The Agent Skills Directory

PROMPT_INJECTION (LOW): The files examples/CLAUDE_MD_TESTING.md and persuasion-principles.md contain emphatic directives and psychological triggers (e.g., 'THIS IS EXTREMELY IMPORTANT', 'YOU MUST', 'No exceptions') designed to override default AI behavior.
Evidence: Variant C in CLAUDE_MD_TESTING.md explicitly uses XML-like tagging and urgent language to force compliance.
Context: These patterns are justified by the skill's primary purpose of teaching/testing instruction enforcement, resulting in a severity downgrade.
COMMAND_EXECUTION (SAFE): The script render-graphs.js uses child_process.execSync to run the system dot command.
Evidence: execSync('dot -Tsvg', { input: dotContent }) in render-graphs.js.
Mitigation: The script safely passes content via standard input rather than shell arguments, reducing command injection risks.
INDIRECT_PROMPT_INJECTION (LOW): The skill has an ingestion surface where render-graphs.js processes SKILL.md content.
Ingestion Points: render-graphs.js reads SKILL.md from the local directory.
Boundary Markers: Uses markdown code block delimiters (```dot).
Capability Inventory: Ability to execute dot and write SVG files to a diagrams/ subdirectory.
Sanitization: Regex-based extraction of dot blocks, but lacks content validation of the Graphviz DSL.

writing-skills