What this prompt does
This prompt does feature engineering for tabular data, generating 25 candidate features a team could compute and test — each with a falsifiable hypothesis, not generic ideas. It casts the AI as a senior ML engineer. You provide your [columns], the [target], the [time_horizon], and known [constraints]. It returns 25 features spread across temporal, aggregation, interaction, ratio, categorical-encoding, text, and external categories, each with a name, definition, and a hypothesis for why it predicts the target, plus a leakage-risk flag, an expected-lift ranking, and the three to test first.
Tying every feature to a falsifiable hypothesis is what separates this from a brainstorm dump — each idea states why it should predict the [target] so you can actually test it. The leakage-risk flag respects your [time_horizon], catching the classic trap of using data computed after the prediction moment. Ranking by expected lift and naming three to test first keeps you from building all 25 when most lift comes from a few.
When to use it
- Engineering features for a tabular model from a known set of
[columns] - Generating hypotheses you can falsify rather than a vague feature wishlist
- Guarding against leakage by respecting a strict
[time_horizon] - Prioritizing which features to build first by expected lift
- Honoring data
[constraints]like no-PII or nullable columns - Covering feature categories you might not brainstorm alone (text, external, interactions)
Example output
You get a table with columns for feature, definition, hypothesis, leakage risk, and expected lift, spanning all seven feature categories. Below it, the top three features to test first are called out with reasoning. Each row is specific enough to compute, and the leakage flags are tied to your stated time horizon so risky features are obvious at a glance.
Pro tips
- Spell out the
[time_horizon]precisely — "features must use data available before the 30-day window" is what makes the leakage flags accurate - List
[columns]with their meanings, not just names, so the hypotheses are grounded in what the data actually represents - State
[constraints]likeno PII featuresorcountry may be null for 8% of rowsso flagged features respect them - Build the top three first and measure lift before constructing the full 25; most of the signal usually lives in a few features
- Scrutinize every feature flagged for leakage — anything computed after the prediction moment will inflate offline scores and fail in production
- Re-run with a different
[target]to reuse the same columns for a second model without starting from scratch - Use the seven feature categories as a checklist; the text and external buckets are the ones engineers most often forget to brainstorm
