alchaincyf/huashu-excel

*“Give me an Excel, and I’ll give you a professional report.” * *A truly skilled data analysis expert. *

Full-process data analysis and Excel · Health checkup → cleaning → alignment → analysis → reconciliation → delivery

**“How did you calculate this number?” ** — Everyone who has handed in a number has been asked this. If you can’t answer, it’s usually not because of poor memory, but because the number was never traced from the start.

** And when AI makes a mistake, it won’t report an error. ** If you write code and make a mistake, it throws an exception; When you do data analysis, nothing happens if you make a mistake. A +161% error, When delivered, it looks exactly like the correct answer.

Throw it a messy Excel file—the title takes up the first row, the header level is two, the region column uses merged cells, Amounts are in the thousandth place, with ‘East China Subtotal’ in between, and the three lines at the end are ‘Total/Percentage/Year-on-Year.’ The mainstream method calculates the monthly total for this table, which is 161% higher than the true value, with zero errors, zero NaN, and zero warnings.

The whole reason this skill exists is to make that rhetorical question answerable. It reads the original cells before moving pandas treats the “total” line on the table as a free verification tool and reconciles accounts, running it once before delivery Master check—If you don’t pass, you won’t get a number.

Cross-agent universality—Claude Code, Cursor, Codex, OpenClaw, Hermes can all be installed.

Check the effect comparison · Install and use it immediately · Eight-step process · Five different ones · Chart · Report

---

Compare the results

The same table, the same question: "What was the total sales in January? **」

Mainstream Approach: pd.read_excel()

If you find the meter header position is wrong, add header=; if the numbers can’t be read, clear the millenths. It already seems pretty comprehensive.

There are three mistakes: none of them will be reported incorrectly:

• 合计 line 15,368,317 was treated as a store
• 华东小计 6,157,150 was once again treated as a store
• One line is a complete duplicate caused by the pasting accident, counted twice

huashu-excel

During the medical exam, they listed all the questions without guessing a single one:

Then take the cleaned details and check the summary line inside the table:

**The 10 yuan difference error was made by the person who created the test data and filled in the sub-timer error. ** Relative error of 0.00%, which is something the naked eye can never detect, and the script is extracted in one go.

Panko (1998) reported that 86% of spreadsheets contain errors. European spreadsheet risk interest group Since 1995, EuSpRIG has continuously documented incidents in the media—most notably Reinhart-Rogoff In the paper, the Excel selection area was selected 5 times less, so the +2.2% GDP growth rate is calculated as −0.1%, At that time, that paper was the main academic basis for global austerity policies.

---

Ready to use as soon as you install it

Or directly clone to the directory where your agent reads skills:

| Runtime | Path |