phoebefu6
UserData skills you can install, not just read. Claude skills that run a full analysis on your own schema - 233,835 real rows, executed charts, flaws planted on purpose with an answer key, and a review pass that rewrites the code. by Phoebe Fu
Categories
Indexed Skills (12)
how-to-data-quality
Put a data-quality gate on a warehouse load - map every defect to one of the six DQ dimensions (completeness, validity, consistency, timeliness, uniqueness, accuracy), run a real FAIL -> fix -> PASS cycle with a quarantine ledger that reconciles every row, and generate a self-contained executive DQ scorecard where every number is computed. Use when asked to "check data quality", "validate this load", "build DQ checks", "data quality scorecard", "should we trust this data", "set up a quality gate", "quarantine bad rows", or when a pipeline needs defects caught at ingestion instead of in a board meeting. Walks the 6-step pipeline - input, sample data, objective, find-skills, build (gate + fixes + scorecard), expert review.
how-to-eda-codex
Run a trust-first exploratory data analysis (EDA) with Codex + Python on any schema, assembling an open-source profiling / integrity / forensic toolbox before writing code. Use when handed a new dataset from an unproven pipeline and asked to "can we trust this data", "profile this", "run EDA", "check data quality", "find integrity issues", or before building any dashboard/model on unfamiliar tables. Walks the same 6-step pipeline as how-to-eda (the Claude edition) but leads with referential integrity, entity resolution, and forensic checks - and outputs findings ranked by dollar and trust impact.
how-to-eda
Run a decision-first exploratory data analysis (EDA) with Claude + Python on any schema. Use when handed a data schema or new dataset and asked to "explore the data", "run EDA", "profile this", "what's in this data", "can we trust this data", or before building any dashboard/model on unfamiliar tables. Walks the 6-step pipeline - input, sample data (if no real data yet), objective, find-skills, code with real executed charts, expert review - and outputs findings ranked by dollar impact.
how-to-metric-to-scorecard
Turn raw marts into a board-ready executive scorecard - define metrics explicitly, validate each against the naive version an analyst ships by accident, then generate a self-contained HTML scorecard where every number is computed, not hand-typed. Use when asked to "build a scorecard", "executive/board dashboard", "KPI report", "define our metrics", "revenue dashboard", "what do we put in front of the board", "metric definitions", or when a stakeholder wants numbers they can trust and audit. Walks the 6-step pipeline - input, sample data, objective, find-skills, build (metric definitions + validation + HTML scorecard), expert review - and outputs an auditable scorecard, not a pile of aggregates.
how-to-rfm
Run a decision-first RFM customer segmentation with Claude + Python on any transaction history. Use when handed customer/order data and asked to "segment customers", "run RFM", "who are our best customers", "who is churning", "where should we spend retention/CRM budget", "find high-value customers", or before designing any lifecycle/loyalty/win-back campaign. Walks the 6-step pipeline - input, sample data (if none yet), objective, find-skills, code with real executed charts, expert review - and outputs segments ranked by dollar at stake with a budget recommendation.
how-to-sales-forecasting
Forecast seasonal demand honestly and turn it into a signable inventory buy - run a data-quality gate before measuring anything, separate what the data can prove (weekly seasonality, backtested accuracy) from what it cannot (a peak observed once), then price the unprovable part as a costed newsvendor decision instead of a fake confidence interval. Use when asked to "forecast sales/demand", "how much inventory for <month>", "plan the seasonal buy", "predict next quarter's volume", "is this spike real", or when a forecast must survive a finance review. Walks the 6-step pipeline - input, sample data, objective, find-skills, build (DQ gate + backtests + spike sizing + Monte Carlo + costed commitment), expert review - and outputs a buy paper, not a point estimate.
how-to-schema-and-warehouse
Turn a raw operational data dump into a queryable star-schema warehouse with a data contract, using DuckDB + pandera. Use when handed raw exports from one or more source systems and asked to "design a schema", "build a warehouse", "model this data", "set up a data lake / warehouse", "create a star schema", "reconcile these sources", or before any analytics/BI is built on unmodeled data. Walks the 6-step pipeline - input, sample data, objective, find-skills, build (ERD + DDL + contract + real queries), expert review - and outputs a genuinely queryable warehouse, not just a diagram.
how-to-schema-consolidation
Audit a sprawling multi-server data estate (many cloned schemas that have drifted) and safely consolidate the copies of a table into one, using a column-level schema diff. Use when handed many servers/schemas that each cloned the same tables, or asked to "consolidate schemas", "merge duplicate tables", "compare schemas across servers", "schema diff", "which columns differ across environments", "dedupe tables", "measure data dictionary coverage", or before merging any table that exists in more than one database. Walks the 6-step pipeline - input, sample data, objective, find-skills, build (coverage + redundancy + drift audit, schema-diff matrix, merged superset DDL), expert review - and never merges on the intersection.
how-to-agentic-on-your-warehouse
Put a text-to-SQL agent on a warehouse without putting a wrong number in the board pack - ground it on a signed-off metric registry, keep every number the model reports traceable to a real executed result row, and let a deterministic five-check gate reconcile each answer against the trusted definition before anyone sees it. Use when asked to "let people ask the warehouse questions", "build a text-to-SQL agent", "chat with our data", "natural language BI", "why does the AI give different numbers than the dashboard", or when an analytics agent has to survive a finance review. Walks the 6-step pipeline - input, sample data, objective, find-skills, build (registry + agent + gate + retry loop), expert review - and outputs a runnable agent plus a captured fail-to-fix transcript, not a demo.
how-to-biz-monitoring
Monitor business metrics so that the alerts are worth reading - build a causal seasonal baseline on the log scale, calibrate the null before trusting any threshold, control the false discovery rate across every series watched on the same day, and score the whole thing against changes planted at known times so recall is measured rather than claimed. Then attribute a real move exactly, splitting a ratio metric into mix and rate so a healthy business does not get escalated. Use when asked to "set up alerting on our metrics", "why did revenue drop", "our dashboard alerts too much", "nobody reads the alerts", "detect anomalies in our KPIs", "is this dip real", "build a weekly metrics digest", or when someone proposes a percent-change threshold on a noisy series. Ships five standalone modules - baseline, detect, attribute, bench, digest - a runnable baseline script, and an injection bench that scores any detector you write.
how-to-data-dictionary
Audit a data dictionary against the warehouse it claims to describe, then generate one that cannot rot - measure every structural fact from the live database, read guarantees from the data contract rather than from today's rows, hand-write only the meaning, and put a drift gate in CI so the schema can never move without the document failing a build. Every check that cannot decide says so instead of passing. Use when asked to "write a data dictionary", "document our tables", "our docs are out of date", "set up a data catalog", "what does this column mean", "which column is revenue", "onboard analysts onto the warehouse", or when someone is about to hand-maintain a column spreadsheet. Ships five standalone modules - dbprofile, claims, audit, gate, render - plus a self-contained browsable HTML dictionary and a mutation suite that proves the detectors fire.
how-to-feature-to-model
Turn a relational schema into a supervised model that survives production - build point-in-time features that respect each fact's own availability lag, cut the label where the outcome clock has not finished, screen the pipeline for leakage instead of trusting the held-out score, measure against a defensible ceiling and a one-line baseline, and set the decision threshold in money. Use when asked to "predict churn / returns / fraud / conversion", "build an ML model on our data", "why is my AUC so high", "is this feature leaking", "which customers should we intervene on", or when a model that looked good offline is about to ship. Walks the 6-step pipeline and ships four standalone modules - asof, leakcheck, censor, economics - each usable on its own.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.