What problem does it solve? Teams routinely make decisions on reported test results that lack effect sizes, confidence intervals, or adequate sample sizes. This Skill turns a reported result into a verdict that is actually supported by the data, before anyone agrees that something "worked". ## Core Features & Use Cases - Structured readout sequence: Reports the effect in decision-relevant units, a 95% interval, the assumptions that must hold, and a final verdict of supported, unsupported, or underpowered. - Underpowered detection: Computes the sample size required for the claimed effect at 80% power, so "not significant" becomes an actionable number instead of a dead end. - Guardrails against misuse: Refuses to report a p-value without an effect size, explains the cost of stopping tests early, and names the specific confounder blocking any causal claim from observational data. - Use Case: A PM reports "conversion rose 2% after the redesign, p=0.08". The Skill returns the effect with its interval, states the test is underpowered, and gives the per-arm sample size needed to detect a 2pp change. ## Quick Start Ask the assistant to evaluate a reported experiment result, for example: "Our A/B test showed a 3% conversion lift with n=200 per arm, is this significant?"
