The Agent Skills Directory

[SAFE]: The skill consists of instructional content and reference documentation for designing evaluation suites. It does not contain any executable scripts, remote code downloaders, or commands that interact with sensitive system files.
[INDIRECT_PROMPT_INJECTION]: The skill guidelines involve processing external data such as agent transcripts and production logs. While this represents a potential surface for indirect prompt injection, the documentation explicitly recommends redacting and sanitizing sensitive data as a best practice. The instructions focus on measurement policy rather than automated execution of untrusted code.
[DATA_EXPOSURE]: There are no hardcoded credentials or attempts to access private user data. The skill focuses on professional benchmarking using well-known public families like SWE-bench and WebArena.

agent-evaluation