data-science-project

data-science-project — The Front Door

Overview

This is the auto-triggering counterpart to the /ds command. It fires when a user starts a tabular supervised-learning task in plain language ("help me build a churn model", "predict this column", "let's look at this dataset") without knowing the pipeline exists.

Its job is to make sure framing, an honest baseline, and a leakage-safe validation strategy happen before the headline model number is reported — without stopping the run to ask permission first. Frame inline, then keep going: the gates (baseline, validation, slices) are what this plugin is for, not a pause for orientation.

When to Use

• A tabular ML / data-science task is beginning and no .last-ds-mile/stages/ directory exists yet — the user hasn't entered the pipeline.
• The user describes a predictive goal ("classify", "predict", "forecast a column", "score these rows") or an exploratory one ("look at", "explore", "EDA on") for row-and-column data.

Do not use when:

• .last-ds-mile/stages/ already exists — the pipeline is underway; defer to /ds, which routes to the actual next stage.
• The user asked a direct factual question about the data (columns, row count, dtypes) with no modeling or evaluation ask attached — just answer it. See ds-method's guard against escalating a plain question into a framing exercise.
• The task is text, vision, recommenders, or time-series forecasting — outside this plugin's scope (see README → Scope).

Core Process

1. Check whether the pipeline already started. Glob .last-ds-mile/stages/*.md. If any stage file exists, do not re-onboard — run the /ds router logic instead (print the map, mark stages done/next, route to the first missing stage) and stop.
2. Frame in-line, in one or two sentences, then move on. State the target, the decision it feeds, and the success metric as your own best read of the request — do not ask the user to confirm before proceeding. Only ask a question here if the answer would change which column is the target or invalidate the whole run; note assumptions instead of pausing on anything else.
3. Carry the request through the pipeline in this same turn, applying each stage's gate as you reach it (honest baseline, leakage-safe validation, slice performance) rather than stopping to hand off. Pick an artifact mode per ds-method — express (one .last-ds-mile/run.md for a single-shot ask, the default here) or full per-stage .last-ds-mile/stages/*.md files (a genuine multi-session project) — and say which you picked. Either way, the artifact is a record of what you did, not a checkpoint to wait at.
4. Never end the turn asking permission to begin. A request to build or evaluate a model is carried through to a model, a scored baseline, and a verdict — not a pipeline map and a question. If you are missing a Hard Gate artifact ds-method requires, produce it inline (see ds-method's discipline-gate handling) and say so; don't stop and ask the user to go run a separate command first.