The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to read, audit, and convert other skill files, creating an attack surface where a malicious file being processed could influence the auditor agent. Ingestion points include the audit (Step 4) and verification (Step 6) workflows in SKILL.md. No explicit boundary markers, delimiters, or sanitization steps are defined for the content being audited. The skill's capability inventory includes file system operations (glob, grep) and file reading/writing.
[DYNAMIC_CONTEXT_INJECTION]: The reference/conventions.md file documents the platform-specific !shell command syntax for preprocessing content at skill load time. It includes a proactive and important security warning advising authors to never combine this preprocessing syntax with $ARGUMENTS to avoid command injection vulnerabilities.
[PROMPT_INJECTION]: The skill provides educational content in reference/persuasion-principles.md regarding the use of assertive prompt engineering techniques (e.g., 'Authority', 'Commitment', 'No exceptions') to improve agent compliance with rules. While these are methods designed to influence or override agent reasoning, they are presented as a methodology for enforcing engineering discipline (like TDD) and are consistent with the skill's purpose as an authoring toolkit.

writing-skills