人工智能科学家的可重复使用技能
概览
可重复使用AI辅助科学研究的技能。 编码代理会在不同会话间丢失上下文，信任不应信任的结果，并在一次诊断运行时提出扫描。这些技能编码了重复的工作流程和判断，避免每次会话都重新学习。 每个技能都刻意紧凑。共享版本编码了跨项目通用的决策规则和工作流，而集群特定、实验室和仓库特定事实则保留在本地策略文件中。 Agentcomm为多个处理同一问题的代理提供一个共享邮箱，这样他们可以交接、互相提问，并在没有人转达的情况下告知完成任务。 它在不同线束之间最为有用。Claude Code 已经可以直接给其他 Claude 会话发送消息，因此两个 Claude 代理无需额外帮助。
概述（AI 翻译）
可重复使用AI辅助科学研究的技能。
编码代理会在不同会话间丢失上下文，信任不应信任的结果，并在一次诊断运行时提出扫描。这些技能编码了重复的工作流程和判断，避免每次会话都重新学习。
每个技能都刻意紧凑。共享版本编码了跨项目通用的决策规则和工作流，而集群特定、实验室和仓库特定事实则保留在本地策略文件中。
技能
| 技巧 | 它的作用 |
|---|---|
slurm |
提交、检查和调试Slurm作业，避免调度器问题与运行时故障混淆 |
experiment-design |
提出改变决策的最小实验 |
training-audit |
诊断早期验证衰退并选择受控优化检查 |
sensemaking |
在信任某个发现之前，先检查它是否合理 |
skeptical-labmate |
压力测试声明、基线及后续计划 |
lab-notebook |
保持记录可重启和可审计 |
quarto-presentations |
创建、渲染、检查和导出Quarto reveal.js幻灯片 |
本页由 AI 翻译，技术细节请以原文为准。
README
以下为原文 README 节选，完整内容请查看原文。
ai-sci-skills
Reusable skills for AI-assisted scientific research.
Coding agents lose context across sessions, trust results they shouldn’t, and propose sweeps when a single diagnostic run would suffice. These skills encode recurring workflows and judgment calls so they don’t have to be re-taught every session.
Each skill is intentionally compact. The shared versions encode decision rules and workflows that generalize across projects, while cluster-specific, lab-specific, and repo-specific facts stay in local policy files.
Skills
| Skill | What it does |
|---|---|
slurm |
Submit, inspect, and debug Slurm jobs without mixing scheduler problems with runtime failures |
experiment-design |
Propose the smallest experiment that changes a decision |
training-audit |
Diagnose early validation decline and choose controlled optimization checks |
sensemaking |
Check whether a finding makes sense before trusting it |
skeptical-labmate |
Stress-test claims, baselines, and follow-up plans |
lab-notebook |
Keep records restartable and auditable |
quarto-presentations |
Create, render, check, and export Quarto reveal.js slide decks |
job-babysitting |
Classify job health and choose the next action |
mission-control |
Coordinate workers, reconcile steering, and intervene before directions become stale |
rut-breaker |
Detect low-yield loops and force a better reset move |
handoff |
Create restart-ready session handoffs |
bounded-auto-loop |
Run a short autonomous loop with a fixed budget and stop rule |
maintenance-pass |
Trim dead code, vestiges, and duplication without churn |
research-code-parsimony |
Reuse existing owners and cut over cleanly instead of growing code |
claude-code-worker |
Delegate isolated Claude Code tasks with resumable sessions and quota-aware recovery |
agentcomm |
Let agents coordinate with each other through a shared mailbox |
quota-saving-mode |
Cut the model requests spent on polling, waiting, and status chatter |
humanize |
Edit prose so it sounds like the author and reads plainly |
Agents talking to each other
agentcomm gives several agents working on one problem a shared mailbox, so
they can hand off, ask each other questions, and say when something is done
without a person relaying it.
It is most useful across harnesses. Claude Code can already message other Claude sessions directly, so two Claude agents need no help from this. A Claude agent and a Codex agent cannot see each other at all, and the mailbox is what lets them coordinate: a durable name each, a roster showing who is reading, and read state that tells a peer whether a question has actually been picked up.
quota-saving-mode uses it for the same reason, to deliver a wake-up when a
long job finishes rather than having an agent poll for it. That skill ships
bin/await-notify and bin/await-event for harnesses without built-in
推荐工具
换一个关键词，或者移除筛选条件。
安装
npx skillfish add uchicago-dsi/ai-sci-skills