Outcome: Turn a spreadsheet question into a reviewable analysis that preserves the source data, documents cleaning and joins, checks the result, and explains limits before anyone acts on it.

Advanced spreadsheet work is more than asking for a chart. Start with one decision question, inspect the inputs, and make each transformation visible. ChatGPT can help clean data, compare sources, create tables or charts, and explain findings, but the final workbook and conclusion still need human review.

Use this six-step workflow

1. Define the decision question. Name the population, measure, time period, comparison, and output. Ask one primary question before exploring secondary patterns.
2. Inventory before calculating. Have ChatGPT describe every sheet, column, data type, unit, date range, likely identifier, missing-value pattern, duplicate, and obvious inconsistency. Do not request conclusions yet.
3. Protect the raw data. Keep source sheets unchanged. Put cleaned data, assumptions, mappings, and excluded records on separate, clearly named sheets.
4. Validate joins and formulas. Identify the key and expected relationship between tables. Report matched, unmatched, and duplicated keys; reconcile row counts and totals before and after each join.
5. Analyze and challenge the finding. Produce the simplest table or chart that answers the question. Check whether missing records, outliers, alternative groupings, or date choices materially change the result. Do not present correlation as causation.
6. Deliver the audit trail. Put the answer first, then methods, data-quality issues, checks, assumptions, and limitations. Ask for an editable workbook plus a concise summary.

Use less capacity: Luna can handle a clearly specified inventory, formatting pass, or column classification. Use Terra for joins, data-quality reasoning, sensitivity checks, and the final explanation. Give one correction at a time before escalating further.

Copy and adapt this prompt

Analyze the attached workbook to answer: [one decision question].
Population and period: [scope]. Output: [workbook, chart, and/or brief].
First inventory sheets and columns; report missing values, duplicates, units, date ranges,
and likely keys. Do not change the raw sheets or invent missing values.
Create separate cleaned-data and analysis sheets. Document every exclusion, mapping,
join, and formula. Reconcile row counts and totals before and after each join.
Show the main result, then test whether outliers, missing records, or a reasonable
alternative grouping changes it. Explain assumptions and limits in plain language.

Illustrative SBS example

Example: A program manager uploads a de-identified attendance export and an event roster to ask which session formats had the strongest attendance rate. ChatGPT inventories the files, discovers duplicate registration IDs and inconsistent session names, proposes a mapping, reports unmatched rows, and calculates rates only after the manager approves the cleaning rules. It then creates a labeled chart and notes that attendance alone does not explain why one format performed differently.

Troubleshooting

• Check key uniqueness and join type. Request before-and-after row counts, distinct-key counts, unmatched keys, and duplicate-key examples before accepting the analysis.
• Ask for counts behind every percentage, show missing records, rerun without influential outliers, and compare a reasonable alternative grouping.
• Request separate Raw, Cleaned, Analysis, and Read Me sheets; visible formulas; a data dictionary; and a short change log.

Related SBS guides

• Work with documents, spreadsheets, presentations, and PDFs
• Review, correct, and improve an AI response
• Choose a model, reasoning effort, and speed without wasting your limits

Official guidance

• Analyze datasets and ship reports | ChatGPT use cases
• Work with files | ChatGPT Learn

Official guidance checked September 16, 2026. Features vary by product surface and workspace settings.