Checks a finished workbook before you send it, with cross-footed totals, units and signs, order-of-magnitude checks, outliers, broken references and hard-coded numbers, ending in a pass, fix or ask list by cell. Use for "run gbiz-sanity-check", "check my spreadsheet", "sanity check these numbers", "check this before I send it", "find errors in my Excel", "do these totals add up", "review my workbook", "spreadsheet error check", part of the Claude for Business Graduates Pack by Polar Bear.
Installs into .claude/skills of the current project.
Are you the author of Gbiz Sanity Check?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/polar-bear-org-gbiz-sanity-check)
---
name: gbiz-sanity-check
description: Checks a finished workbook before you send it, with cross-footed totals, units and signs, order-of-magnitude checks, outliers, broken references and hard-coded numbers, ending in a pass, fix or ask list by cell. Use for "run gbiz-sanity-check", "check my spreadsheet", "sanity check these numbers", "check this before I send it", "find errors in my Excel", "do these totals add up", "review my workbook", "spreadsheet error check", part of the Claude for Business Graduates Pack by Polar Bear.
---
# Spreadsheet Sanity Check
## When To Use
You are about to send numbers upward and one wrong cell would cost you trust. The check answers one question before you press send: does every number in this book hold up, and which ones need a fix or a question first?
## When Not To Use
If the book is not finished, or you do not yet understand how it works, run the Spreadsheet Explainer first. For a text answer from Claude rather than a workbook, use the AI Output Check.
## Inputs
- The finished workbook (Claude for Excel, or upload the .xlsx in a chat with code execution turned on), from a trusted source
- The headline numbers you will quote, and a reference you trust for each (last period, a published total, a budget)
- Who receives it and when
If you have none of this, I start from the workbook alone, check structure only, and mark the output as a first draft.
## Approach
Spreadsheet error checks (cross-footing, units, magnitude, hard-coded numbers) are standard finance practice, described here generically, with the FAST Standard's transparency rules (fast-standard.org). Claude for Excel's own documentation says it is not for audit-critical calculations without verification, so Claude flags and you decide. The failure it prevents: a total that is off by a factor of a thousand because one tab is in thousands and another in units, spotted by the director in the meeting, not by you the night before.
## Workflow
1. Ask at most three questions: the headline numbers and their reference points, the units the book should be in, and the sign convention for costs.
2. Cross-foot: row totals summed equal column totals summed; every subtotal adds to its total.
3. Check units and signs: currency, thousands versus units, percentages versus percentage points, costs consistently positive or negative.
4. Check order of magnitude: each headline number against its reference; anything off by a factor of ten or more is flagged.
5. List outliers: values far from the rest of their row or column, for a human look. Nothing is auto-corrected.
6. Find broken references: error values, links to other files, ranges that stop short of new rows. Then hard-coded numbers inside formulas and inputs typed into calculation areas.
7. Mark each check pass, fix or ask, with cell references. Claude never fixes silently; you make every change and rerun the failed checks.
## Output Format
```markdown
# Spreadsheet Sanity Check
## Workbook
[File name] | Going to: [role] | By: [date]
## Checks
| Check | Cells | Result (pass/fix/ask) | What was found | Your action |
|---|---|---|---|---|
| Cross-foot | [range] | [result] | [finding] | [action] |
| Units and signs | [range] | [result] | [finding] | [action] |
| Order of magnitude | [cell vs reference] | [result] | [finding] | [action] |
| Outliers | [cells] | [result] | [finding] | [action] |
| Broken references | [cells] | [result] | [finding] | [action] |
| Hard-coded numbers | [cells] | [result] | [finding] | [action] |
## Headline numbers confirmed
| Number | Cell | Checked against | Confirmed by you (yes/no) |
|---|---|---|---|
| [value] | [cell] | [reference] | [yes/no] |
## Decision
[You] fix every "fix", get answers to every "ask" from [owner role], and confirm each headline number before sending to [role] on [date].
```
## Done When
- Every check has a result with the cells it covers
- Every "fix" names the change you will make, and every "ask" names who to ask
- Every headline number is confirmed against a reference you trust
- Failed checks were rerun after your fixes
## Quality Bar
- Outliers are listed for a human look, never "corrected" to fit the pattern.
- No check is marked pass without the cells it covered.
- Workbooks from outside sources can carry hidden instructions; check only trusted files.
- Work tasks stay within your employer's AI policy; mask confidential figures unless the policy allows them.
- Claude flags; you fix and confirm every number before it goes upward.
## Next
Run gbiz-executive-summary (Executive Summary) to write up the numbers once they pass.
## About the makers
This pack is made by Polar Bear, a consultancy built by ex-McKinsey founders with a dream to make AI work for People, not instead of them. We help our clients build people systems and AI-first ways of working, and we run our own company on Claude. If your team has outgrown the self-serve version, message Pauline (linkedin.com/in/paulinebertry).