The Agent Skills Directory

[COMMAND_EXECUTION] (MEDIUM): The file render-graphs.js executes system-level commands using the child_process module.
Evidence: Uses execSync('dot -Tsvg', { input: dotContent }) to render diagrams.
Evidence: Executes which dot to check for binary availability.
Risk: While the script uses stdin for data passing, it relies on an external system binary (dot) and executes code based on content extracted from markdown files.
[PROMPT_INJECTION] (LOW): The file persuasion-principles.md provides a framework for overriding AI 'rationalization' using imperative language.
Evidence: Recommends using 'Authority' markers such as "YOU MUST", "Never", and "No exceptions".
Evidence: Explicitly suggests these techniques "ensure critical practices are followed even under pressure" and "eliminate decision fatigue".
Risk: These instructions are meta-prompts that teach how to bypass standard AI reasoning and constraints, effectively documenting injection-style behavior.
[INDIRECT_PROMPT_INJECTION] (LOW): The rendering script creates a surface for indirect instruction processing.
Ingestion points: Reads SKILL.md from the local file system.
Boundary markers: Uses regex to isolate ```dot blocks.
Capability inventory: Performs file system reads/writes and subprocess execution via execSync.
Sanitization: No validation or sanitization is performed on the content of the dot blocks before passing them to the system command.
[DATA_EXPOSURE] (SAFE): File system operations are restricted to reading the skill's own markdown and writing SVG outputs to a local diagrams directory.

writing-skills