excel-assistant

Excel Assistant (inspect → clean → answer)

Never modify the user's original file. Copy first, clean the copy, show before/after evidence for every change, and end with a written findings note.

Inputs

Input	Required	Default	Notes
file path	yes	—	xlsx / xls / csv
goal	yes	—	e.g. 汇总各月销售额 / 找出重复客户 / 画趋势
constraints	no	—	columns to preserve, output format

If anything required is missing, ask ONCE:

请提供：① 表格文件路径；② 你想得到什么结果（一句话即可）。 可选：哪些列不能动、希望输出 xlsx 还是 csv。

Preflight self-check

python -c "import pandas, openpyxl; print('xl-ok')"

• xl-ok → automated path below.
• ImportError → tell the user exactly which import failed and offer pip install pandas openpyxl; without consent, proceed in guide-only mode: give precise manual steps instead of running code, and say so plainly.

Workflow

Step 1: Inspect before touching anything

import pandas as pd
df = pd.read_csv(PATH, encoding="utf-8-sig")   # or read_excel(PATH)
print(df.shape); print(df.dtypes); print(df.head(3))
print(df.isna().sum())

Expected: shape, dtypes, sample rows, null counts. Record these numbers — they are your before-evidence. CSV garbled? Retry encodings in order: utf-8-sig → gbk → gb18030.

Step 2: Clean with one change per step

Apply at most one fix per step, re-running inspection after each: ①去重（df.duplicated() 先看再删）→ ②补/标缺失（填充规则要写进交付说明） → ③日期列转 datetime（pd.to_datetime(col, errors="coerce") 后检查 NaT 数） → ④数值列剥离单位字符再转类型。Each step's expected result: null/dup counts move exactly as predicted; if not, undo and investigate — never chain blind fixes.

Step 3: Answer the goal

Compute the requested aggregation/trend/ranking. Expected: a number-or-table that directly answers the user's sentence from Inputs, not adjacent trivia.

Step 4: Deliver

df.to_excel(PATH_stem + "_cleaned.xlsx", index=False)

Plus a short findings.md: what was wrong, what you changed, the answer, and any rows you had to drop (count them).