If we had to cut the course to ten lessons

Someone asked which lessons they should read if they only had an afternoon. We could not agree, so we each wrote a list of ten and compared.

Six appeared on everyone's list:

1. Rows, grain, and the shape of a dataset — unanimous, and first on six of nine lists. Nearly every wrong number traces here.
2. Turning a request into a question — the most under-taught skill in the job.
3. Joins that do not fan out — the specific mechanism behind most wrong numbers.
4. Describing a column without lying — because the mean is the default and usually wrong.
5. Idempotency, backfills, and yesterday arriving twice — the property that separates a pipeline you can operate from one you babysit.
6. The first hour with an unfamiliar dataset — the routine that catches most errors before they exist.

The contested four, in rough order of how many lists they appeared on:

7. Window functions — highest leverage per hour of any technical topic here.
8. Choosing the form before the colours — the visualization side campaigned hard and was right.
9. Designing an experiment that can succeed — mine, and I will defend it: most experiments fail at the arithmetic stage.
10. When not to build a model — from the modelling side, and the one that saves the most time.

What is striking is what did not make it. Nothing about algorithms. Nothing about specific tools. Nothing about architecture. The consensus list is almost entirely about knowing what you are holding and what you are being asked — and those are the parts that transfer between jobs, stacks and decades.

If you have an afternoon, that is the order.