The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides instructions for creating AI agent skills that interact with external data (e.g., PDF rotation, database queries). This creates an attack surface where malicious input in those external files could influence agent behavior. However, this is inherent to the skill's purpose as a developer guide for building functional tools.
Ingestion points: User-defined skill names/paths and files processed by example utility scripts (e.g., scripts/rotate_pdf.py).
Boundary markers: The guide does not explicitly instruct on the use of boundary markers for external data.
Capability inventory: Local file system operations (mkdir, touch) and execution of bundled Python/Bash scripts.
Sanitization: The guide focuses on procedural reliability via scripts but does not detail input sanitization for those scripts.
[DYNAMIC_CONTEXT_INJECTION]: The skill documentation addresses a known platform behavior involving the dynamic context injection syntax. It explicitly warns developers to avoid using this syntax in SKILL.md to prevent unintended execution during skill load. The documentation itself avoids triggering the behavior by using descriptive text rather than literal payloads.
[COMMAND_EXECUTION]: The skill references standard shell commands (mkdir, touch, cc) used in the creation and testing of plugin skills. These commands are localized to the developer's environment and are necessary for the documented workflow.

skill-development