Writing Good Prompts for Excel: Clarity, Context & the Clear-vs-Bad Guide

A good prompt is a clear, complete instruction that gets a useful answer on the first try. This tutorial covers what makes a prompt work for Excel, how to tell a clear prompt from a bad one, and how to fix a prompt when the AI gets it wrong.

This tutorial is written for Excel with AI — the example prompts below are things you type into an AI assistant such as Microsoft 365 Copilot in Excel or ChatGPT.

Two ways to send these prompts:

• Copilot in Excel works inside your workbook. It can already see your sheet, so you can name columns directly and it often applies the result (a formula, a fix) for you.
• ChatGPT (or any chat AI) runs outside Excel. You paste your layout and a few sample rows into the chat, then copy the answer back into your sheet.

The prompting skills are the same either way — only how the AI reaches your data changes.

Table of Contents

• What Makes a Prompt Good
• The Four-Part Prompt Formula
• Clear vs Bad Prompts
• Include Your Actual Data
• Specify the Output Format
• Iterating: Fixing a Bad Answer

What Makes a Prompt Good

A prompt is the instruction you send the AI. A good one tells the AI exactly three things: what you have, what you want, and how it should be delivered.

The AI has not opened your workbook. It cannot see your columns, your numbers, or your layout. Everything it knows comes through your words, so a good prompt is a complete briefing — like handing the task to a smart colleague who has never seen your file.

Why it matters: a vague prompt returns a vague answer that wastes your time. A specific prompt often returns a ready-to-paste formula or chart on the first try.

Example — the same request, two ways:

Explanation: the clear prompt names the sheet, the columns, and the exact task. The AI can answer immediately; the vague one would need several follow-up questions.

The Four-Part Prompt Formula

Most good prompts contain four parts. You do not need all four every time, but each one improves the answer:

1. Role — who the AI should act as. "Act as an Excel expert."
2. Context — your data and situation. "My sheet has 500 rows of sales data, row 1 is headers."
3. Task — the exact thing you want done. "Write a formula that..."
4. Format — how the answer should look. "Give me the formula, then explain it in one sentence."

Example — a complete prompt using all four:

Act as an Excel expert.

I have a sales table with columns Date, Region, Product,
Units Sold, Revenue. Column A contains dates.

Write a formula that totals Revenue for the West region
in the last 90 days.

Give me the formula first, then a one-line plain-English
explanation of what it does.

Explanation: the role sets the tone, the context supplies your data, the task says exactly what to do, and the format tells the AI how to deliver it so you can paste it straight into a cell.

Note: when the task is mechanical (formulas, cleanup), the context part is the most important. Without your layout, the AI can only guess at column letters and ranges.

Clear vs Bad Prompts

The fastest way to learn good prompting is to see both sides. A bad prompt is usually bad for one of four reasons: no data, no task, no format, or too many requests at once.

Explanation: every clear prompt gives the AI a concrete job with a concrete shape. Notice what the bad versions lack: they describe a feeling ("help me", "make it better") instead of an action.

Best Practice: before sending a prompt, ask yourself — could a stranger complete this task with only my words and a copy of the spreadsheet? If not, add detail.

Include Your Actual Data

The most common reason an Excel prompt fails is that the AI never saw the data. Name your columns, tell it where the headers are, and mention anything unusual.

What to include:

• The sheet name and what each column holds.
• How many rows there are.
• Which row has headers.
• Anything messy — "column B has blank cells," "dates are stored as text."

Example:

My sheet "Orders" has 1,200 rows. Columns:
A = Order ID, B = Date (text format, like "03/12/2025"),
C = Customer, D = Amount.
Row 1 is headers.

Write a formula to count orders from the customer "Acme".

Explanation: the note that dates are text matters. Without it, the AI might suggest date functions that return errors. A single line about your messy data prevents a whole round of debugging.

Best Practice: paste 2–3 real sample rows into the prompt. Real examples tell the AI more than a paragraph of description, especially when data is inconsistent.

Specify the Output Format

Tell the AI how you want the answer. Without this, you might get a paragraph when you wanted a formula, or a formula when you wanted steps.

Common output requests:

• "Give me the formula only."
• "Formula first, then a short explanation."
• "Answer as a table with columns X and Y."
• "List the steps 1 to 5."
• "Use bullet points, one idea each."

Example:

Give me a VLOOKUP to find the Price (column E) of a Product
(column B) in the sheet "Products" by name.
Return the formula, and then 3 one-line notes on when
this formula breaks.

Explanation: the AI matches whatever structure you name. If you say "formula only," you get clean copy-paste text. If you say "table," you get a table. Formatting instructions are cheap to add and save you from reformatting the response yourself.

Iterating: Fixing a Bad Answer

The first answer is rarely the best one. Treat a bad answer as information, not failure — then adjust and ask again.

How to fix a bad answer:

• Add context — "I forgot to mention there are blank cells in column C."
• Add constraints — "This must work in Excel for Windows, no array formulas."
• Report the result — "That formula returns #VALUE! — what could cause it?"
• Rephrase the task — sometimes the question is the problem, not the model.

Example:

You: "The formula you gave me returns #VALUE!."
AI: "That usually happens when a range contains text.
Try wrapping the range with =SUM(VALUE(...)) or using
SUMPRODUCT. Here is the fixed version..."

You: "That works. Now explain why it works in one line."

Explanation: each exchange adds information the model did not have. Iterating turns one failed answer into a short conversation that ends with the exact result you need.