Excel Prompts for AI: Formulas, Analysis, and Verification

Useful Excel prompts tell the AI assistant what each column means, how the workbook is organized, what result you need, and how you will check it. Ask for a proposed formula or analysis with test cases before applying it to important data. A plausible explanation is not proof that the workbook calculates correctly.

Keep reusable instructions separate from the workbook’s private contents. fenbs can hold reviewed prompts, agreed data definitions, and lessons from corrected formulas so connected assistants can read the same working rules. The workbook and its calculations remain in your spreadsheet tools.

What workbook context should every prompt include?

Start with the Excel version, the sheet and table names, the relevant column headers, and a few fictional rows. Say whether dates are real date values or imported text, what a blank means, and whether amounts are whole dollars or cents. These details often determine whether a proposed answer matches your data.

Also explain the unit represented by a row. One row per order is different from one row per order line. Counting rows in the second case does not count orders. If the assistant does not know this distinction, a well-formatted summary can still answer the wrong question.

Excel version and platform: [facts]
Relevant sheet/table: [names]
One row represents: [unit]
Columns and meanings: [headers, types, definitions]
Blank, zero, and missing-value rules: [rules]
Date format and reporting period: [facts]
Fictional sample rows: [small representative sample]
Needed result and destination: [goal and cell/column]
Allowed changes: propose first; do not overwrite source data.
If any required detail is missing, ask before choosing a method.

If you use an Excel table, give its actual name and headers. Microsoft’s structured-reference guide↗ (opens in a new tab) explains how formulas can refer to table and column names. Ask for a result compatible with your workbook rather than accepting a formula that assumes unrelated names or a different layout.

How do you ask for a formula you can verify?

Describe the result in ordinary language first. Include cases where the answer is obvious, then state how to handle exceptions. “Total the amount for completed orders” is incomplete if the data also contains refunds, repeated order numbers, and blanks. Tell the assistant which of those records belong in the total.

Using the workbook context above, propose a formula for [result].
Include records only when [explicit conditions].
Treat blank values as [rule] and zero as [rule].
Expected examples:
- [small input] should return [known result].
- [boundary or missing-value case] should return [known result].
Return the formula, where to place it, and an explanation of references.
List compatibility assumptions and three manual checks.
Do not hide errors with a replacement value unless I approve that rule.

For a date filter, specify whether the start and end boundaries are included and whether timestamps may appear. For a lookup, say what to do with no match and multiple matches. Those are data decisions for you to supply, not details the assistant should guess from a column title.

Test on a copy and compare a few rows by hand before filling the result down. Microsoft’s formula overview↗ (opens in a new tab) explains references and calculation building blocks. Use it to check the proposed construction when an explanation feels unclear, rather than adding layers to a formula you cannot yet inspect.

What prompt helps explain or debug an existing formula?

Provide the formula exactly as written, its location, the relevant input values, the current result, and the result you expected. Include the workbook context so the assistant can distinguish a reference problem from a misunderstood data rule. A screenshot alone may hide the formula or the type of a stored value.

Explain this formula using the supplied workbook context: [formula].
Location: [cell]. Relevant inputs: [values and types].
Actual result: [result]. Expected result: [result and reason].
Trace the references and intermediate calculations.
Separate confirmed problems from possible causes needing a check.
Suggest the smallest correction and test it against the supplied cases.
Do not replace errors with zero merely to make the output look clean.

Ask for the smallest explanation that resolves the discrepancy. If the issue is a number stored as text, changing the entire calculation may mask it. Keep a failing row and the corrected expectation as a reusable test case, using fictional values when the example will be shared.

Once the cause is confirmed, preserve the lesson with the prompt rather than trusting everyone to remember it. fenbs keeps Lessons learned and Skills and prompts with the shared work record, so another assistant can read the corrected rule before repeating the analysis.

How do you prompt for cleanup without losing valid rows?

Request a cleanup plan before a deletion or replacement. Define a duplicate using the columns that identify a record; the same customer name on two legitimate orders is not necessarily a duplicate. Ask the assistant to list ambiguous cases separately and preserve a route back to the original values.

Propose a cleanup plan for [table] using [definitions].
Preserve the source and propose changes in a separate output.
Check whitespace, mixed data types, missing values, and duplicate keys.
A duplicate means [specific key and rule], not simply repeated text.
Return issue type, affected example, proposed action, and review check.
Flag ambiguous records for a person instead of deleting them.
Reconcile the row count and key totals before and after any approved change.

Excel offers ways to highlight duplicate values so you can review them before removal; Microsoft’s duplicate guide↗ (opens in a new tab) describes that distinction. In your prompt, ask for the same separation between identifying a possible issue and deciding that a row should be removed.

Be especially careful with text identifiers. A leading zero may be meaningful in a product code even though it looks unnecessary in a number. State which fields are labels and which are quantities before asking an assistant to standardize formats across the table.

How do you ask for analysis that answers the real question?

Give the business question and define the measure before requesting a chart. For example, “Which product category had the most returns?” could mean return count, returned units, or returned amount. Choose the measure and denominator before comparing categories, and make the reporting period explicit.

Question: [specific decision or comparison].
Measure: [definition, unit, and included/excluded records].
Group by: [field]. Reporting period: [boundaries].
Use only the supplied data. Do not infer causes from a correlation.
First list missing fields or definitions that block the analysis.
Then propose the calculation and a suitable summary table or chart.
Show totals, denominators, excluded records, and limitations.
Provide checks that reconcile the summary with the source.

Ask for descriptive findings separately from possible explanations. A higher return rate can be visible in the data without the reason being known. The next action may be checking a source record or collecting another field, rather than presenting an invented explanation as the conclusion.

A reusable work prompt needs this same discipline: stated inputs, a bounded output, and a review standard. The system prompt examples cover continuing instructions, while project management prompts show how to turn unresolved checks into explicit work.

How do you use these prompts with ChatGPT or Claude?

For ChatGPT prompts for Excel, start with the context block and ask the assistant to state what data it can actually inspect. If it cannot access the workbook, paste fictional rows and the relevant formulas. Require a proposed answer with assumptions instead of language suggesting the workbook was already changed.

For Claude prompts for Excel, use the same facts and review checks. You can label separate blocks as workbook context, task, and expected output to keep them distinct. In either app, ask the assistant to distinguish calculations it executed from results it only predicts, and verify important outputs in Excel.

For related questions, see Excel AI tools, the focused formula guide, and separate prompts for copywriting, research, and SEO.

How to do this with fenbs

1. Create a free account and add the project that uses the recurring workbook.
2. Save one reviewed Excel prompt in Skills and prompts. Include required column definitions, sample inputs, and checks without copying private workbook data.
3. Add agreed terms to the Glossary, and record continuing review requirements in Decisions and rules.
4. Connect a supported assistant and ask it to read the project context and relevant prompt before proposing a workbook change.
5. After checking the result in Excel, record the outcome on the task and add any useful lesson. Save progress with the next check and unresolved question.

fenbs does not recalculate your workbook or replace Excel’s review tools. A single formula question may only need a direct chat and a manual check. Keep the shared library when the same work recurs, and use the AI prompts hub to find the next useful template.