Verified against Claude for Excel · 2026-07-30
Get Claude for Excel to show its formulas, not just narrate a number
A spreadsheet-analysis prompt for Claude for Excel and Sheets that requires real live formulas written into named cells, a chat-side manifest of what went where, and explicit data-quality flags, so the analysis can be re-run and audited instead of trusted on faith.
The prompt
Ready to copy — highlighted parts are example details you can swap.
Work directly in the spreadsheet, not just in the chat panel — write the actual formulas into cells so the analysis lives in the workbook and can be re-run if the underlying data changes, not just narrated as a one-time answer. QUESTION Which three sales regions had the highest quarter-over-quarter growth in Q2 2026, and by how much? WHERE THE DATA LIVES Sheet 'Sales_2026' has columns: Date, Region, Rep, Amount, Product. Roughly 4,800 rows covering Jan-Jun 2026. WHERE TO PUT THE WORK A new sheet called 'Q2_Growth_Analysis' — don't add anything to Sales_2026 itself. Do not overwrite any existing data or formulas outside this range — if the analysis needs a helper column or a summary area, add it in the specified location, not wherever happens to be empty. EXISTING FORMULAS TO PRESERVE Columns F and G on Sales_2026 already contain commission formulas referencing this data — don't restructure the source sheet in a way that would break those references. WHAT TO PRODUCE 1. The direct answer to the question, as a number or short statement, stated in the chat first, before touching the sheet. 2. The actual formulas written into the cells — real Excel formulas (SUMIFS, INDEX/MATCH, a pivot, whatever fits), not a value pasted in as static text where a formula should be, so the result updates if the source data changes later. 3. Back in the chat, name every formula you wrote and which cell it lives in, so I can find and audit it without hunting through the sheet myself. 4. A DATA QUALITY note: any blank cells treated as zero, duplicate rows, inconsistent date or number formats, or mismatched units you noticed while building the formulas — flag these even if you handled them a reasonable way, since I need to know the handling happened, not just trust that it did. 5. One sentence on how sensitive the answer is to the issues in part 4 — would resolving them plausibly change the conclusion, or is the answer robust either way? CONSTRAINTS - If a column or field the question depends on is ambiguous — two columns that could both plausibly be "revenue," for instance — stop and ask which one before writing any formula, rather than picking one and proceeding. - State any filtering applied (date range, category, excluded rows) explicitly in the chat, even if it seemed like the obviously correct filter to apply — do not silently narrow the dataset without saying so. - If a formula would be materially clearer as several simpler steps across a few helper cells rather than one dense nested formula, prefer the simpler version — a formula I have to reverse-engineer defeats the point of an auditable analysis. - Currency figures formatted with the existing workbook's currency style, growth percentages to one decimal place. BEFORE FINISHING Re-check that every written formula actually references the correct range — a formula that looks right but points one row off produces a plausible-looking wrong answer that is the hardest kind of spreadsheet error to catch by eye.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
Claude for Excel and Sheets can act directly in the workbook rather than only describing an answer in chat, and asking for narration only wastes that specific capability, producing a number nobody can re-run when the source data changes next month — this prompt's core bet is forcing the mechanism to be used for what it is actually for: real formulas that live in the file and stay live. Naming an exact output location and forbidding overwriting outside it addresses a real risk unique to a tool with write access to a shared file: spreadsheets accumulate other people's formulas and formatting nobody wants disturbed, and an unscoped 'analyze this' ask gives the model no signal about where it is safe to write, which matters far more here than it would for a chat-only assistant that can only describe changes rather than make them. Requiring a chat-side manifest of which formula lives in which cell provides the actual audit trail a spreadsheet user needs — 'what changed and where,' not just 'the answer is X' — because without it, the value of writing real formulas into cells is undermined by nobody knowing where to go check them without manually scanning the whole sheet cell by cell. Data-quality flagging even when handled reasonably, plus the sensitivity sentence, matters more in a live-editing tool than in a narration-only one: a tool that quietly picks one convention — blank-as-zero versus blank-excluded — changes the actual numbers written into a shared file that someone else might open next, not just a chat answer someone might forget within the hour, which raises the real stakes of an unflagged assumption. The closing off-by-one-row check targets a genuinely spreadsheet-specific failure class: a formula referencing the wrong range looks completely normal syntactically and produces a plausible number, making it the error most likely to survive a casual glance at the finished sheet.
Verified against
Claude for Excel Beta (Sonnet 4.6) · 2026-07-30
Changelog
- 2026-07-30 — Initial publish, verified against Claude for Excel (beta) on Sonnet 4.6.
Building this for real?
This is a free starting point. If you'd rather have what Scult builds built and running for your business, that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
