Tooluck logoTooluck
SKILL.md

science-experiment-runbook

Agent skill fromyzlin286-wq/Deepscient

Use when turning an approved plan into an actual run — when the design and sample size are settled and someone needs to produce data, execute code, or reproduce a result; when a run must be recordable (inputs, outputs, environment, seed) so that six months later the numbers in the paper can be traced back; when a result cannot be reproduced and the difference has to be explained rather than waved away; also use when deciding where raw data, analysis code and generated artifacts belong.

How to install science-experiment-runbook

install science-experiment-runbook
# install the Deepscient pack with the skills CLI
npx skills add yzlin286-wq/Deepscient

# or add only this skill to Claude Code: copy the folder

dsh-science/skills/science/science-experiment-runbook/

# into

~/.claude/skills/science-experiment-runbook/ (all projects)

.claude/skills/science-experiment-runbook/ (one project)

dsh-science/skills/science/science-experiment-runbook/SKILL.md

MITFull file
1# 从方案到一次可复现的执行
2
3## 这一层要守住的三件事
4
51. **原始数据只进 `data/`,只读。**
6 清洗结果放 `data/clean/`,分析产物放 `experiments/<slug>/<runId>/`。
7 一旦在原始文件上直接改,你永远回不到采集当天的状态——而审稿人问的第一个问题往往就是那个状态。
82. **每次执行都要能被别人重跑。**
9 `science_exp_run` 记的是:脚本内容指纹、输入文件 sha256、输出文件 sha256、
10 Python 与关键库版本、seed、是否允许网络。缺任何一项,这次执行就只是"你跑过一次"。
113. **Python 层默认无网。**
12 数据下载在采集阶段做完,分析阶段不该联网。要联网就显式 `network: true`,
13 记录里会写明——这是给自己留的证据,不是给别人看的仪式。
14
15## 顺序
16
17```
18science_study_coach { projectId } ← 先确认 S2/S3 的门过了
19science_exp_plan { projectId, effect: … } ← 样本量与前置条件(若还没做)
20把数据放进 data/
21写 code/run.py ← 一个脚本一件事,参数从 argv 进来
22science_exp_run { projectId, slug, script: 'code/run.py', inputs: ['data/raw.csv'], seed: 42 }
23science_exp_query { projectId, what: 'runs' } ← 确认状态、产物、耗时
24```
25
26## `science_exp_run` 的字段怎么填
27
28| 字段 | 什么时候用 | 坑 |
29| --- | --- | --- |
30| `script` | 已有脚本 | 路径必须在课题目录内;`../../x.py` 会被拒 |
31| `code` | 一次性验证 / 小实验 | 会写成 run 目录里的 `run.py`,同样留痕 |
32| `inputs` | 有输入文件 | 相对课题根;文件不存在直接报错,不静默跳过 |
33| `seed` | 任何有随机性的分析 | 写进 `seed.txt` 并导出 `DSH_SEED`;不传就用默认值并记录下来 |
34| `network` | 需要下载/调用 API | 默认 `false`。Python 层会被 hook 拦住;`curl` 子进程拦不住,这一点要在记录里自己注意 |
35| `timeoutMs` | 长任务 | 超时按失败记,不会假装成功 |
36
37## 失败了怎么办
38
39失败的执行**也进溯源链**,这是刻意的:一份只记录成功的历史没有诊断价值。
40
411. 先读 `stderrTail` 和 run 目录里的 `stderr.log`。
422. 分类:代码错(改脚本重跑)/数据错(回去看采集)/环境错(`probeEnv` 对比版本)/超时(拆任务)。
433. **不要**为了让结果出来而降低检验标准、删样本、换主结局。那一步一旦迈出去,后面所有数字都要重做。
44
45## 复现性检查(交给别人之前)
46
47- 同一 seed 跑两次,输出文件的 sha256 应完全一致。不一致就说明有未受控的随机源(时间、字典序、并行度、浮点归约顺序)。
48…

Excerpt shown under the repository's MIT license.

More skills in Deepscient

Keep exploring

Similar skill repositories

thedotmack /

claude-mem
116k

Persistent Context Across Sessions for Every Agent – Captures everything your agent does during sessions, compresses it with AI, and injects relevant context back into future sessions. Works with Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode + More

CCCXGHGM+135 skills

Leonxlnx /

taste-skill
113k

Agent skills that give coding agents taste: art direction, reference images and a browser check instead of generic AI pages

CCCXCU1 skill

Newsletter

Join the Community

Subscribe to our newsletter for the latest news and updates