Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add slowlyC/agent-gpu-skills --skill cuda-skillgit clone --depth 1 https://github.com/slowlyC/agent-gpu-skillsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/slowlyc/agent-gpu-skills/cuda-skill)<a href="https://agentmods.dev/skills/slowlyc/agent-gpu-skills/cuda-skill"><img src="https://agentmods.dev/badge/skills/slowlyc/agent-gpu-skills/cuda-skill.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00127 | $0.01643 |
| Opus 5 | $0.00063 | $0.00822 |
| Sonnet 5 | $0.00025 | $0.00329 |
| Haiku 4.5 | $0.00013 | $0.00164 |
Grade A, and why
cuda-skill scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 7d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 156 lines — stays where its author put it; the contents beside it link to each section on GitHub.
NVIDIA CUDA Reference
Use this skill as the source of truth for CUDA, PTX, NVIDIA GPU architecture, and NVIDIA profiling or debugging tools. Prefer the local official-document snapshots, then verify against NVIDIA's current online documentation when a fact is version-sensitive or absent locally.
For DSL- or library-specific implementation, use the corresponding skill first:
- Triton or Gluon kernel code: triton-skill
- CUTLASS, CuTe, or CuTeDSL code: cutlass-skill
Add this skill when those tasks require CUDA API, PTX ISA, architecture, or NVIDIA tool facts.
Locate the references
Resolve the directory containing this SKILL.md, then use its references/ child. Do not assume a Cursor, Claude, or Codex-specific install path.
In examples below, set a task-scoped variable to the resolved absolute path:
CUDA_REFS=/absolute/path/to/cuda-skill/references
Read MANIFEST.md before making version claims. It records the snapshot version, source URL, and document inventory.
Source routing
| Question | Primary source |
|---|---|
| PTX syntax, semantics, ISA or target requirements | ptx-docs/ |
| CUDA Runtime functions, errors, and structs | cuda-runtime-docs/ |
| CUDA Driver functions, contexts, modules, VMM | cuda-driver-docs/ |
| CUDA programming model and feature behavior | cuda-guide/ |
| General CUDA optimization guidance | best-practices-guide/ |
| Nsight Compute metrics, sections, and CLI | ncu-docs/, ncu-guide.md |
| Nsight Systems tracing and CLI | nsys-docs/, nsys-guide.md |
| Correctness tools and cuda-gdb | debugging-tools.md |
| NVTX instrumentation | nvtx-patterns.md |
| Frequent performance mistakes | performance-traps.md |
The short guide files are search maps, not substitutes for the full official snapshots.
Query workflow
Start with file discovery. Do not load a large chapter or the whole specification when a focused page exists.
# Discover focused PTX pages.
rg -l -i 'wgmma\.mma_async' "$CUDA_REFS/ptx-docs"
# Read the relevant lines with context.
rg -n -C 12 'Target ISA Notes|PTX ISA Notes|wgmma\.mma_async' \
"$CUDA_REFS/ptx-docs/9-instruction-set"
What ships with it
60 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- references/best-practices-guide/1-overview.md 494 B
- references/best-practices-guide/10-memory-optimizations.md 59 KB
- references/best-practices-guide/10.1-data-transfer-between-host-and-device.md 14 KB
- references/best-practices-guide/10.2-device-memory-spaces.md 44 KB
- references/best-practices-guide/10.3-allocation.md 260 B
- references/best-practices-guide/10.4-numa-best-practices.md 812 B
- references/best-practices-guide/11-execution-configuration-optimizations.md 15 KB
- references/best-practices-guide/11.1-occupancy.md 5.1 KB
- references/best-practices-guide/11.2-hiding-register-dependencies.md 660 B
- references/best-practices-guide/11.3-thread-and-block-heuristics.md 3.5 KB
- references/best-practices-guide/11.4-effects-of-shared-memory.md 1.5 KB
- references/best-practices-guide/11.5-concurrent-kernel-execution.md 1.3 KB
- references/best-practices-guide/11.6-multiple-contexts.md 2.4 KB
- references/best-practices-guide/12-instruction-optimization.md 24 KB
- references/best-practices-guide/12.1-arithmetic-instructions.md 22 KB
- references/best-practices-guide/12.2-memory-instructions.md 932 B
- references/best-practices-guide/13-control-flow.md 2.9 KB
- references/best-practices-guide/13.1-branching-and-divergence.md 1.7 KB
- references/best-practices-guide/13.2-branch-predication.md 1.2 KB
- references/best-practices-guide/14-deploying-cuda-applications.md 846 B
- references/best-practices-guide/15-understanding-the-programming-environment.md 4.8 KB
- references/best-practices-guide/15.1-cuda-compute-capability.md 1.6 KB
- references/best-practices-guide/15.2-additional-hardware-data.md 727 B
- references/best-practices-guide/15.3-which-compute-capability-target.md 496 B
- references/best-practices-guide/15.4-cuda-runtime.md 1.6 KB
- references/best-practices-guide/16-cuda-compatibility-developers-guide.md 27 KB
- references/best-practices-guide/16.1-cuda-toolkit-versioning.md 3.0 KB
- references/best-practices-guide/16.2-source-compatibility.md 1.0 KB
- references/best-practices-guide/16.3-binary-compatibility.md 6.5 KB
- references/best-practices-guide/16.4-cuda-compatibility-across-minor-releases.md 15 KB
- references/best-practices-guide/17-preparing-for-deployment.md 15 KB
- references/best-practices-guide/17.1-testing-for-cuda-availability.md 2.2 KB
- references/best-practices-guide/17.2-error-handling.md 1.6 KB
- references/best-practices-guide/17.3-building-for-maximum-compatibility.md 2.8 KB
- references/best-practices-guide/17.4-distributing-the-cuda-runtime-and-libraries.md 8.8 KB
- references/best-practices-guide/18-deployment-infrastructure-tools.md 5.3 KB
- references/best-practices-guide/18.1-nvidia-smi.md 2.4 KB
- references/best-practices-guide/18.2-nvml.md 1.1 KB
- references/best-practices-guide/18.3-cluster-management-tools.md 355 B
- references/best-practices-guide/18.4-compiler-jit-cache-management-tools.md 765 B
- references/best-practices-guide/18.5-cuda_visible_devices.md 643 B
- references/best-practices-guide/19-recommendations-and-best-practices.md 2.3 KB
- references/best-practices-guide/19.1-overall-performance-optimization-strategies.md 2.2 KB
- references/best-practices-guide/2-preface.md 8.2 KB
- references/best-practices-guide/2.1-who-should-read-this-guide.md 890 B
- references/best-practices-guide/2.2-assess-parallelize-optimize-deploy.md 4.3 KB
- references/best-practices-guide/2.3-recommendations-and-best-practices.md 1.4 KB
- references/best-practices-guide/2.4-assessing-your-application.md 1010 B
- references/best-practices-guide/20-nvcc-compiler-switches.md 1.1 KB
- references/best-practices-guide/20.1-nvcc.md 1.1 KB
- references/best-practices-guide/21-notices.md 5.2 KB
- references/best-practices-guide/21.1-notice.md 4.8 KB
- references/best-practices-guide/21.2-opencl.md 99 B
- references/best-practices-guide/21.3-trademarks.md 254 B
- references/best-practices-guide/3-heterogeneous-computing.md 7.7 KB
- references/best-practices-guide/3.1-differences-between-host-and-device.md 2.5 KB
- references/best-practices-guide/3.2-what-runs-on-a-cuda-enabled-device.md 4.5 KB
- references/best-practices-guide/4-application-profiling.md 7.7 KB
- references/best-practices-guide/4.1-profile.md 7.7 KB
- references/best-practices-guide/5-parallelizing-your-application.md 954 B
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 7d ago First seen · 156 lines · 127 tokens per session scan A bbfd1c56b6d2
cuda-skill is a skill published in the GitHub repository slowlyC/agent-gpu-skills (165 stars, last pushed 29d ago), licensed MIT. It adds 127 tokens to every session and 1,643 once invoked, about $0.0006 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
cuopt-numerical-optimization-api
LP, MILP, and QP (beta) with cuOpt — Python, C, and CLI. Use when the user is solving LP, MILP, or QP with any cuOpt interface.
cuopt-developer
Modify, build, test, debug, and contribute to NVIDIA cuOpt (C++/CUDA, Python, server, CI). Use for solver internals, PRs, DCO, and code conventions.
huawei-cloud-ascendc-operator-performance-optim
Develop and optimize custom operators using AscendC programming language. Analyze operator performance bottlenecks and conduct optimization validation. Based on AscendC and CANN toolkit Use this skill when the user wants to: (1) optimize performance-critical operators on Ascend NPU, (2) develop custom operators for…
cuda-c-basics
A beginner’s guide to CUDA C, the C/C++ programming model for running work in parallel on NVIDIA GPUs. It covers kernels, threads, blocks, grids, and the GPU’s memory types.
torchtalk-analyzer
Analyze PyTorch internals across Python, C++, and CUDA layers using the TorchTalk MCP server. Use when asked about how PyTorch operators work internally, where functions are implemented, what would break if code is modified, or finding tests for PyTorch operators.
torchtalk-trace
Trace a PyTorch function's cross-language binding chain (Python -> C++ -> CUDA).