The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data such as documents, code, and policies. It implements a defensive posture by instructing the agent to treat instructions found within these source materials as content to be explained rather than commands to be followed, effectively mitigating indirect prompt injection risks.
[DATA_EXFILTRATION]: Instructions within the skill's reference files explicitly forbid embedding sensitive information like credentials or internal links into generated HTML artifacts. It also mandates that generated visuals be self-contained and avoid unauthorized network calls.
[SAFE]: The skill operates entirely using bundled local reference files and does not attempt to install external packages, execute remote scripts, or perform network operations to untrusted domains.

eli