# Workbook Doctor > Audits an existing or AI-generated Excel workbook for hidden risks before it is used in decisions. Runs four passes — structure map, formula risk scan, repair plan, verification memo — and produces a pass/fail checklist on formula consistency, source traceability, assumption clarity, and data freshness. Use before relying on any AI-generated spreadsheet for financial decisions, reporting, or client deliverables. - Skill: `m2ai-portfolio/workbook-doctor` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add m2ai-portfolio/workbook-doctor` - Raw SKILL.md: https://api.skillmd.com/api/skills/m2ai-portfolio/workbook-doctor/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Data & Analytics - Author: m2ai-portfolio (https://skillmd.com/u/m2ai-portfolio) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/m2ai-portfolio/workbook-doctor --- # Workbook Doctor Performs a structured audit of a workbook before it drives decisions. AI tools generate clean-looking spreadsheets that frequently contain broken formula logic, hardcoded outputs masquerading as live calculations, and assumptions never labeled as such. This skill runs four diagnostic passes and produces a clear verdict: safe to use, needs repair, or do not use. ## When to Invoke Trigger on: "workbook-doctor", "audit this spreadsheet", "check my Excel", "verify my workbook", "are my formulas right", "is this model safe to use", "QA my financial model", "check before I share this", or any time the user is about to use an AI-generated workbook for a consequential decision. ## Phase 1: Workbook Map Before evaluating formulas, build a structural map of the workbook. For each sheet, identify and document: ``` Sheet: [name] Purpose: [what this sheet does] Inputs (assumption cells): [list key variable cells and their current values] Outputs (result cells): [list key output cells and their values] Dependencies: [which other sheets does this sheet pull from?] Hardcoded regions: [areas that appear to be manually entered rather than formula-driven] ``` Ask the user to paste or describe the workbook structure if direct file access is unavailable. The map does not need to be exhaustive — cover every sheet that drives a key output. ## Phase 2: Formula Risk Scan Run each check below on the formulas and structure identified in Phase 1. For each finding, record: sheet name, cell reference (if known), finding description, and severity. ### Check 1 — Formula Consistency - Are formulas consistent across rows in the same column, or do some rows have different formula logic? - Are there cells in formula areas that contain plain numbers instead of formulas? - Do SUM ranges cover all rows they should, or do some stop short? ### Check 2 — Live vs. Hardcoded Outputs - When input assumption cells change, do output cells change accordingly? - Are any "calculated" cells actually hardcoded constants that will never update? - Are there cells formatted to look like formulas but containing static values? ### Check 3 — Circular References - Are there circular dependencies between cells? - Does the workbook rely on iterative calculation to resolve them? ### Check 4 — Reference Integrity - Do formulas reference cells that exist, or are there broken references (#REF! errors)? - Do cross-sheet references point to the correct sheet and range? - Are there named ranges that have been deleted but are still referenced? ### Check 5 — Aggregation Accuracy - Do summary totals equal the sum of their detail rows? - Do percentages calculated from subtotals match independently calculated percentages? - Are there rounding differences that accumulate to material discrepancies? ### Check 6 — Stale Data - Are there hardcoded dates, prices, or rates that are unlikely to have been updated? - Does the workbook reference external data (market rates, FX, headcount) that may be out of date? ## Phase 3: Repair Plan For each finding from Phase 2, classify and recommend an action: ``` Finding [F01]: Sheet / Cell: [location] Issue: [description] Severity: must-fix / should-fix / optional Recommended action: [specific fix] Verification method: [how to confirm the fix worked] ``` **Severity definitions:** - **Must-fix**: the workbook will produce materially wrong outputs if this is not corrected. Do not use until resolved. - **Should-fix**: the workbook may produce correct results now but is brittle or misleading. Fix before sharing. - **Optional**: cosmetic or low-risk. Fix at your discretion. ## Phase 4: Verification Memo Issue a final verdict on four dimensions: ``` ## Workbook Verification Memo Workbook: [name] Audited: [date] ### Formula Consistency Status: PASS / FAIL / CONDITIONAL Notes: [key findings or "formulas consistent across all reviewed ranges"] ### Source Traceability Status: PASS / FAIL / CONDITIONAL Notes: [key findings or "all input figures traced to identified sources"] ### Assumption Clarity Status: PASS / FAIL / CONDITIONAL Notes: [key findings or "assumption cells clearly labeled and separated from outputs"] ### Data Freshness Status: PASS / FAIL / CONDITIONAL Notes: [key findings or "all data verified current as of [date]"] --- Overall verdict: SAFE TO USE / CONDITIONAL USE / DO NOT USE Conditions (if CONDITIONAL): [list what must be resolved before use] Open items (if DO NOT USE): [list must-fix items blocking use] ``` **PASS**: no issues found in this dimension. **CONDITIONAL**: issues exist but do not block use if the stated conditions are met. **FAIL**: issues block use of this workbook for any consequential decision. ## Verification - [ ] All six formula risk checks run — none skipped - [ ] Every finding has a location, not just a category - [ ] Repair plan includes verification method for each must-fix item - [ ] Verification memo issued with explicit overall verdict - [ ] User informed of any must-fix items before they use the workbook ## Source Nate Jones newsletter (2026-05-27) — "The deck got forwarded with a wrong number inside. The Trust Layer's two-model review is built to catch exactly that." The workbook doctor implements the Trust Layer Guide's Prompt Kit 1 (Workbook Doctor). Core insight: AI-generated spreadsheets look correct but frequently have hardcoded outputs that do not respond to changed assumptions — the formula risk scan is the only reliable way to catch this without running every scenario manually.