What does this skill do, and when should you use it?
This Agent Skill is for developing NVIDIA cuOpt itself, not for using its solver APIs. It covers the C++/CUDA core, Python bindings, server, build and test workflows, CI, and contribution practices. It emphasizes user-space conda environments, CUDA-driver compatibility, dataset preparation, minimal diffs, DCO sign-off, and complete CI checks. It fits maintainers and contributors working on the cuOpt codebase.
It prompts the agent to clarify the component, change, tests, contribution intent, and target branch; inspect nearby code and follow existing patterns; check the CUDA version reported by nvidia-smi; create and activate ./.cuopt_env; set PARALLEL_LEVEL when needed; download test data and export RAPIDS_DATASET_ROOT_DIR; run ./build.sh, ctest, pytest, and pre-commit; diagnose CUDA-driver mismatches, missing datasets, Cython rebuilds, OOM conditions, missing nvcc, and CI failures; and guide changes to C++/CUDA code, Python bindings, server endpoints, dependencies, CUDA kernels, and VRP dimensions, including DCO and PR practices.
- A cuOpt core maintainer needs to change a C++/CUDA solver component and run the relevant build and tests.
- A Python binding maintainer needs to trace Cython-to-C++ parameter flow and update tests.
- A server developer needs to add or modify a cuOpt REST endpoint while matching the OpenAPI schema.
- A contributor is preparing a cuOpt pull request and needs branch, DCO, pre-commit, and CI guidance.
How do you install this skill?
- The skill guides conda/pip installation, test-data downloads, and commit or push workflows; confirm directories, network sources, branches, and write scope before execution.
- It assumes Linux, an NVIDIA GPU, a compatible CUDA driver, and conda; alternatives should be confirmed when these assumptions do not hold.
- Chinese-language support and mainland-China network reachability are not documented; validate availability separately for GitHub or other overseas downloads.
- BENCHMARK.md is a statically supplied evaluation report, not independent execution evidence from this review.
- Shell / CLI
- Network access
- Local filesystem
NVIDIA CUDA-compatible driverconda or mambacuOpt source repository
Use the default installation flow documented by the repository README: run npx skills add nvidia/skills, then select cuopt-developer and an installation destination when prompted. The supplied source does not document a dedicated non-interactive command for this skill or require manually cloning the repository or copying skill folders.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/cuopt-developer" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
After loading the skill in an Agent Skills-compatible client, make a specific development request such as: Help me modify the cuOpt C++/CUDA solver, build it, run the relevant tests, and prepare a contribution. State the component, intended change, required tests, contribution intent, and target branch. If you only want to use cuOpt, choose the relevant problem skill instead.
What are this skill's strengths and limitations?
- Covers environment setup, building, testing, debugging, and contribution workflows.
- Clearly prohibits sudo, CI bypasses, and unconfirmed destructive commands.
- Includes CUDA compatibility checks, test-data preparation, and common failure diagnostics.
- Addresses C++/CUDA, Python bindings, server work, and CI.
- Requires an NVIDIA CUDA-compatible driver, conda or mamba, and a usable cuOpt source checkout.
- Tests depend on MPS/data files outside the repository and require the documented dataset setup.
- The referenced auxiliary documents are not included in the supplied source, so some component-level details require consulting the cuOpt repository.
- It is explicitly for modifying cuOpt, not for merely calling or using its APIs.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| NVIDIA cuOpt Developer this page ✓ NVIDIA · Official | 62 · Recommended | ★ 3.6k | 13d ago | Apache-2.0 |
| Holoscan Conda Installer ✓ NVIDIA · Official | 52 · Use with care | ★ 3.6k | 13d ago | Apache-2.0 |
| cuOpt Numerical Optimization API ✓ NVIDIA · Official | 54 · Use with care | ★ 3.6k | 13d ago | Apache-2.0 |
| cuOpt Python Routing ✓ NVIDIA · Official | 53 · Use with care | ★ 3.6k | 13d ago | Apache-2.0 |
| cuOpt REST Server Deployment & Python Client ✓ NVIDIA · Official | 45 · Not recommended | ★ 3.6k | 13d ago | Apache-2.0 |
Unlike cuOpt problem skills such as cuopt-routing and cuopt-lp-milp, this skill focuses on modifying, building, testing, debugging, and contributing to cuOpt source code; it does not guide business-problem solver API usage.
How did FollowSkills review this skill?
The skill explicitly refuses sudo, root, system-file changes, and risky destructive operations, and requires confirmation for commits, pushes, and third-party code; it also promotes user-space setup and minimal diffs. Deducted 5 points because credential handling, sensitive-data flows, dependency-source verification, network disclosure, and rollback procedures are not systematically specified.
The material is internally coherent and covers setup checks, builds, tests, diagnostics, CUDA conventions, and failure feedback; the benchmark file reports limited evaluation results. Deducted 11 points because this review is static, key paths were not executed, some referenced files are not included in the supplied evidence, and edge-case and environment-specific failure coverage is limited.
The audience, development scenarios, and boundary against merely using cuOpt are clear, with component, build-target, and trigger guidance. Deducted 4 points because it primarily assumes Linux, NVIDIA GPU, CUDA, and conda; Chinese-language support and mainland-China reachability are not addressed, and overseas-service dependence is not evaluated.
The skill provides a version, Apache-2.0 license, author, layered reference documentation, setup/build/test notes, and maintenance-oriented guidance; the benchmark report records schema findings. Deducted 5 points for missing recommended Instructions and Examples sections, nonconforming author format, no clear changelog or update path, and unexpected-directory findings.
The documentation addresses practical tasks including building, testing, DCO, CUDA memory, parameters, endpoints, and dependencies; BENCHMARK.md reports correctness and effectiveness results across multiple tasks. Static calibration limits this to 7; points are withheld because execution and independent reproduction were not performed here, and some outputs still depend on other repository documents and the user's environment.
Versioned documents, concrete paths, commands, evaluation tasks, and metrics provide useful audit trails. Deducted 5 points because the supplied material lacks CI workflows, test execution logs, raw benchmark artifacts, or independent cross-source corroboration; key claims therefore rely largely on the author-provided evaluation report.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →