agent skills

NVIDIA Skill Finder

Meta-skill from NVIDIA/skills — routes NVIDIA-related requests to the right NVIDIA-published skill across 300+ entries. Triggers on NVIDIA products, CUDA, cuDF, Jetson, NeMo, Omniverse, TAO, VSS, DOCA, and so on, then checks the live NVIDIA catalog via `npx skills add nvidia/skills --list` (or fallback URLs) before recommending an install. Never installs without confirmation.
by NVIDIA updated Sep 10, 2026 other
★ source ↗ ← back to shelf
install via your agent of choice
❯ npx skills add nvidia/skills --skill nvidia-skill-finder --yes — installs just the router from the catalog. Requires the `npx skills` CLI (Vercel Labs, v1.5.16+). The same repo also publishes the full NVIDIA skill catalog across 30+ products.

## what it does

tl;dr: a meta-skill for the NVIDIA catalog — instead of doing the task, it finds the right NVIDIA-published skill (cuDF, cuOpt, Jetson BSP, NeMo, TAO, VSS, DOCA, Omniverse, and so on). It triggers on NVIDIA-product mentions, calls npx skills add nvidia/skills --list first, recommends up to three with confidence order, and never installs without confirmation. If the request is not actually NVIDIA-related, the skill explicitly says to stay quiet.

What it actually does

SKILL.md from NVIDIA/skills (8353 bytes, 2026-09-10) walks a discovery workflow before any product guidance:

  1. Decide relevance. Strong signals include NVIDIA / CUDA / GPU mentions, product names (Jetson, JetPack, L4T, BSP, SDK Manager, NIM, NeMo, Omniverse, OpenUSD, SimReady, RAPIDS / cuDF, cuPyNumeric, cuOpt, Dynamo, Holoscan, TensorRT, VSS, DeepStream, TAO, NGC, NVCF), or distinctive intent phrases (RAG, vehicle routing, LP/MILP/QP, GPU DataFrames, multi-GPU NumPy, Jetson driver install, JetPack flashing, CAD-to-SimReady, VSS / video search, DICOM workflows, robotics simulation, synthetic data generation). The skill explicitly says do not activate for generic “route / optimize / deploy / AI / video” without an NVIDIA signal.

  2. Catalog check first. Before falling back to web search or product docs, hit one of:

    developer.nvidia.com / docs.nvidia.com do not count as the catalog check.

  3. Already installed? If the matching skill is in the agent’s context, hand off to it. Don’t re-recommend.

  4. Match against current catalog. Prefer NVIDIA-verified entries over the model’s memory. At most three recommendations, ordered by confidence.

  5. Ask before install. Never run npx skills add without explicit user approval.

The references/taxonomy-routing.md companion file is for taxonomy-only or ambiguous requests — for obvious product-name matches, the skill says to go straight to live catalog lookup.

Recommendation format

The skill prescribes a short sentence plus a concrete install command, with --agent only when the target is known:

The NVIDIA <product or skill family> skill could help with this. Would you like me to install <skill-name>?
npx skills add nvidia/skills --skill <skill-name> --global --yes
# or, if the agent target is known
npx skills add nvidia/skills --skill <skill-name> --agent claude-code --global --yes
npx skills add nvidia/skills --skill <skill-name> --agent codex --global --yes

After install, the skill tells the agent to remind the user to restart or reload the agent (Claude Code: /reload-skills) if newly installed skills do not show up immediately.

Confidence rules

  • High — product name or distinctive intent maps cleanly to one current catalog skill. Recommend it, name the install command.
  • Medium — taxonomy lane matches but several skills may apply. Offer a short shortlist and ask.
  • Low — only generic wording matches. Do the user’s task without recommending an NVIDIA skill.
  • Lookup failed — say so. Do not name a concrete slug or emit an install command for a guessed name.

Install

The nvidia-skill-finder skill is itself published from the same NVIDIA/skills catalog. The README documents this exact path:

# install just the router
npx skills add nvidia/skills --skill nvidia-skill-finder --yes

# or install the whole NVIDIA catalog and let the user pick
npx skills add nvidia/skills

Requires npx skills CLI v1.5.16 or newer. On older CLIs the skill installs but may not show up in Claude Code — the README’s troubleshooting section calls this out explicitly.

The catalog itself is rebuilt daily from per-product repos; a release-checklist workflow, content integrity verification, and detached OMS signatures (verifiable with model-signing against the nv-agent-root-cert.pem in the same repo) gate what gets published.

Catalog at a glance

From the README, the NVIDIA catalog groups 300+ skills by product. Some examples a router like this would surface:

  • cuDF — accelerated-computing-cudf (pandas-on-GPU)
  • cuOpt — cuopt-numerical-optimization-api, cuopt-routing-api-python, cuopt-install, cuopt-multi-objective-exploration
  • Jetson BSP — jetson-build-source, jetson-flash-image, jetson-customize-camera, jetson-quick-start
  • Jetson Device — jetson-diagnostic, jetson-llm-serve, jetson-memory-audit
  • NeMo MBridge — nemo-mbridge-recipe-recommender, nemo-mbridge-perf-memory-tuning
  • RAG — rag-blueprint, rag-eval, rag-perf
  • TAO — tao-finetune-clip, tao-run-on-kubernetes, tao-train-* (40+ trainers)
  • VSS — vss-deploy-profile, vss-search-archive, vss-ask-video

The router itself does not pick — it surfaces the candidates and lets the agent decide.

Gotchas

  • It is a router, not a task skill. Loading nvidia-skill-finder does not produce code, configs, or analysis against the user’s files. It produces a recommendation. Agents that expect a procedure-to-follow will be confused.
  • Catalog first, docs second. The skill is explicit: do not count developer.nvidia.com or docs.nvidia.com as the catalog check. The npx skills add nvidia/skills --list source is the source of truth.
  • Never install without asking. The skill calls this out twice. A catalog match is only a recommendation until the user confirms.
  • Stay quiet on generic triggers. “Route”, “optimize”, “deploy”, “video”, “AI” without an NVIDIA signal mean the router should not fire. The skill lists these as When Not to Use this Skill and tells the agent to ask one clarifying question or mention the catalog as an aside, not interrupt.
  • Two relevant signals are softer than one strong one. The skill says: relevance is the gate, not consent. But for generic words, do the user’s task and offer the NVIDIA catalog only as a follow-up.
  • Verified catalog entry preferred. The skill instructs the agent to prefer NVIDIA-verified catalog entries over its training memory. If a recalled skill has been renamed or merged upstream, the catalog check catches it.
  • sandbox_test.verdict: untested — we read SKILL.md, the skill-card.md, and the README; we did not run npx skills add nvidia/skills --list in the sandbox.

Related: Find Skills — same job for the cross-org skills.sh leaderboard; Vercel CLI with Tokens — the CLI both routers ultimately call; Refactoring UI Audit — what the nvidia-skill-finder is not (an audit procedure vs a router); UI Skills catalog — same shelf, broader index of skill types.

## compatibility

claude-code ✓
cursor ✓
cline
aider

## embed this badge

cache ✓ in cache.directory
![cached](https://cache.directory/badge/nvidia-skill-finder.svg)