SkillEvaluator is an open-source framework from NVIDIA for evaluating agent skills, the folders of instructions and supporting files defined by the Agent Skills specification that extend coding agents like Claude Code and Codex. As skill libraries grow, teams need to check that a new skill is safe, well-formed, non-redundant, and actually useful before agents rely on it, and SkillEvaluator does this with a three-tier pipeline that can be entered at any stage. Tier 1 runs deterministic validation gates: schema conformance, PII scanning, license checks, quality rubrics, Unicode safety, script linting, and security scanning via the companion SkillSpector tool. Tier 2 checks overlap against an existing skill library through semantic similarity and intra-skill redundancy detection. Tier 3 performs live evaluation, generating a synthetic dataset for a skill and running it against a real agent inside a sandbox to measure whether the skill measurably changes agent behavior, using the open-source Harbor framework to power the runs. The CLI supports OpenAI, Anthropic, Amazon Bedrock, NVIDIA Build, and OpenAI-compatible endpoints. It is part of the broader NVIDIA Verified Skills pipeline and ships under Apache 2.0 as a community-supported, experimental-status project.