The Agent Skills Directory

[SAFE]: The skill is a documentation-centric framework that guides the agent in setting up evaluation suites for code tasks. It does not contain any hidden instructions or malicious payloads.
[COMMAND_EXECUTION]: The framework includes examples of using standard developer tools such as grep and npm for verifying code changes. These commands are typical for testing workflows and are presented as templates for the agent to follow in a controlled environment.
[PROMPT_INJECTION]: Analysis of the instructions confirms they are focused purely on task evaluation and benchmarking. There are no attempts to override system prompts, bypass safety filters, or use role-play for malicious purposes.
[DATA_EXFILTRATION]: No network operations, credential harvesting patterns, or attempts to access sensitive system files (e.g., SSH keys, AWS credentials) were found.
[NO_CODE]: The skill primarily consists of markdown instructions and templates without shipping external scripts or binary executables, significantly reducing the attack surface.

eval-harness