clean-data-xls

作成者： anthropic

雑多なスプレッドシートデータを整理 — 空白のトリム、大文字小文字の不統一を修正、文字列として保存された数値を変換、日付を標準化、重複を削除、フラグを付ける…

npx skills add https://github.com/anthropics/financial-services --skill clean-data-xls

Clean Data

Clean messy data in the active sheet or a specified range.

If running inside Excel (Office Add-in / Office JS): Use Office JS directly (Excel.run(async (context) => {...})). Read via range.values, write helper-column formulas via range.formulas = [["=TRIM(A2)"]]. The in-place vs helper-column decision still applies.
If operating on a standalone .xlsx file: Use Python/openpyxl.

If a range is given (e.g. A1:F200), use it
Otherwise use the full used range of the active sheet
Profile each column: detect its dominant type (text / number / date) and identify outliers

Issue	What to look for
Whitespace	leading/trailing spaces, double spaces
Casing	inconsistent casing in categorical columns (`usa` / `USA` / `Usa`)
Number-as-text	numeric values stored as text; stray `$`, `,`, `%` in number cells
Dates	mixed formats in the same column (`3/8/26`, `2026-03-08`, `March 8 2026`)
Duplicates	exact-duplicate rows and near-duplicates (case/whitespace differences)
Blanks	empty cells in otherwise-populated columns
Mixed types	a column that's 98% numbers but has 3 text entries
Encoding	mojibake (`Ã©`, `â€™`), non-printing characters
Errors	`#REF!`, `#N/A`, `#VALUE!`, `#DIV/0!`

Show a summary table before changing anything:

Column	Issue	Count	Proposed Fix

Prefer formulas over hardcoded cleaned values — where the cleaned output can be expressed as a formula (e.g. =TRIM(A2), =VALUE(SUBSTITUTE(B2,"$","")), =UPPER(C2), =DATEVALUE(D2)), write the formula in an adjacent helper column rather than computing the result in Python and overwriting the original. This keeps the transformation transparent and auditable.
Only overwrite in place with computed values when the user explicitly asks for it, or when no sensible formula equivalent exists (e.g. encoding/mojibake repair)
For destructive operations (removing duplicates, filling blanks, overwriting originals), confirm with the user first
After each category of fix (whitespace → casing → number conversion → dates → dedup), show the user a sample of what changed and get confirmation before moving to the next category
Report a before/after summary of what changed