What problem does it solve? Teams routinely make causal claims from weak evidence — opt-in pilots, naked before/after comparisons, and metrics chosen after the data arrives. This Skill provides a rigorous workflow for designing experiments that actually isolate cause and effect, so decisions about process changes, tooling rollouts, and product mechanics rest on valid evidence rather than noise. ## Core Features & Use Cases - Falsifiable Hypothesis Framing: State hypotheses with a mechanism and secondary predictions, so you can distinguish "it worked" from "something worked". - Comparison Design & Confound Hunting: Choose the strongest feasible design (randomized control, staggered rollout, matched comparison, diff-in-diff) and walk a confound gallery covering selection effects, seasonality, novelty, instrumentation drift, and contamination before running. - Pre-Registration & Decision Rules: Lock the primary outcome, decision rule, guardrails, detectable effect size, and stop date before data collection to prevent post-hoc rationalization. - Use Case: A platform team wants to know whether AI-assisted code review reduces escaped defects. The Skill guides them to randomize at the team level, pre-register the defect-rate decision rule, blind the graders, and check the mechanism (pre-merge findings per PR) before adopting. ## Quick Start Ask the agent to design an experiment testing whether a specific change causes a measurable outcome, providing your constraints on randomization, units, and the smallest effect worth detecting.
