メニュー
Use when planning or revisiting an A-B test, A-A validation, split methodology, or quasi-experiment and decision quality depends on the hypothesis, randomization unit, exposure unit, guardrails, or contamination risk
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
Use when a research question spans multiple SQL, pandas, or notebook steps and you need explicit cohorts, windows, metrics, validation checks, outputs, and statistical methods before touching queries or cells
Use when you already have a written analysis plan and want to execute SQL, notebook, or validation tasks in a separate session with review checkpoints and rerun-based evidence between batches
Use when a metric definition, denominator, CUPED covariate, linearization choice, invariant, outlier treatment, missingness pattern, or leakage risk may distort an experiment or validation readout
Use when creating, editing, reviewing, or finalizing analytical notebooks whose code, assumptions, transformations, diagnostics, or conclusions need to be understandable to another analyst
Use when notebook results depend on parameters, caches, run order, hidden state, seeds, external helper code, or exported tables and the analysis must be rerunnable by another analyst
Use when an experiment analysis, metric study, notebook investigation, or methodology proposal is close to decision-ready and you need an independent check of design, statistics, interpretation, and reproducibility
| name | ds-experiment-design |
| description | Use when planning or revisiting an A-B test, A-A validation, split methodology, or quasi-experiment and decision quality depends on the hypothesis, randomization unit, exposure unit, guardrails, or contamination risk |
Design the experiment before touching the notebook. A clean statistical readout cannot rescue a weak unit of randomization, a biased metric, or a vague decision rule.
If metric definitions, denominators, CUPED, or linearization are still uncertain, switch to ds-metric-validation before trusting the design.
Write these down before analysis:
Start from causal structure, not from available tables.
If these answers point to different entities, randomization and analysis units are probably misaligned.
