菜单
Create, run, and record AI research experiments with manifests, run records, artifact paths, metrics, and experiment index updates.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Turn a research agenda into one cold-readable `/goal` prompt for autonomous AI research loops, with evidence-graph notebook rules, experiment cadence, verification, scope edges, and a finite, batch, or continuous completion contract.
Maintain the research workspace belief layer: open questions, claims, decisions, confidence levels, evidence links, and revisit conditions.
Rank candidate experiments, decide what not to run next, maintain the factory queue, and choose one decisive next research goal before autonomous loops spend compute.
Compress an AI research experiment batch or branch into lessons, updated claims/questions, and a continue/stop/pivot decision.
Route AI research workspace work across the evidence graph. Use for broad research tasks, choosing which research sub-skill applies, or updating the dashboard after experiment, claims, synthesis, or reference work.
Remote CUDA development workflow for the local NVIDIA DGX Spark at spark-f0a3.local. Use when Codex needs to probe the DGX Spark, choose a developer experience, mirror a research workspace with rsync, run CUDA/PyTorch/NAC/Le-WM training remotely while the agent stays on the Mac, use tmux/screen, manage TensorBoard tunnels, sync artifacts back, or decide between direct SSH, Docker, NVIDIA AI Workbench, and native Python setup.
| name | research-experiment |
| description | Create, run, and record AI research experiments with manifests, run records, artifact paths, metrics, and experiment index updates. |
Use when starting or updating an experiment, running a training/eval command, or recording metrics/artifacts.
program: RP### exists in programs/index.yml;
a new research direction needs its program card first (see AGENTS.md).Q### in OPEN_QUESTIONS.md.H###.experiments/NNN_short_name/README.md
experiments/NNN_short_name/manifest.yml
experiments/NNN_short_name/notes.md
experiments/index.yml
artifacts/ or data/.After a command, write:
experiments/NNN_short_name/runs/YYYY-MM-DD_HHMM_short_name.md
Prefer tools/create_run_record.py when a summary.json exists. It drafts the
run from the manifest, command, and metrics, reducing notebook tax.
Every completed serious experiment needs:
E###: one-line result.planned, running, complete, failed, paused, superseded.