원클릭으로
hpc
Launch and manage jobs on the HPC cluster (INVOKE ONLY when the user explicitly runs /hpc or asks to use this skill).
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
메뉴
Launch and manage jobs on the HPC cluster (INVOKE ONLY when the user explicitly runs /hpc or asks to use this skill).
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Make the current work resumable from a clean checkout — commit/push, record in-flight async work, flag what isn't durable. INVOKE ONLY when the user explicitly runs /handoff or asks for a handoff.
Analyze files for quality and consistency within the codebase
Deep analysis using parallel reasoning with iterative refinement
Deep verification that all changes meet requirements and preserve correctness
Identify files needed to accomplish a task
Read files in context
| name | hpc |
| description | Launch and manage jobs on the HPC cluster (INVOKE ONLY when the user explicitly runs /hpc or asks to use this skill). |
You run the hpc launcher yourself to put the user's work on the cluster. That is
the point of the tool — do not hand commands to the user to run, unless asked.
Look for its path in the user's global ~/.claude/CLAUDE.md, under an
## hpc-launcher location heading.
If it isn't there, discover it: run command -v hpc with the Bash tool
(Git Bash) and take the directory of that path. If that fails, ask the user
where the hpc-launcher folder is.
Record it so later calls skip discovery — append to ~/.claude/CLAUDE.md:
## hpc-launcher location
The `hpc` launcher (for the /hpc skill) is at: `<absolute path>`
<path>/README.md — what it does, setup, every command. Always.<path>/clusters/<HPC_CLUSTER>.md — facts for the cluster this repo targets
(HPC_CLUSTER in .hpc.env; hpc clusters lists them). Read it when running or
configuring a job — partitions, filesystem, gotchas (e.g. Roihu's split login
hosts + 24 h SSH cert).<path>/AGENTS.md, DESIGN.md, TODO.md — only when you need rationale,
limitations, or current status.Read the live files. Do not rely on remembered defaults (partitions, modules,
torch module/wheel, walltimes) — they live in the cluster's clusters/<name>.env
profile and hpc-env.example, not in your memory.
hpc is bash; use the Bash tool (Git Bash / MSYS), not PowerShell.hpc run, hpc test, hpc status, hpc logs,
hpc collect, hpc upload, hpc wait, hpc init, etc. Don't delegate them to
the user. (For unattended/scripted use, hpc wait blocks until a job is done and
exits its rc, and hpc status --json / hpc logs --once are non-interactive —
see the README.)hpc call fails on SSH
auth, ask the user to load it once (ssh-add their key), then keep running the
commands yourself..hpc.env (run hpc init) before hpc run / hpc test.