' before it is passed to the system command.

writing-skills