Claude Skills HPC
szaghi/claude-skills-hpc
Padrões HPC sob demanda para Claude Code, sem viés
Ver no GitHubConjunto curado de skills para Claude Code que ensina ao agente IA os padrões HPC (paralelismo, aceleradores, linguagens) baseado em especificações técnicas. Inclui referências de MPI 5.0, OpenMP 6.0, OpenACC 3.4, Fortran 2023 e C23, estruturadas em capítulos navegáveis para respostas precisas sem alucinação.
Instalação
npx convez-ai szaghi/claude-skills-hpc1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Consulta de semântica MPI e OpenMP durante desenvolvimento paralelo
- Verificação de conformidade com especificações HPC
- Resolução de problemas de deadlock e sincronização em código paralelo
- Aprendizado de diretivas de offload (OpenACC, OpenMP device)
- Síntese de código Fortran/C moderno respeitando padrões
Veredito

Ferramenta rigorosa para quem trabalha com HPC em escala; elimina alucinação em questões de conformidade ao vincular respostas a specs oficiais. Instalação simples, peso leve no contexto (carrega capítulos sob demanda). Impacto real em ambientes de pesquisa e computação de alto desempenho.
README
A tradução para o português deste README ainda está na fila.
claude-skills-hpc
HPC Reference Skills for Poor People
a curated, growing pack of Claude Code agent skills that teach your AI the HPC standards — parallelism, accelerators, languages — so it answers from the spec, not from vibes.
| 📚 Spec-grounded answersEach skill distills a full standard — or a practitioner's playbook — into frameworks, decision tables, and worked examples; the agent reads the chapter before answering, instead of hallucinating a flag | 🧭 Navigable, not a brain-dumpA SKILL.md entry point + topic index routes to on-demand chapters/; only the relevant chapter is loaded, so context stays cheap | 🔬 Conformance- and craft-firstFor "what does the standard actually require?" — constraint IDs, version deltas, send-mode semantics — and for "what's the right way to build this?" — decomposition, roofline, the parallel-pitfall catalogue | 🛠️ Synthesis, not a copyOriginal explanatory prose; only API names, code, and numeric spec values coincide with the source — each skill ships a NOTICE |
|---|---|---|---|
⚡ Drop-in for Claude CodeOne ./install.sh copies them into ~/.claude/skills/; restart and they trigger automatically on the right questions | 🤝 Plays nice with dotfilesSkips any skill you already manage via symlink/stow — won't clobber an existing managed setup | 📦 Standards + playbooksThe core HPC standards (message passing, offload, languages) plus cross-cutting practitioner playbooks (parallel design, numerics, cluster tooling) — see the full list below, expanding over time | 🆓 Free & openCC-BY-4.0 on the original content — share and adapt with attribution |
Built with book-to-skill
These skills are generated, not hand-typed: book-to-skill reads a specification or a technical book and extracts its structure — named frameworks, API semantics, anti-patterns, worked examples — into a navigable skill.
Skills
hpc-indexis the router, not a knowledge base: a thin navigation layer that decides which of the skills below to consult when a question plausibly spans several (reference vs. applied vs. theory). It kicks in only when the target is ambiguous — a query that already names one technology routes straight to that skill.
| Skill | Distils | Reach for it when… |
|---|---|---|
mpi-5.0 | MPI: A Message-Passing Interface Standard, v5.0 | message passing, collectives, RMA, MPI-IO, datatype matching, deadlock hunts |
openmp-6.0 | OpenMP API v6.0 (+ Nov-2025 errata) | directives/clauses, tasking, device offload, the flush memory model |
openacc-3.4 | OpenACC API v3.4 | !$acc directives, data clauses, gang/worker/vector, async queues |
fortran-2023-standard | Fortran 2023 (J3/23-007r1) | what the standard requires/permits; modern-Fortran features; Cxxx constraints; conformance |
iso-c-9899-2024 | ISO/IEC 9899:2024 — C23 (N3220 draft) | what the standard requires/permits; undefined/unspecified behavior; C23 features; integer promotions; the memory model |
iso-cpp-2023 | ISO/IEC 14882 — C++23 (N4950 draft) | well-formed vs UB/IFNDR; value categories & move semantics; concepts/ranges/coroutines; overload resolution; the memory model |
cuda-programming | NVIDIA CUDA Programming Guide, Release 13.3 | CUDA C++/Python kernels, SIMT/tile, streams, graphs, unified memory, multi-GPU |
gpu-multithreading | Cross-technology parallel-programming playbook — design, performance laws, optimization | choosing a decomposition (PCAM); Amdahl/Gustafson/roofline; threads/MPI/CUDA/OpenMP/OpenCL/Thrust; load balancing; diagnosing parallel pitfalls |
python-hpc | Python performance-engineering playbook — CPU and GPU | profiling (cProfile/Scalene/py-spy); NumPy vectorization; compiling (Numba/Cython, the GIL); concurrency (asyncio/multiprocessing/Dask); Polars; Numba-CUDA/CuPy/RAPIDS/JAX |
cpp-hpc | C++ HPC playbook — toolchain, idioms, parallel models, ecosystem | CMake/Spack/SLURM; modern C++ & the STL; parallel-STL/OpenMP/MPI/CUDA; Kokkos portability; HPC hardware & roofline; parallel I/O (HDF5/MPI-IO); debugging/profiling; BLAS/LAPACK/PETSc |
hpc-cluster-tooling | Cluster workflow & command-line tooling playbook | Unix shell for HPC; Make & CMake; git discipline; GDB & sanitizers/Valgrind; profiling (gprof/perf/PAPI/TAU); SLURM batch jobs (sbatch/squeue, arrays, dependencies) |
hpc-numerics | Numerical & algorithmic theory of scientific computing | floating-point/round-off & stability; ODE/PDE discretization & CFL; numerical linear algebra & Krylov/multigrid solvers; roofline & cache blocking; N-body & Monte Carlo |
Each skill lives self-contained under skills/<name>/: a SKILL.md entry point, on-demand chapters/, and supporting glossary.md · patterns.md · cheatsheet.md.
What they're for
These exist to make the agent a competent pair for real HPC work. They were forged alongside ADAM — a multi-physics AMR SDK for high-performance computing, from laptop to exascale device-accelerated supercomputer — where "is this standard-conforming Fortran, C, or C++?" and "what's the right MPI / OpenACC / CUDA incantation?" are daily questions. Point the agent at a kernel and it can check the actual rule instead of guessing.
Install
git clone https://github.com/szaghi/claude-skills-hpc
cd claude-skills-hpc
./install.sh
This copies each skill into ~/.claude/skills/. Restart Claude Code (or reload skills) and they trigger automatically. Update later with git pull && ./install.sh.
| Command | Does |
|---|---|
./install.sh | install / refresh all skills |
./install.sh -n | dry-run — show what would happen |
./install.sh --uninstall | remove the skills this repo installed |
CLAUDE_SKILLS_DIR=/path ./install.sh | install to a non-default location |
install.sh refuses to overwrite a target that is already a symlink (e.g. one you manage through your own dotfiles/stow), so it won't trample an existing setup.
Authors
Stefano Zaghi · stefano.zaghi@cnr.it
Chief Yak Shaver, Accidental Research Scientist, and HPC Farmer — decided that re-reading the MPI standard for the hundredth time was one time too many, and taught the robot to do it instead.
Claude (Anthropic)
Omniscient Code Oracle & Tireless Rubber Duck — read the specs so you don't have to, and now quotes constraint IDs back at you.
Contributions welcome — corrections and new skills via pull request. Keep edits to a skill's own expression (no verbatim spec prose); new skills must ship a NOTICE.
Copyrights
Original content — the syntheses, structure, explanatory text, tables, and indices — is licensed under CC-BY-4.0 (see LICENSE): share and adapt freely, with attribution.
This does not relicense third-party material. Each skill carries a NOTICE naming its upstream source and terms; API names, code examples, and numeric spec values remain subject to their owners' terms. These skills are independent works — not affiliated with, endorsed by, or reproductions of any specification or vendor documentation. For authoritative behavior, consult the upstream source named in each skill's NOTICE.
"CUDA" and "NVIDIA" are trademarks of NVIDIA Corporation; other product and standard names are trademarks of their respective owners.