The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to interact with external binaries, including soffice (for document validation) and git (for redlining comparison). These calls are constructed using lists for arguments, which is a secure method that prevents shell-based command injection vulnerabilities.
[PROMPT_INJECTION]: Instructional steering markers, such as "MANDATORY
READ ENTIRE FILE," are used within SKILL.md and related documentation. These are functional directives designed to ensure the AI agent processes the full context of technical specifications and do not attempt to bypass core safety or ethical constraints.
[DATA_EXFILTRATION]: While the skill ingests and processes potentially sensitive document data (.docx), it employs the defusedxml library during unpacking and editing to mitigate risks associated with XML External Entity (XXE) attacks. The data processing occurs locally within the defined workspace.

docx