Teaching AI agents to wield CUDA without hallucinating

What it does — This repository is a daily-mirrored catalog of agent skills: compact instructions that teach AI coding agents how to use NVIDIA libraries and platform tools correctly. Rather than hoping an LLM guesses the right cuOpt API or DeepStream pipeline, you point it to a verified skill. The actual skill files live in their respective product repositories and sync here automatically.

The interesting bit — NVIDIA is treating its developer documentation as agent-native curriculum. Rather than optimizing for human readers, they are packaging expertise so Claude, Cursor, or Codex can consume it directly—a bet that the primary consumer of complex technical docs is becoming another AI.

Key highlights —

• Covers a sprawling range: quantum computing with CUDA-Q, medical imaging via MONAI, weather forecasting with Earth2Studio, GPU DataFrames via cuDF, and distributed training with Megatron-Core
• Skills are narrowly scoped (e.g., cuopt-numerical-optimization-api-python, dicom-series-preflight) rather than broad “learn CUDA” prompts
• Installs via the open skills CLI spec and targets specific agents including Claude Code, Cursor, Codex, and Kiro
• Governed in source product repos and mirrored here daily; NVIDIA is building the catalog in the open with a public roadmap

Caveats —

• Skills are maintained in their respective product repos and mirrored here daily, so this catalog is not the source of truth for the latest content
• The README is dominated by installation commands and a product catalog; it does not explain how skills are structured, tested, or versioned

Verdict — Worth exploring if you are building agent-based workflows on NVIDIA hardware and want your AI to stop improvising CUDA-X APIs. Skip it if you do not use agent coding tools or NVIDIA’s ecosystem.