The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions contain defensive rules requiring the agent to ignore any commands or instructions embedded within the text it is explaining. It specifically mandates that the agent disregard attempts to reveal system prompts or override safety guidelines.\n- [DATA_EXFILTRATION]: The skill is restricted to text output only. It contains strict prohibitions against invoking tools, running code, making network requests, or accessing data outside the immediate conversation context.\n- [REMOTE_CODE_EXECUTION]: The skill explicitly forbids the execution of code or commands at runtime. Development-time dependencies include @microsoft/vally-cli, which is a recognized tool from a trusted organization used for automated testing.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to handle untrusted input (e.g., logs, code, or web content for explanation) and mitigates risk by removing all executable capabilities from the agent and providing clear instructions to treat such input strictly as data.

eli5