The Agent Skills Directory

[SAFE]: The skill instructions define a process for evaluating other skills in isolated environments. It mandates 'blind' runs where subagents are denied access to the main conversation's context, which is a security-positive design pattern.
[SAFE]: The workflow includes explicit cleanup steps, such as monitoring 'git status' and using throwaway sandbox directories, to identify and remove any side effects or leaks created during the execution of target skills.
[SAFE]: No evidence of prompt injection, data exfiltration, obfuscation, or unauthorized remote code execution was detected within the instructions.

eval-skills