/explore-my-data

name: explore-my-data
description: Use when the work is hands-in-the-data right now — a number moved, an open question needs exploring, a picture of the estate needs drawing, a change needs its blast radius known. An OPEN-ENDED look at data with no wrongness symptom - find what is there, and make the findings survive scrutiny later. Pre-registers the questions and the finding-bar BEFORE results are examined, logs every cut (N cuts => ~N/20 false hits expected), keeps effect sizes and bases ahead of significance talk, labels findings Exploratory-found until a hold-out check YOU run confirms, and records dead ends. Detects: "find insights", "explore this dataset", "what drives [metric]", "any patterns here", "dig into the segments". Within this family: a number already wrong or moved is triage-my-number; drawing the structure is map-my-estate. No question chosen yet — "what could our data tell us?" — is worth-knowing. Read-only: writes the exact cuts and checks for you to run and paste back; never runs the analysis.
allowed-tools: Read, Write

explore-my-data

The colleague who keeps your exploration honest: registers what you're looking for before you look, counts every cut you take, and won't let the lucky cell become the headline.

When to use

Fire at the START of an open-ended investigation — "find insights in this", "explore the data", "what drives [metric]", "any patterns in the segments", "dig into why X varies" — with no wrongness symptom in hand. Works whether the cuts are still to be run or a results table is already on the desk (then everything already seen is registered as post-hoc). Do NOT fire when a number is already wrong or moved (`triage-my-number`), when a controlled causal result needs validating (`audit-my-experiment`), when the premises of a SOURCE need clearing (`audit-my-assumptions`), when no question has been chosen yet and the agenda itself is the ask — "what could our data tell us?", sources listed but nothing in hand (`worth-knowing` charters the questions; this room runs a chosen one), or to communicate findings (`brief-my-findings`, once confirmed). `groundwork` stops before analysis — this is that analysis, harnessed.

The trap this exists to beat

Asked to "find insights," a capable model dredges — fluently. It slices until something looks striking, then hypothesizes backwards from the hit (HARKing); it never counts the slices, so the one-in-twenty fluke reads as a discovery; it leads with "+96% lift!" over "5 conversions on a base of 85"; it narrates correlation as a driver; and it ships the lucky cut as THE insight — confident, decision-shaped, and unreplicable. The garden of forking paths, walked at machine speed. This skill explores too — but it registers the questions first, logs every fork taken, keeps magnitude and base ahead of excitement, and lets nothing be called confirmed by the data that generated it.

The loop

1. **Warm start, then frame + pre-register (before looking).** If a `knowledge-base/` exists, read what powers the look: a `question-charter.md` candidate routed here arrives pre-registered — adopt its question, its decision, and its how-to-read as the finding-bar, and note the candidate id so the charter can sync; prior `exploration-log.md` dead ends are inherited, not re-walked; a contracted metric's `kpi-contract.md` pins the definition you're cutting against. Then pin the decision the exploration serves, the questions/hypotheses, the population, window, grain, and the finding-bar (what magnitude on what base would matter). Write them into the hypothesis ledger BEFORE results are examined. Results already in hand? Register the questions as they stand and label everything already seen **post-hoc** — honestly, not retroactively "predicted." 2. **Direct the cuts — and count them.** Write the exact cuts/queries for the user to run (paste-back spine). EVERY cut examined — yours or theirs, hit or miss — increments the cut log. The counter never quietly resets. 3. **Read results with magnitude first.** For each pattern: effect size, base (n), scope, THEN any significance talk — with the multiplicity line beside it: "N cuts examined ⇒ ~N/20 false hits expected at α≈.05; this could be one of them." Correlation is stated as correlation. 4. **Grade what was found.** **Exploratory — found** (a hit, unconfirmed) · **Robust pattern** (consistent across related cells / dose-response / stable in the pre-period — still unconfirmed) · **Dead end** (recorded, not deleted — dead ends are findings too). Nothing is Confirmed at this step. 5. **Write the confirmation checks.** For each finding worth pursuing: a pre-specified hold-out — a fresh window, an untouched slice, a replication cut — that the user runs and pastes back. **Confirmed only on that paste-back, never on the generating data.** A causal "X drives Y" claim needs a design, not a cut → route to `audit-my-experiment`. 6. **Emit + route.** Write `exploration-log.md` (template: `references/exploration-log.md`); a dredge-mirage stopped gets its `catches.md` line; confirmed findings hand to `brief-my-findings`; a definition wobble surfaced mid-cut routes to `kpi-contract`. A question that came from a charter candidate gets its outcome (confirmed / dead end, candidate id, evidence) recorded in the log and `timeline.md` — `worth-knowing` syncs the charter from there on its next re-fire. Then stop.

The signature output: the hypothesis ledger + cut log

An exploration record where every question is dated relative to the data (pre-registered vs post-hoc), every fork is counted, and every finding carries its grade and its confirmation path — the analog of the bench's other graded artifacts, pointed at the act of looking. Engine, the robust-vs-lucky tests, and the worked mirage live in `references/exploration-engine.md`.

Bright lines (non-negotiable; inherits groundwork's read-only line)

• **Register before looking.** Questions written after the results are post-hoc and labeled so;