Explore an unfamiliar dataset without losing track of what the rows mean. Reconcile sources and grain, test quality and patterns, then hand off evidence-linked findings another analyst can rerun.
Exploratory analysis can look convincing while using the wrong grain, a partial period, or a join that quietly doubles the population. This skill catches those problems before a chart becomes a decision.
It starts with the source population, cutoff, grain, keys, metric definitions, privacy boundary, and sampling plan. It then guides read-only profiling, distribution checks, time and segment comparisons, relationship analysis, anomaly review, and sensitivity checks. Every material observation stays linked to its population, denominator, method, artifact, limitations, and plausible alternative explanations.
The handoff includes source controls, a field and metric map, a data-quality register, reproducibility log, finding cards, supporting tables or figures, and the next analyses worth running. It keeps exploratory patterns separate from causal or confirmatory claims, and it blocks decision use when the source or evidence cannot support it.
Input
Explore these event and account tables before we choose onboarding experiments. The extracts have different cutoffs, account IDs repeat, and nobody has confirmed whether trial restarts create a new account row.
Output
A source and cutoff reconciliation, declared-versus-observed grain, join fanout checks, field and metric map, event coverage profile, duplicate and missingness register, bounded time and segment findings, competing explanations, a reproducibility log, and a blocked or review-ready verdict.
Input
Find useful patterns in this 20,000-row customer sample. It contains email addresses, plan history, support outcomes, and revenue. We do not know how the sample was selected.
Output
Provide a CSV, spreadsheet, Parquet file, JSON export, SQL table, or warehouse reference. Add the question or decision, source owner, expected grain and cutoff, metric definitions, privacy restrictions, allowed tools, and required deliverables. Ask the agent to explore the data and return the source controls, field map, profile, quality issues, evidence-linked findings, method log, limitations, next tests, and verdict.
A CSV, spreadsheet, Parquet file, JSON export, SQL table, or warehouse connection with enough access to inspect metadata and run bounded read-only checks.
The question or decision, expected population and grain, cutoff, key metric definitions, units, timezone, source owner, and any known control totals.
Rules for sensitive fields, row-level examples, masking, retention, compute limits, artifact sharing, and the person who reviews the analysis.
SKILL.md; analysis contract and check design; output, finding-card, reproducibility, and review templates; OpenAI interface metadata
No reviews yet.
A privacy-minimized analysis plan, sample-frame limitation, sensitive-field handling, stable denominators, distribution and segment tables, evidence-linked finding cards, no unsupported population claims, and the exact sampling evidence needed before broader use.
Input
The dashboard says customers who contact support retain better. Check the pattern and tell us whether adding proactive support will improve retention.
Output
A reconstructed metric and cohort definition, missingness and selection checks, time-order and segment comparisons, sensitivity to contact and retention windows, plausible confounding explanations, a clear statement that exploration does not establish causality, and a confirmatory test plan.
Creator
Eesvano8
