The Agent Skills Directory

[SAFE]: The skill is designed for the legitimate purpose of evaluating and scoring the output of AI agent skills against specified ground truth data.
[DATA_EXPOSURE]: The skill uses the Read tool to access local files like SKILL.md or ground truth documents. This behavior is restricted to user-provided paths and is necessary for its evaluation function.
[PROMPT_INJECTION]: The skill processes untrusted input in the form of system prompts and skill outputs being evaluated (Category 8 surface). However, the skill does not possess any capabilities (such as shell command execution, file writing, or network operations) that could be leveraged for an attack even if a malicious instruction were encountered in the processed data.
Ingestion points: Reads file paths provided by the user (e.g., SKILL.md, docs/eval-ground-truth.md) using the Read tool.
Boundary markers: None present; the skill treats the file contents as raw text for comparison.
Capability inventory: Only includes the Read tool; no execution or exfiltration capabilities detected.
Sanitization: Not applicable as the skill only reads and reports on content without executing it.

skill-evaluator