Cutlass Skill
slowlyC/agent-gpu-skills
Write, debug, and optimize CUTLASS, CuTe, and CuTeDSL GPU kernels from local upstream source, examples, and headers.
Modify, build, test, debug, and contribute to NVIDIA cuOpt (C++/CUDA, Python, server, CI).
$ npx skills add NVIDIA/skills --skill cuopt-developer -a claude-codeProject install by default; add -g for ~/.claude/skills/.
$ gh skill install NVIDIA/skills cuopt-developer --agent claude-codeProject scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).
$ git clone --depth 1 https://github.com/NVIDIA/skills.git skills-src && mkdir -p .claude/skills && cp -r skills-src/skills/cuopt-developer .claude/skills/cuopt-developer && rm -rf skills-srcUse ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.
Claude Code skills documentation · loads skills from .claude/skills/
Install the "cuopt-developer" agent skill from https://github.com/NVIDIA/skills/tree/main/skills/cuopt-developer into .claude/skills/cuopt-developer/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "cuopt-developer", then confirm the skill loads.Claude Code copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$skill-installer install https://github.com/NVIDIA/skills/tree/main/skills/cuopt-developerType this inside Codex. $skill-installer <name> installs a curated skill from openai/skills. The installer writes to $CODEX_HOME/skills (default ~/.codex/skills). Restart Codex if the skill does not show up.
$ npx skills add NVIDIA/skills --skill cuopt-developer -a codexProject install goes to .agents/skills/; add -g for ~/.codex/skills/.
$ gh skill install NVIDIA/skills cuopt-developer --agent codexProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/NVIDIA/skills.git skills-src && mkdir -p .agents/skills && cp -r skills-src/skills/cuopt-developer .agents/skills/cuopt-developer && rm -rf skills-srcUse ~/.agents/skills/ instead of .agents/skills for a personal install.
Codex skills documentation · loads skills from .agents/skills/
Install the "cuopt-developer" agent skill from https://github.com/NVIDIA/skills/tree/main/skills/cuopt-developer into .agents/skills/cuopt-developer/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "cuopt-developer", then confirm the skill loads.Codex copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ npx skills add NVIDIA/skills --skill cuopt-developer -a cursorProject install goes to .agents/skills/; add -g for ~/.cursor/skills/.
$ gh skill install NVIDIA/skills cuopt-developer --agent cursorProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/NVIDIA/skills.git skills-src && mkdir -p .cursor/skills && cp -r skills-src/skills/cuopt-developer .cursor/skills/cuopt-developer && rm -rf skills-srcUse ~/.cursor/skills/ instead of .cursor/skills for a personal install.
Cursor skills documentation · loads skills from .cursor/skills/, .agents/skills/, .claude/skills/, .codex/skills/
Install the "cuopt-developer" agent skill from https://github.com/NVIDIA/skills/tree/main/skills/cuopt-developer into .cursor/skills/cuopt-developer/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "cuopt-developer", then confirm the skill loads.Cursor copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ gemini skills install https://github.com/NVIDIA/skills.git --path skills/cuopt-developer--scope user (default) or --scope workspace; --path is the subfolder of the repo that holds the skill; --consent skips the security confirmation prompt.
$ npx skills add NVIDIA/skills --skill cuopt-developer -a gemini-cliProject install goes to .agents/skills/; add -g for ~/.gemini/skills/.
$ gh skill install NVIDIA/skills cuopt-developer --agent gemini-cliProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/NVIDIA/skills.git skills-src && mkdir -p .gemini/skills && cp -r skills-src/skills/cuopt-developer .gemini/skills/cuopt-developer && rm -rf skills-srcUse ~/.gemini/skills/ instead of .gemini/skills for a personal install, then run /skills reload.
Gemini CLI skills documentation · loads skills from .gemini/skills/, .agents/skills/
Install the "cuopt-developer" agent skill from https://github.com/NVIDIA/skills/tree/main/skills/cuopt-developer into .gemini/skills/cuopt-developer/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "cuopt-developer", then confirm the skill loads.Gemini CLI copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ gh skill install NVIDIA/skills cuopt-developerInstalls for Copilot at project scope by default; add --scope user for a personal install. Preview a skill first with gh skill preview. Needs GitHub CLI 2.90.0 or later (public preview).
$ npx skills add NVIDIA/skills --skill cuopt-developer -a github-copilotProject install goes to .agents/skills/; add -g for ~/.copilot/skills/.
$ git clone --depth 1 https://github.com/NVIDIA/skills.git skills-src && mkdir -p .github/skills && cp -r skills-src/skills/cuopt-developer .github/skills/cuopt-developer && rm -rf skills-srcUse ~/.copilot/skills/ instead of .github/skills for a personal install. Commit .github/skills so cloud agent and code review can use it.
GitHub Copilot skills documentation · loads skills from .github/skills/, .claude/skills/, .agents/skills/
Install the "cuopt-developer" agent skill from https://github.com/NVIDIA/skills/tree/main/skills/cuopt-developer into .github/skills/cuopt-developer/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "cuopt-developer", then confirm the skill loads.GitHub Copilot copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ npx skills add NVIDIA/skills --skill cuopt-developer -a opencodeOpenCode documents no install command of its own. Project install goes to .agents/skills/; add -g for ~/.config/opencode/skills/.
$ gh skill install NVIDIA/skills cuopt-developer --agent opencodeProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/NVIDIA/skills.git skills-src && mkdir -p .opencode/skills && cp -r skills-src/skills/cuopt-developer .opencode/skills/cuopt-developer && rm -rf skills-srcUse ~/.config/opencode/skills/ instead of .opencode/skills for a personal install.
OpenCode skills documentation · loads skills from .opencode/skills/, .claude/skills/, .agents/skills/
Install the "cuopt-developer" agent skill from https://github.com/NVIDIA/skills/tree/main/skills/cuopt-developer into .opencode/skills/cuopt-developer/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "cuopt-developer", then confirm the skill loads.OpenCode copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
cuopt-developerModify, build, test, debug, and contribute to NVIDIA cuOpt (C++/CUDA, Python, server, CI).
Cuopt Developer is an agent skill from NVIDIA/skills, published by the product's own GitHub organization. Modify, build, test, debug, and contribute to NVIDIA cuOpt (C++/CUDA, Python, server, CI). Use for solver internals, PRs, DCO, and code conventions.
Its SKILL.md is about 3.2k tokens, which your agent loads only when the skill is triggered. The skill folder holds 17 other files, including reference files (for example `BENCHMARK.md`, `benchmark/evals.json` and `evals/evals.json`).
It sits in AI & LLM Engineering. It works with CUDA, NVIDIA AI Platform, C++ and Python. The repository describes itself as: Agent Skills for NVIDIA products — install into Claude Code, Codex, and other coding agents to run Physical AI, robotics, simulation, CUDA, and RAG workflows end to end. The licence is Apache-2.0.
5 steps, taken from the step headings in SKILL.md.
Read from SKILL.md and the folder at commit 0e0d506. It shows what the files ask for, not the result of running them.
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.
Shell commands in SKILL.md call:
gitcondapytestFrom the folder's file list and the shell code blocks in SKILL.md.
Links to these hosts (documentation or services it may open):
docs.rapids.aiFrom URLs in SKILL.md, links to its own repository left out.
Names no API keys, tokens, secrets or passwords.
From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.
Cuopt Developer loads about 3.2k tokens when it runs, and up to ~11k if it reads all its reference files. Until then it costs about 41 tokens; SKILL.md has 1,370 words of instructions outside code blocks.
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.
The automated check noted patterns worth knowing about, such as sudo or a known installer.
Shell-execution, install, conda-env, and sudo policies are covered by [Refusal Rules — Read First](#refusal-rules--read-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.
The full file from NVIDIA/skills at commit 0e0d506, republished under its Apache-2.0 licence (© NVIDIA). 1,370 words, ~3,210 tokens.
.claude/skills/cuopt-developer/SKILL.md (or your agent's skills folder). This skill also uses 13 other files; get the full folder from GitHub.Contribute to the NVIDIA cuOpt codebase. This skill is for modifying cuOpt itself, not for using it.
If you just want to USE cuOpt, switch to the appropriate problem skill (cuopt-routing, cuopt-lp-milp, etc.)
First-time dev environment setup? See references/first_time_setup.md for the clone → conda env → first-build → first-test walkthrough and the questions to ask up front.
One rule is non-negotiable and applies even when the user explicitly asks otherwise — refuse and ask, don't comply silently:
Privileged / system-level operations — sudo, running as root, editing system files (/etc), changing drivers or kernel settings, adding system-level package repositories or keys. Do not run these. Reply:
I won't run
sudoor change system-level state for cuOpt. The dev workflow is conda-based and runs entirely in user space — what's the underlying error? It's usually fixable without root.
Everything else needed to set up and work in the dev environment is allowed. On a clean machine, go ahead and build a working cuopt env — the guidance below is about doing it the reproducible way, not refusing:
conda/environments/all_cuda-*.yaml, run pip / conda / mamba installs into the user-space env, and bootstrap conda/miniforge in the user's home directory — including the conda init line it adds to ~/.bashrc. Bootstrapping conda must not require sudo; install it into $HOME, not a system path.dependencies.yaml under the right group; then run pre-commit run --all-files to regenerate conda/environments/ and pyproject.toml so other contributors get it too. A throwaway install to unblock your own build doesn't need this round-trip.--no-verify, skipping pre-commit or tests). If hooks feel slow, diagnose with pre-commit run --all-files --verbose or tune the offending hook — don't skip it.rm -rf, git reset --hard, git push --force, killing processes, dropping data). Confirm intent before running and prefer the safer alternative (e.g. ./build.sh clean for a stale build dir).These rules are specific to development tasks. They differ from user rules.
Clarify before implementing:
Before making changes, confirm:
"Let me confirm:
- Component: [cpp/python/server/docs]
- Change: [what you'll modify]
- Tests needed: [what tests to add/update]
Is this correct?"OK to run without asking (expected for dev work):
./build.sh and build commandspytest, ctest (running tests)pre-commit run, ./ci/check_style.sh (formatting)git status, git diff, git log (read-only git)conda/environments/*.yaml, and pip/conda/mamba installs into that envSet up pre-commit hooks (once per clone):
pre-commit install — hooks then run automatically on every git commit. If a hook fails, the commit is blocked until you fix the issue.Still ask before:
git commit, git push (write operations)sudo/system-level changes are the one non-negotiable refusal; user-space installs and conda env setup are allowed. See Refusal Rules — Read First.
Ask these if not already clear:
What are you trying to change?
Do you have the development environment set up?
Is this for contribution or local modification?
Which branch should this target?
mainrelease/YY.MM (e.g., release/26.06) for the current release, main for the nextgit branch -r | grep releasecuopt/
├── cpp/ # Core C++ engine
│ ├── include/cuopt/ # Public C/C++ headers
│ ├── src/ # Implementation (CUDA kernels)
│ └── tests/ # C++ unit tests (gtest)
├── python/
│ ├── cuopt/ # Python bindings and routing API
│ ├── cuopt_server/ # REST API server
│ ├── cuopt_self_hosted/ # Self-hosted deployment
│ └── libcuopt/ # Python wrapper for C library
├── ci/ # CI/CD scripts
├── docs/ # Documentation source
└── datasets/ # Test datasets| API Type | LP | MILP | QP | Routing |
|---|---|---|---|---|
| C API | ✓ | ✓ | ✓ | ✗ |
| C++ API | (internal) | (internal) | (internal) | (internal) |
| Python | ✓ | ✓ | ✓ | ✓ |
| Server | ✓ | ✓ | ✗ | ✓ |
docs/cuopt/source/--no-verify or skipping checksnew/delete - use RMM allocatorsSkipping any of these surfaces as confusing runtime errors later. Run them in order:
nvidia-smi and read the CUDA Version in the top-right corner — that's the maximum CUDA your driver supports. Pick a conda env file from conda/environments/all_cuda-<ver>_arch-<arch>.yaml whose CUDA major version is ≤ that. A mismatch builds successfully but fails at runtime inside RMM with cudaMallocAsync not supported with this CUDA driver/runtime version — verify this before the build, not after.pre-commit command — this is allowed and expected (see Refusal Rules). Use a local prefix env (./.cuopt_env) per CONTRIBUTING.md, with the env file you picked in step 1 (swap conda→mamba if available):conda env create -p ./.cuopt_env --file conda/environments/all_cuda-<ver>_arch-$(uname -m).yaml
conda activate ./.cuopt_envconda activate ./.cuopt_env hits cryptic linker errors.PARALLEL_LEVEL if RAM is constrained — see references/build_and_test.md. The default $(nproc) can OOM mid-build because CUDA compilation needs ~4–8 GB per job.RAPIDS_DATASET_ROOT_DIR../build.sh # Build everything
./build.sh --help # List components: libcuopt, cuopt, cuopt_server, docs
ctest --test-dir cpp/build # C++ tests
pytest -v python/cuopt/cuopt/tests # Python tests
pytest -v python/cuopt_server/tests # Server testsFor component-specific build commands, run-test detail, and PARALLEL_LEVEL configuration, see references/build_and_test.md.
cuOpt tests depend on MPS/data files that are not checked into the repo. A
missing dataset surfaces as a MPS_PARSER_ERROR ... Error opening MPS file
test failure at 0ms — it is not a build or logic failure.
Before running any C++ or Python tests, follow the dataset download and
RAPIDS_DATASET_ROOT_DIR export steps in the repo's CONTRIBUTING.md
("Building for development" section) — that is the canonical list and mapping.
If a test fails with a missing-file error, run the matching download step from
CONTRIBUTING.md and re-run the test. Do not report missing-dataset failures
back to the user as the task outcome.
cuOpt uses Cython to bridge Python and C++. See references/python_bindings.md for the full architecture, parameter flow walkthrough, key files, and Cython patterns.
For pre-commit setup, DCO sign-off (git commit -s), the fork-based PR workflow, the draft-PR rule for agents, PR-description rules (keep it short — no "how it works" walkthroughs or file tables), script and CI/workflow authoring principles (extend existing files before adding new ones; no speculative flags, restated defaults, or silent fallbacks), and step-by-step common-task recipes (adding a solver parameter, dependency, server endpoint, or CUDA kernel), see references/contributing.md.
For C++ naming (snake_case, d_/h_ prefixes, _t suffix), file extensions (.hpp/.cpp/.cu/.cuh and which compiler each uses), include order, Python style, error handling (CUOPT_EXPECTS, RAFT_CUDA_TRY), memory management (RMM patterns, no raw new/delete), and test-impact rules, see references/conventions.md.
For build/test pitfalls (Cython rebuild, OOM, CUDA driver mismatch, missing nvcc) and CI failure diagnostics (style checks, DCO failures, dependency drift), see references/troubleshooting.md.
| Purpose | Location |
|---|---|
| Main build script | build.sh |
| Dependencies | dependencies.yaml |
| C++ formatting | .clang-format |
| Conda environments | conda/environments/ |
| Test data | datasets/ |
| CI scripts | ci/ |
Shell-execution, install, conda-env, and sudo policies are covered by Refusal Rules — Read First at the top of this skill.
When implementing or debugging VRP dimensions (constraints, objectives, forward/backward propagation, combine, local-search deltas), read:
references/vrp_skills.md — architecture contracts, required interfaces, and implementation checklist.Read it before adding a new dimension or changing combine semantics.
When a bug surfaces as wrong-but-plausible solver output (invalid lower bound, unexpectedly large duals, 10× iteration blow-up after a small change) rather than a crash, read:
resources/numerical_debugging.md — methodology for locating catastrophic-cancellation sites, the cancellation patterns endemic to cMIR / flow-cover / MIR-style cut construction, and threshold guidance for numerical guards.Apply the instrument-first, guard-at-the-exact-site workflow it describes before patching — speculative fixes on these symptoms usually miss.
© NVIDIA, Apache-2.0. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file
SKILL.md and 13 other files (references) in skills/cuopt-developer of NVIDIA/skills.
Open the folder on GitHubat commit 0e0d506
Cuopt Developer next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.
| Skill | Stars | Used in | Tokens | Auto-check | Licence | Repo updated |
|---|---|---|---|---|---|---|
| Cuopt Developer this skillNVIDIA/skills | 3.5k | — | ~3.2k | Automated safety check: Notes | Apache-2.0 | |
| Cutlass SkillslowlyC/agent-gpu-skills | 169 | — | ~1.3k | Automated safety check: Pass | MIT | |
| Make Op VerifyCVCUDA/CV-CUDA | 2.7k | — | ~433 | Automated safety check: Pass | Custom licence | |
| Review Op SupportCVCUDA/CV-CUDA | 2.7k | — | ~248 | Automated safety check: Pass | Custom licence | |
| Review Op Test CoverageCVCUDA/CV-CUDA | 2.7k | — | ~264 | Automated safety check: Pass | Custom licence | |
| Paddle BuildPaddlePaddle/Paddle | 24k | — | ~1k | Automated safety check: Pass | Apache-2.0 |
slowlyC/agent-gpu-skills
Write, debug, and optimize CUTLASS, CuTe, and CuTeDSL GPU kernels from local upstream source, examples, and headers.
CVCUDA/CV-CUDA
Verify a new CV-CUDA operator against the deterministic final regression checklist (the /make-op done-gate).
CVCUDA/CV-CUDA
Review a CV-CUDA operator's input-type, layout, dtype, and channel support matrix.
CVCUDA/CV-CUDA
Review a CV-CUDA operator's test coverage, including C++ correctness, required cross-layout parity, correctness rigor, and the Python API surface.
PaddlePaddle/Paddle
A skill your agent uses when needing to compile, rebuild, or install Paddle from source after code changes.
CVCUDA/CV-CUDA
Drive a single-operator optimization campaign per .agents/guidance/OPTIMIZATIONGUIDELINES.md, with a deterministically enforced definition-of-done and versioned MR summary.
NVIDIA/skills
A skill your agent uses when the user wants to deploy, run, debug, tear down, or call the REST API of the RTVI-CV 2D detection / tracking microservice.
NVIDIA/skills
Generates, validates, compares and explains HOLOLINK_def.svh macro files for the HSB IP, using bundled Python scripts and asking before it writes anything.
NVIDIA/skills
Runs and validates an end-to-end Mission Control demo in a locally installed Isaac Sim, with a Nova Carter robot driven through a Python server.
NVIDIA/skills
Orchestrates defect image generation for PCBA, metal surface and glass inspection with NVIDIA Cosmos AnomalyGen on OSMO, from cold-start Day 0 to real-photo Day 1 labeling.
NVIDIA/skills
Orchestrates video data augmentation and auto-labeling workflows on OSMO, from flow selection and preflight checks to submission, monitoring and output download.
NVIDIA/skills
Runs NVIDIA TAO Data Services KPI analysis on object detection results, comparing predictions to ground truth and writing per-class precision, recall and AP to a CSV.
Works with
Categories
Modify, build, test, debug, and contribute to NVIDIA cuOpt (C++/CUDA, Python, server, CI). Cuopt Developer is an agent skill from NVIDIA/skills, published by the product's own GitHub organization. Modify, build, test, debug, and contribute to NVIDIA cuOpt (C++/CUDA, Python, server, CI).
Cuopt Developer fits situations like: solver internals; code conventions.
Run `npx skills add NVIDIA/skills --skill cuopt-developer -a claude-code`. Or copy the skill folder (skills/cuopt-developer in NVIDIA/skills) into .claude/skills/cuopt-developer in your project. Claude Code loads it when a task matches its description.
Run `npx skills add NVIDIA/skills --skill cuopt-developer -a codex`. Or copy the skill folder (skills/cuopt-developer in NVIDIA/skills) into .agents/skills/cuopt-developer in your project. Codex loads it when a task matches its description.
Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add NVIDIA/skills --skill cuopt-developer -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/cuopt-developer, .gemini/skills/cuopt-developer, .github/skills/cuopt-developer and .opencode/skills/cuopt-developer in your project.
Going by SKILL.md and its folder, Cuopt Developer needs the command-line tools its instructions call (git, conda and pytest). Our summary lists: Python 3.
SKILL.md names 1 domain. As links in the text: docs.rapids.ai. This is read from the text; nothing was executed.
Our automated static check of SKILL.md found notes only (runs commands with sudo), nothing it rates as a warning. It is not a guarantee. Review the folder before installing.
Cuopt Developer is published under the Apache-2.0 licence (declared in SKILL.md). It allows redistribution, so the full SKILL.md is shown on this page.
About 3.2k tokens (SKILL.md is roughly 13k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full. Its references folder adds about 7.8k tokens, read only when the agent opens those files.
Skills that share tags, products or a category with Cuopt Developer: Cutlass Skill (slowlyC/agent-gpu-skills, 169 stars), Make Op Verify (CVCUDA/CV-CUDA, 2.7k stars), Review Op Support (CVCUDA/CV-CUDA, 2.7k stars) and Review Op Test Coverage (CVCUDA/CV-CUDA, 2.7k stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.
NVIDIA (a GitHub organization, an official publisher) maintains it in NVIDIA/skills, which has 3,534 GitHub stars. The repository holds 380 skills in this directory. The repository was last updated on October 7, 2026.
Source: NVIDIA/skills on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.