data-analysis

Start from the value delivered to users, not revenue, then build bottom-up: identify behaviors that move the north star and find leading indicators of those behaviors. The Skill produces a hierarchy document with L1 input metrics, L2 levers, and guardrail metrics to prevent gaming.

Define each step precisely with a maximum time window, count unique users not sessions, and report step conversion with 95% confidence intervals. Segment by channel, device, or cohort to locate where the biggest drop occurs and whether it is consistent across segments.

A valid test pre-specifies a single primary metric, computes required sample size from baseline rate and minimum detectable effect at 80% power, and runs for a fixed duration covering a full business cycle. Avoid peeking, early stopping, and declaring underpowered tests as showing no effect.

Yes, it supports hands-on execution mode where you provide data and it validates for nulls, duplicates, and schema issues, explores distributions before aggregating, and delivers a structured report. Findings are labeled as exploratory hypotheses or confirmatory results.

Training only on closed deals creates survivorship bias — the model learns what converted deals looked like at close, not what leads looked like at scoring time. Train on all leads that entered the funnel with outcome labels attached within a defined time window.

Skip analysis when the decision is already made, the sample is too small to distinguish signal from noise, the data does not measure the phenomenon of interest, or collecting credible data takes longer than the decision's relevance window.