What should I check in the first 30 minutes with a new table?

At the start of the course you answered four questions before you knew the method: one on the welcome page, about a sales table, and three about the Rueda table. Here they are again, word for word. Answer without looking back at your earlier answers; at the end you can compare them with your Maestro.

Check yourself: In the trip table, 73 rows have no birth year. What do you do first?

Check yourself: The average trip duration is 34 minutes. What does that tell you about a typical trip?

Check yourself: One trip lasts 1,452 minutes: more than 24 hours. What do you do?

Check yourself: Someone sends you a new sales table and asks for a chart of price versus units sold by the end of the day. What do you do first?

Compare today's answers with the ones from the start. If one changed, notice what changed: usually it's that you now look for evidence in the other columns before you decide. You can go over them with your Maestro and see what changed.

With any new table, in this order:

What a row is. Count rows and distinct identifiers, and look at rows from the top and from the middle.
Shape and types. How many rows and columns, what type each one has, and what period it covers.
What's missing. How much is missing per column, whether gaps go together, and where they cluster.
Each column. A histogram and a describe() for the numeric ones; mean versus median.
Categories. value_counts() for each text column, looking for labels that shouldn't be there.
Extremes. Each far-out value, judged by its other columns: probable error, real, or pending.
Relationships. Two columns at a time, looking for the group that behaves differently.

Everything you find goes in the log, with how many rows and what you suspect. One way to split half an hour: about 5 minutes for the row and the shape, 5 for what's missing, 10 for the columns and their categories, 5 for the extremes, and 5 for relationships and tidying up the log. You don't have to finish everything: whatever you don't get to, write down as pending.

A first look doesn't answer the question you were asked: it tells you whether the table can answer it, and with what caveats. Sometimes the most useful result of half an hour is a list of three questions for whoever sent you the data, before you make a single chart for the report. And the next time you open a table from the same source you'll be faster, because your log tells you where to look first.

This cheat sheet has the code for each step, with print() on every line so it shows everything even when you run it as a script (python primer_vistazo.py). Copy it and swap in your own file and column names.

primer_vistazo.py

import pandas as pd

datos = pd.read_csv("tu_tabla.csv", parse_dates=["fecha"])

# 1. What a row is: rows versus distinct identifiers
print(len(datos), datos["id"].nunique())
print(datos.head(10))
print(datos.sample(8, random_state=1))

# 2. Shape, types and what period it covers
print(datos.shape)
print(datos.dtypes)
print(datos["fecha"].min(), datos["fecha"].max())

# 3. What's missing
print(datos.isna().sum())

# 4. Each numeric column: compare mean with 50%
print(datos["monto"].describe())

# 5. Categories
print(datos["categoria"].value_counts())

# 6. Extremes: look at the whole row
print(datos[datos["monto"] > 10000])

# 7. Relationships
print(datos["monto"].corr(datos["unidades"]))

One line per step. Replace tu_tabla.csv, fecha, id, monto, categoria and unidades with your own table's names.

This course finds and logs problems; it doesn't fix them. Two free Maestro courses pick up from here: one on deciding what to do with the gaps, duplicate labels and extremes you found, and one for when the trouble starts earlier, as you open the file: garbled accents, commas that don't separate columns, leading zeros that vanish.

And if you already work with an AI agent, or want to start, the next course is Agentic AI for Data. There you hand data work to an agent in your terminal and check what it did with the same judgment you practiced here: what a row is, what's missing, which extreme is an error. An agent does the arithmetic of a first look fast; the questions in your log are still yours.

Agentic AI for Data: Your Judgment, at Scale

Reflect: Think of a table from work or from your own life that you'll be opening soon. What would a row be, what three things would you check first, and what do you already suspect before you open it?

Only you and your Maestro read this. It's kept on your account and deleted with it.

Write your answer