The Agent Skills Directory

[PROMPT_INJECTION]: The instruction 'The request or $ARGUMENTS is authoritative' explicitly tells the agent to treat external, potentially untrusted input as the primary source of truth, which could be exploited to bypass other instructions or safety constraints.
[INDIRECT_PROMPT_INJECTION]: The skill processes external 'source material' (modules, incidents, facts) and user-provided arguments to generate explanations, creating a surface where malicious instructions embedded in that material could influence agent behavior.
Ingestion points: Uses $ARGUMENTS and external source material for explainers.
Boundary markers: The skill lacks explicit prompt delimiters for external content, though it instructs the agent to separate facts and inference.
Capability inventory: Generates HTML/JavaScript artifacts, writes to the file system, and opens files in a browser or previewer.
Sanitization: The skill includes an instruction to 'Escape untrusted text', though the efficacy depends on the agent's implementation.
[DYNAMIC_EXECUTION]: The skill is designed to generate HTML and JavaScript artifacts at runtime based on external data and then 'Open it in an isolated browser or host preview', creating a risk associated with executing dynamically generated code.
[COMMAND_EXECUTION]: The instructions suggest using the mktemp utility to create temporary directories, which involves the execution of shell commands on the host system.

eli5