The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted shorthand notation (()=>) found within files the agent interacts with. This creates a vulnerability where an external contributor could embed malicious instructions in a comment block that the agent then interprets as valid technical tasks.
Ingestion points: Content located between ${language:comment} start-shorthand and ${language:comment} end-shorthand markers in any file.
Boundary markers: The skill defines specific delimiters but lacks instructions to sanitize or validate the intent of the shorthand before execution.
Capability inventory: The skill documentation explicitly grants the agent authority to perform "non-code actions" including running shell commands, creating files, and fetching data from APIs.
Sanitization: There are no mentioned mechanisms for filtering or escaping the commands derived from the shorthand notation.
[COMMAND_EXECUTION]: The skill documentation encourages the agent to translate shorthand into actual system actions.
Evidence: The Troubleshooting and Advanced Usage sections explicitly state: "Shorthand requests non-code actions (run commands, create files, fetch data)" and "Execute the requested action (run commands, create files, fetch data) and remove shorthand."
[DYNAMIC_EXECUTION]: The skill centers on script generation where the agent creates production-quality code based on high-level, potentially ambiguous natural language descriptions.
Evidence: Instructions to "Replace shorthand lines with production-quality code" and "Apply appropriate syntax for the target file type."

quasi-coder