NVIDIA SkillSpector on 4 popular Agent Skills repos: read the findings, not the score

viest_me · reddit · 2026-11-16

The author ran NVIDIA SkillSpector v2.12.0 in static-only mode (--no-llm) inside a throwaway container against four trending Agent Skills repos: anthropics/skills (20 skills, 250 findings, 112 HIGH), addyosmani/agent-skills (25 skills, 59 findings, 32 HIGH), mattpocock/skills (39 skills, 18 findings, 9 HIGH), and google/skills (only 27 of 203 skills scanned; the 176-skill cloud folder returned an empty report after 267s).

Findings fell into three groups:

Takeaways: read findings rather than scores; prioritize flags about agent config, credential files, and bundled scripts; scan hook scripts that actually execute, not just skill folders; a single static run may produce noise the LLM stage could filter (untested).

Original post →

More from coding & agent

coding & agent channel →