// the take
A catalog of Agent Skills (the Claude Code / Codex SKILL.md convention) covering NVIDIA's entire product surface — CUDA libraries, Omniverse/Physical AI, Jetson, DOCA, TAO, NeMo, and dozens more. It's for developers already working inside a specific NVIDIA stack who want their coding agent to know the right API calls and flags without re-deriving them from docs each time.
Each skill ships as a package, not a loose markdown file: SKILL.md plus BENCHMARK.md and an evals/evals.json with actual test cases, so there's a way to check whether the skill helps before trusting it. Skills are signed (skill.oms.sig) and run through a content-integrity verification workflow in CI, which is more rigor than the typical skill repo that's just prompts in a folder. The CLI (npx skills add) supports multiple agent targets — Claude Code, Codex, Cursor, Kiro — so you're not locked into one tool. Sync-from-source-repo automation means skills get updated without someone hand-copying files into this repo and letting them drift.
This repo is a mirror, not the source — if a skill is wrong or missing something, you file the issue in the product's own repo, so debugging a bad skill means first figuring out which of 50+ source repos actually owns it. The catalog spans everything from cheminformatics to BlueField DPU firmware; unless you're deep in one specific NVIDIA vertical, the vast majority of these ~400 skills are noise you have to scroll past. It depends on a third-party CLI (vercel-labs/skills) pulled via npx at install time, and the README already flags a version footgun — older CLI versions silently fail to surface skills in Claude Code with no error. There's effectively no application code here despite the Python language tag; it's YAML/markdown/JSON configuration, so don't expect to read source and understand behavior — you're trusting the prompt text and the eval suite.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →