Ds explore
Runs exploratory data analysis — distributions, relationships between each feature and the target, correlations — with a running hypothesis log. Use when someone asks for EDA, wants to explore or visualise a dataset, or wants to plot how features relate to what is being predicted. Use after a dataset has been profiled and before features are engineered.From its SKILL.md
npx -y skills add StamKavid/last-ds-mile --skill ds-exploreAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 3 stars3 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
3.0 KB, as published. Nobody here has run it
ds-explore — Exploratory Data Analysis
Overview
Systematically sweeps a profiled dataset for patterns relevant to the target, logging a hypothesis for each finding instead of plotting aimlessly.
When to Use
- After
/ds-datahas produced a data dictionary and integrity findings. - Before feature engineering begins.
- NOT for: acting on findings by transforming features (that's
/ds-prep) — this stage observes and hypothesizes, it doesn't change the data.
Core Process
- Univariate pass: distribution of each key variable (numeric: histogram and summary stats; categorical: value counts).
- Bivariate pass: relationship of each candidate feature to the target, and obvious collinearity between candidate features.
- Log a hypothesis for every pattern noticed ("higher X seems associated with target=1, hypothesis: because ...") rather than producing plots with no question behind them.
- Flag anything that looks too predictive at this stage (see
ds-method's Red Flags) as a leakage candidate for/ds-prepto resolve. - Export the 1-3 plots that actually back the strongest findings from steps 1-2 (not
every plot considered) to
.last-ds-mile/figures/02-<name>.png— perdata-viz-standards, state the hypothesis before building each one. Typically: the target's distribution (and its transform, if skewed) and the single strongest bivariate relationship found. This is the "critical/important only" bar, not exhaustive EDA output. - Write to
.last-ds-mile/stages/02-explore.md: key findings, the hypothesis log, leakage candidates flagged for follow-up, and a reference to each exported figure next to the finding it illustrates.
Common Rationalizations
| Rationalization | Reality |
|---|---|
| "I'll just make a bunch of plots and see what jumps out" | Aimless plotting produces cherry-picked patterns. Every plot should test a stated hypothesis. |
See ds-method for the shared Rationalizations that apply to every stage.
Red Flags
| Red Flag | What it usually means |
|---|---|
| A feature is almost perfectly separated by target in a bivariate plot | Candidate leakage — don't treat it as a modeling win yet. Flag it for /ds-prep's leakage check. |
See ds-method for the shared Red Flags that apply to every stage.
Verification
- Univariate and bivariate passes both done.
- Hypothesis log has one entry per notable finding.
- Any suspiciously predictive feature is flagged as a leakage candidate, not celebrated.
- The 1-3 most important plots are exported to
.last-ds-mile/figures/, not just described in prose. -
.last-ds-mile/stages/02-explore.mdwritten.