メニュー
Implement research papers from arxiv. Use when the user provides an arxiv link, asks to implement a paper, or wants to reproduce research results.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Interview-driven specification development. Use when starting a new project, after context compaction, when SPEC.md is missing or stale, or when the user needs to clarify project architecture. Triggers on "spec", "interview me", "what are we building", or when substantial work is requested without clear requirements.
Delegate complex tasks to sub-agents for parallel autonomous work. Use when GPU kernel optimization, numerical correctness verification, performance profiling, or long-running validation would benefit from focused independent execution.
Provision and manage Voltage Park H100 GPU instances. Use when the user needs to spin up H100s, SSH into VP instances, transfer files, or terminate cloud GPU instances.
| name | paper-implementation |
| description | Implement research papers from arxiv. Use when the user provides an arxiv link, asks to implement a paper, or wants to reproduce research results. |
Require arxiv link (not PDF). Fetch LaTeX: https://arxiv.org/e-print/XXXX.XXXXX (.tar.gz with .tex files)
Parse LaTeX for: architecture, algorithms, hyperparameters, loss functions, datasets. Search existing GitHub implementations. Note ambiguities.
Gather: Scope (train/inference/finetune)? Scale (model size, compute)? Baseline codebase? Priority (accuracy/speed/memory)? Validation method?
If repo exists: audit against LaTeX source. Check architecture, hyperparameters, training procedure. Identify discrepancies.
Present findings, ask questions, iterate until execution steps are perfectly clear.
Build/modify code. Write correctness tests. Profile performance.
Profile with nsight-systems/torch.profiler. Write custom CUDA/Triton kernels. Benchmark with measurements.
