Cuda Skill

What it does

When your agent uses it

• Direct CUDA C++
• For framework tasks only when they need NVIDIA ISA
• Include inline PTX
• Fabric operations

Example prompts

• “/cuda-skill”

Requirements

• Python 3

What it can do on your machine

Read from SKILL.md and the folder at commit ae02d07. It shows what the files ask for, not the result of running them.

• Tool permissions

  Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.

  From allowed-tools in the SKILL.md frontmatter.

• Runs code

  No scripts in the folder and no shell commands in SKILL.md.

  From the folder's file list and the shell code blocks in SKILL.md.

• Network

  No URLs in SKILL.md.

  From URLs in SKILL.md, links to its own repository left out.

• Credentials

  Names no API keys, tokens, secrets or passwords.

  From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.

Context cost

Cuda Skill loads about 1.8k tokens when it runs, and up to ~2M if it reads all its reference files. Until then it costs about 123 tokens; SKILL.md has 812 words of instructions outside code blocks.

Always · name and description, kept in context so the agent knows when to use it

~123

When it runs · the whole SKILL.md, loaded when a task matches

~1.8k

With references · SKILL.md plus every file in references/, read only if the agent opens them

~2M

Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.

Safety

Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.

SKILL.md

The full file from slowlyC/agent-gpu-skills at commit ae02d07, republished under its MIT licence (© slowlyC). 812 words, ~1,824 tokens.

name

cuda-skill

description

Query current NVIDIA CUDA, PTX ISA, Runtime API, Driver API, Programming Guide, Best Practices, Nsight Compute, and Nsight Systems references. Use for direct CUDA C++ or PTX work, and for framework tasks only when they need NVIDIA ISA, API, architecture, or tool facts. Triggers include inline PTX, WMMA, WGMMA, TMA, tcgen05, mbarrier, fabric operations, CUDA APIs and Graphs, memory ordering, compute capability, Ampere, Hopper, Blackwell, Rubin, nsys, ncu, and compute-sanitizer.

CUDA_REFS=/absolute/path/to/cuda-skill/references

# 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"

# Runtime and Driver API lookup.
rg -l 'cudaStreamSynchronize' "$CUDA_REFS/cuda-runtime-docs"
rg -l 'cuMemMap' "$CUDA_REFS/cuda-driver-docs"

# Programming and optimization concepts.
rg -l -i 'thread block cluster' "$CUDA_REFS/cuda-guide"
rg -l -i 'coalesc' "$CUDA_REFS/best-practices-guide"

PTX ISA version
  → virtual target accepted by the assembler
  → toolkit/compiler support
  → physical GPU capability

rg -n -C 20 'cudaErrorInvalidValue' "$CUDA_REFS/cuda-runtime-docs"
rg -n -C 25 'cudaLaunchKernelEx' "$CUDA_REFS/cuda-runtime-docs"
rg -n -C 25 'cuCtxCreate' "$CUDA_REFS/cuda-driver-docs"
rg -n -C 25 'cuMemCreate|cuMemMap' "$CUDA_REFS/cuda-driver-docs"

compute-sanitizer --tool memcheck ./program
compute-sanitizer --tool racecheck ./program
compute-sanitizer --tool initcheck ./program
compute-sanitizer --tool synccheck ./program

nsys profile -o report ./program
nsys stats report.nsys-rep --report cuda_gpu_kern_sum

ncu --list-sets
ncu --list-sections
ncu --query-metrics
ncu --kernel-name regex:myKernel --launch-count 1 -o report ./program

cd /path/to/agent-gpu-skills
uv run scripts/scrape_docs.py all \
  --output-dir /tmp/cuda-docs-staging \
  --force

diff -qr skills/cuda-skill/references/ptx-docs \
  /tmp/cuda-docs-staging/ptx-docs

python3 scripts/validate_cuda_skill.py

© slowlyC, MIT. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file

Files

SKILL.md and 898 other files (references) in skills/cuda-skill of slowlyC/agent-gpu-skills.

• SKILL.md
• references/MANIFEST.md
• references/best-practices-guide/1-overview.md
• references/best-practices-guide/10-memory-optimizations.md
• references/best-practices-guide/10.1-data-transfer-between-host-and-device.md
• references/best-practices-guide/10.2-device-memory-spaces.md
• references/best-practices-guide/10.3-allocation.md
• references/best-practices-guide/10.4-numa-best-practices.md
• references/best-practices-guide/11-execution-configuration-optimizations.md
• references/best-practices-guide/11.1-occupancy.md
• references/best-practices-guide/11.2-hiding-register-dependencies.md
• references/best-practices-guide/11.3-thread-and-block-heuristics.md
• references/best-practices-guide/11.4-effects-of-shared-memory.md
• references/best-practices-guide/11.5-concurrent-kernel-execution.md
• references/best-practices-guide/11.6-multiple-contexts.md
• references/best-practices-guide/12-instruction-optimization.md
• references/best-practices-guide/12.1-arithmetic-instructions.md
• references/best-practices-guide/12.2-memory-instructions.md
• references/best-practices-guide/13-control-flow.md
• … and 880 more

Open the folder on GitHubat commit ae02d07