Agent Skills
› mohitagw15856/pm-claude-skills
› spreadsheet-audit-live
spreadsheet-audit-live
GitHub在Cowork沙箱中直接打开真实电子表格，追踪公式以检测硬编码、范围错误及单位不一致。生成按损害排名的发现报告、验证账本和修复建议，确保数据准确后再展示。
触发场景
检查电子表格是否安全可用
审计模型中的数字准确性
排查数据汇总错误
展示前验证报表
安装
npx skills add mohitagw15856/pm-claude-skills --skill spreadsheet-audit-live -g -y
SKILL.md
Frontmatter
{
"name": "spreadsheet-audit-live",
"description": "Audit the user's REAL spreadsheet by opening it in the Cowork sandbox — not by reading a description of it. Use when asked to check this sheet before we present it, audit the model in my Drive, why don't these numbers add up, or is this spreadsheet safe to build on. Pulls the file via the Google Drive connector (or an uploaded .xlsx), opens it programmatically in the sandbox to trace formulas, hunts hardcodes \/ broken ranges \/ unit mixes, and produces a ranked findings artifact with a verified-vs-suspect ledger and a fix list."
}
Spreadsheet Audit (Live)
Spreadsheets fail silently — the SUM that stops at row 40, the hardcoded 1.1 typed over a formula, the column that's monthly in one section and annual in another. In Claude Cowork this skill opens the actual file in the sandbox and traces the real formulas, so the mistake is caught before the meeting corrects it.
What This Skill Produces
- Findings ranked by damage — each with cell location, what's wrong, and what it's currently mis-stating
- The verified/suspect ledger — which load-bearing outputs were traced clean and which remain unverified (unverified ≠ wrong; the label is the honesty)
- The fix list — ordered, with make-it-robust upgrades where they matter
- A findings artifact — the above as a shareable report, plus (on request) a corrected copy of the sheet
Required Inputs
Ask for these if not provided:
- The sheet — a Drive file/link or an uploaded
.xlsx; the audit needs the real formulas, not a screenshot - The stakes — what decision it feeds (budget approval? pricing? a board number?) — depth and ranking follow the damage potential
- Growth pattern — does data get appended? The fragility check keys on it
Framework: The Hunt Rules
- Hardcode hunt first — constants typed over formulas: invisible, wrong-forever. Flag every constant where a column is otherwise formulaic.
- Range-edge check on every aggregate — SUM/AVERAGE/LOOKUP vs the data's real extent (the stops-at-row-40 error).
- Unit & time-grain consistency — monthly-vs-annual, currency, thousands-vs-units mixes at every junction (tell: a ratio ~12× or ~1000× off).
- Trace the load-bearing outputs — the 3–5 numbers the sheet exists to produce get full precedent traces.
Execution (Cowork)
- Get the file — via the Google Drive connector, download the sheet (or take the uploaded
.xlsx). Never audit from values alone. - Open in the sandbox — use the spreadsheet tooling (the
xlsxskill / a Python pass with openpyxl) to read formulas, not just cached values. Enumerate sheets, named ranges, and the formula map. - Run the hunt rules programmatically — scan for constants-in-formula-columns, aggregate ranges shorter than the data, and cross-section grain mismatches; trace precedents for the headline outputs.
- Rank by damage — location × what-it-mis-states × the decision it feeds.
- Emit the artifact — findings + verified/suspect ledger + fix list. On request, write a corrected copy (never overwrite the original) with the fixes applied and a change log.
Guardrails: read-only on the source unless a corrected copy is explicitly requested, and then only as a new file; label unverified outputs honestly; if the connector/file is unavailable, say so — don't audit a sheet you couldn't open.
Output Format
A Spreadsheet Audit artifact:
Verdict
Safe to present / Fix before presenting / Do not trust — one line why
Findings (ranked by damage)
| # | Cell/range | Class | What's wrong | Currently mis-stating |
|---|
Verified / suspect ledger
| Output | Traced | Status |
|---|
Fix list
- [fix] — [robust upgrade if relevant]
Quality Checks
- Formulas were read from the real file, not inferred from values
- Every finding names an exact cell/range
- Load-bearing outputs are each marked verified or suspect (none silently assumed)
- The original file was not modified; any corrected copy is a new file with a change log
- The verdict follows the findings, not optimism
Anti-Patterns
- Auditing the screenshot — open the actual formulas.
- Overwriting the source — corrections go to a copy.
- Claiming "looks fine" without tracing the headline numbers.
- Cosmetic tidying dressed up as an audit — protect the decision, not the formatting.
Example Trigger Phrases
- "Audit the budget model in my Drive before the board call."
- "Check this spreadsheet — why don't the totals add up?"
- "Is this sheet safe to build on? Open it and trace it."
- "Someone left me this model; hunt it for hardcodes in Cowork."
版本历史
- 471c606 当前 2026-07-22 14:01
