Prompting for data analysis

Describe the data before you ask the question

"How do I calculate monthly sales?" has a dozen right answers depending on what your data looks like. An AI assistant can write excellent spreadsheet formulas, pandas code, and SQL — but only for data it understands. Your first job is to describe the structure.

A good data description includes:

• Columns and types — names, and whether each is a date, number, text, or category.
• A few sample rows — real-looking but safe (more on privacy below).
• Size and quirks — roughly how many rows, blank cells, duplicate IDs, mixed date formats.
• The tool — Excel or Google Sheets? pandas or SQL? Which SQL database?

Spreadsheet formulas

Excel and Google Sheets share many functions but not all of them, and some functions only exist in newer versions. Always say which one you use. If your locale uses semicolons instead of commas between formula arguments, mention that too.

I'm using [Excel for Microsoft 365 / Google Sheets].
My data is in [sheet name], rows [2–N], with these columns:
[list columns with letters, names, and data types]

I want to [goal, in plain words].
Give me the formula, say exactly which cell it goes in, and explain each part. If there's a simpler alternative using a pivot table, mention it briefly.

pandas and SQL

For code, the column names and types matter even more because a typo means an error. Paste the schema or the output of a quick inspection, like df.dtypes and df.head() in pandas, or the CREATE TABLE statement for SQL.

SQL dialects differ — date functions in PostgreSQL, MySQL, SQLite, and SQL Server aren't the same — so name yours.

I'm using [PostgreSQL / MySQL / SQLite / pandas 2.x in Python].
Here is my table structure:
[paste CREATE TABLE or df.dtypes output]

Here are 5 sample rows (anonymized):
[paste rows]

Write a query/code that [goal]. Handle [known quirk, e.g. NULL amounts, duplicate rows].
Add comments explaining each step, and tell me one quick check I can run to confirm the result is correct.

A typical answer might look like this:

import pandas as pd

df = pd.read_csv("orders.csv", parse_dates=["order_date"])
paid = df[(df["status"] == "paid") & df["amount"].notna()]
monthly = paid.groupby(paid["order_date"].dt.to_period("M"))["amount"].sum()
print(monthly)

Sanity-check the numbers

Language models are not calculators. If you paste a big table and ask "what's the average?", the model may produce a number that looks right but is off. Two safer patterns:

1. Ask for the method, not the answer. Have the AI write a formula or code, then run it yourself on the real data. The computer does the arithmetic.
2. Use AI tools that execute code. Some assistants can run code on uploaded files. That's far more reliable than mental math — but still check the code it ran.

Then sanity-check the result like an analyst would:

• Row counts — does the filtered data have roughly as many rows as you expect?
• Totals — does the sum of the monthly figures match the overall total?
• Spot checks — calculate one month by hand or with a simple filter.
• Plausibility — is any value impossibly large, negative, or zero?

Protect private data

You rarely need real data to get good help. Instead:

• Share column names and types plus a few made-up rows with the same shape.
• Replace identifiers with placeholders (customer_001, customer_002).
• Ask the AI to generate synthetic sample data for testing your formula or code.
• Run the final code on the real data locally, where it never leaves your machine.

The pattern to remember

Describe the structure, name the tool, state the goal and the quirks, and ask for a method you can run. Then verify the output before anyone makes a decision based on it.