NVIDIA/skills: Verified Agent Instruction Sets for Claude Code and Codex

What Agent Skills Are and the Problem They Solve

An agent skill is a portable instruction set that tells an AI coding agent how to use a particular piece of software optimally. Without skills, an agent such as Claude Code or Codex must infer the correct API usage, configuration flags, and workflow steps from general training data, which may be outdated or incomplete for specialized NVIDIA software.

NVIDIA/skills solves this by publishing official, verified instruction sets for its products. A skill for cuOpt, for example, teaches the agent the correct Python API calls for vehicle routing, linear programming, and quadratic programming, rather than the agent guessing at the API from documentation that might have changed.

The catalog is a mirror: each skill is maintained in its own product repository and synchronized here daily via an automated pipeline. This means the source of truth for a given skill is the team that owns the product, not a centralized documentation team.

Skills work with Claude Code, Codex, Cursor, Kiro (as kiro-cli), and Snowflake CoCo (as the cortex agent target).

The Skill Catalog and What It Covers

The catalog covers several NVIDIA product areas. The AIQ section provides skills for deploying local AI-Q services and running shallow or deep research workflows. BioNeMo skills cover GPU-accelerated cheminformatics, molecular fingerprinting, protein structure prediction via OpenFold2, and model pretraining and fine-tuning through the KERMT framework. CUDA-Q skills cover quantum computing simulation, QPU access, and the @cudaq.kernel programming model.

The cuDF skills cover NVIDIA GPU DataFrames, pandas acceleration, dask-cuDF, ETL operations, CSV and Parquet I/O, and multi-GPU workloads. cuOpt covers vehicle routing, linear programming, quadratic programming, installation, and server deployment across multiple skills.

Physical AI and robotics skills include Isaac Lab (reinforcement learning for robot training), Isaac Sim (simulation environment), Isaac ROS (ROS integration), and Newton (a physics engine for robot learning). RAPIDS skills cover cuGraph (GPU graph analytics), cuML (machine learning), and the broader RAPIDS suite.

NIM (NVIDIA Inference Microservices) skills cover blueprint deployment and inference for specific models. Warp covers GPU-accelerated simulation in Python. The catalog is described as growing continuously.

Installing Skills with the npx CLI

Installation uses the skills CLI, which runs through npx without requiring a local install or cloning the repository. The basic install command prompts you to choose a skill interactively:

npx skills add nvidia/skills

To install a specific skill without prompts, pass the skill name and the --yes flag:

npx skills add nvidia/skills --skill cuopt-numerical-optimization-api --yes

To install into a specific agent, use the --agent flag:

npx skills add nvidia/skills --skill cuopt-numerical-optimization-api --agent claude-code

Multiple agents can be targeted in one command by repeating the --agent flag for claude-code, codex, cursor, and kiro-cli. In Claude Code, run /reload-skills after installing to load newly installed skills in the current session.

The CLI requires version 1.5.16 or newer. On older versions (1.5.15 and earlier), skills may install but not appear in Claude Code. Use npx skills@latest to ensure the current version.

Keeping Skills Up to Date and Browsing the Catalog

Because skills are mirrored daily from product repositories, installed versions can become stale when a product team revises or renames a skill. The update command refreshes all installed skills:

npx skills update

Running this interactively also flags skills that were removed or merged upstream, for example when several smaller skills are consolidated into one, and offers to remove stale local copies. To preview what would change before updating:

npx skills check

To see the full catalog before installing anything:

npx skills add nvidia/skills --list

The list option shows available skill names without installing. This is useful for identifying the exact skill name to pass to --skill in non-interactive pipelines such as CI setup steps or developer onboarding scripts.

How Skills Are Structured: Governance and Verification

The README describes skills as official and NVIDIA-verified. The catalog-exceptions.yml file at the root of the repository handles cases where a skill's name in the product repo differs from its catalog entry.

The repository layout includes separate directories: `.agents/` for agent-level configuration, `.claude-plugin/` for Claude Code plugin configuration, `.cursor-plugin/` for Cursor integration, and `skills/` for the actual skill content. A `versions.json` file tracks which version of each skill is currently in the catalog. A `benchmarks.json` records benchmark data.

The nv-agent-root-cert.pem file at the repository root is used for certificate verification in the installation flow. The skills themselves are organized by product name under the `skills/` directory, with each skill in its own subdirectory.

The project uses a dual license: Apache-2.0 for code and CC-BY-4.0 for documentation, as declared in the file headers and the LICENSE-APACHE and LICENSE-CC-BY-4.0 files.

Limitations: Skill Quality Depends on the Product Team

The catalog is a passive mirror. NVIDIA/skills itself does not write or maintain the skill content; that responsibility belongs to each product team. A skill for a product that has thin internal documentation will be thin. A skill that was last updated before a major API change in its product may give incorrect guidance until the product team pushes an update and the daily sync runs.

The CLI requires the skills package at version 1.5.16 or newer. For teams running pinned tool versions in CI, the version constraint is something to track explicitly.

There is no offline install path documented in the README for air-gapped environments. The npx flow requires network access to both the npm registry and the GitHub repository. The advanced install documentation at docs/advanced-install.mdx describes fallback manual copying for cases where the CLI cannot be used, but this is not the primary flow.

Maintenance, License, and Repository Activity

The last push to NVIDIA/skills was on 2026-09-25. The project is maintained by NVIDIA under Apache-2.0 for code and CC-BY-4.0 for documentation. A CHANGELOG.md documents catalog changes. A SECURITY.md and CODE_OF_CONDUCT.md are present.

The repository has no GitHub releases; the primary distribution channel is the npm-based skills CLI. The skills are versioned via the versions.json file rather than through npm package versioning.

An alternative approach to agent-assisted NVIDIA workflows is direct documentation via MCP (Model Context Protocol) servers, which expose API documentation as tool results at query time rather than as pre-installed instruction sets. Skills are static instruction sets installed at agent setup time, while MCP retrieves documentation dynamically. The tradeoff is that skills work offline once installed and do not add latency at inference time, while MCP requires a running server but can serve fresher documentation.

Editorial conclusion

NVIDIA/skills is worth installing for any engineer using Claude Code or Codex alongside NVIDIA's simulation, GPU computing, or robotics software. The skills are maintained in their respective product repositories and mirrored daily, so the quality of a given skill reflects how much the relevant product team invests in documenting its agent interface. Before relying on a skill in production, run npx skills check to confirm the installed version matches the current catalog.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository