The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests external 'golden cases' (untrusted data such as prompts or files) to evaluate the behavior of target skills. This creates a surface where a malicious test case could attempt to influence the evaluation logic or the resulting skill revisions.
Ingestion points: 'Golden cases' provided as input in SKILL.md.
Boundary markers: The skill mandates 'blind' runs to isolate the runner agent from the judge's context, though it does not explicitly define prompt delimiters for the inputs.
Capability inventory: The skill can invoke subagents for testing and grading, and possesses file-write capabilities for revising target skills.
Sanitization: No explicit content sanitization is mentioned, relying instead on environment isolation and 'blind' context management.
[COMMAND_EXECUTION]: The workflow involves shell interaction via git status to identify and clean up files that may have been leaked by the skill under test. This is an expected integrity check for a developer-oriented evaluation tool.

eval-skills