One retail platform - Everrest - carried across the full data stack: raw dump to warehouse to analytics to AI agent. Each skill walks one complex case through a 6-step pipeline with real executed artifacts and an expert-review panel. Clone the repo, install the skill, run it on your own data.
The full CDAIO stack on one Everrest dataset - Infrastructure at the base, Gen AI at the top. Click any live card for its visual walkthrough.
Turn a raw operational dump into a queryable DuckDB star schema with a data contract - the layer everyone fakes, done honestly.
9 servers, 8,600 cloned tables, drifted apart. Audit the sprawl and schema-diff across servers before merging copies into one.
Which column is revenue, and can you trust the doc? Audit a dictionary against the warehouse, then generate one a CI gate keeps true.
Pipeline SLOs: freshness, latency and error monitoring.
What should Everrest's category team act on this quarter? Decision-first exploratory data analysis.
What eight numbers does the board need - and can we defend each one? Metrics defined, validated, and a self-contained scorecard where nothing is hand-typed.
Which of last week's movements are real enough to act on? A seven-rung detector ladder scored against changes planted at known times - so recall is measured, not claimed - and the measured answer is that no rung beats chance on precision.
Does Monday's load ship? A real FAIL -> fix -> PASS gate: quarantine ledger that reconciles every row, fabrication held not "corrected", computed DQ scorecard.
Six pipelines predict returns; five leak and only two look wrong. A computed ceiling, point-in-time features, and a threshold set in dollars instead of F1.
Can the data be trusted? Codex assembles a profiling + integrity + forensic toolbox and digs out six defects.
Where should the CRM team spend its Q3 retention budget? RFM customer segmentation.
How many units - and dollars - for November? One spike is real, one is a retry bug; a DQ gate, honest backtests, and a costed newsvendor buy.
Perfect SQL, wrong number. A grounded agent whose first answer is 809% too big, and the five-check gate that rejects it and makes it try again.
Automate a real pipeline task end to end.
A scored AI use-case backlog from a discovery workshop.
Same spine every time: input → sample data → objective → find-skills → code → expert review. Real executed outputs only - every chart on every page came from a real run with a fixed seed.
Add the marketplace once, then install any skill on its own. Each ships a runnable baseline you tune with Claude.
/plugin marketplace add phoebefu6/phoebe-data-skills /plugin install how-to-eda@phoebe-data-skills