NVIDIA's experimental SkillEvaluator gives developers a three-stage way to assess AI-agent skills for security, specification fit, overlap and task performance, with command-line, continuous-integration and agent integrations.
NVIDIA's repository and documentation describe one stage that examines a skill's structure and security and checks whether it follows the Agent Skills specification. Another looks for overlapping guidance across skills. The final stage evaluates task results in Harbor sandboxes, comparing agent performance with a skill enabled against performance without it.
Developers can run the tool from a command line or include it in continuous-integration workflows. They can also add the SkillEvaluator skill to Claude Code or Codex. NVIDIA marks the project experimental and releases it under Apache License 2.0.


