version: "1.0.0" name: aetox-clean-data-xls description: ล้างข้อมูลที่เละก่อนเอาไปคำนวณ รวมยอด หรือ pivot ช่องว่างหัวท้าย ตัวพิมพ์ไม่ตรงกัน ตัวเลขที่ถูกเก็บเป็นข้อความ วันที่คนละฟอร์แมตในคอลัมน์เดียว แถวซ้ำ และตัวอักษรเพี้ยน source: https://github.com/anthropics/financial-services-plugins license: Apache-2.0 copyright: Copyright Anthropic, PBC. Full terms in the source repository

Clean up messy data

Data arrives from an export, a form, a system somebody else runs, or fifty rows somebody typed by hand. Your job is to get it to where every column has one type, every value is what it looks like, and nothing has to be retyped before it can be summed.

The failure to avoid is a cleaned file that is tidier than the original and no more trustworthy: rows quietly dropped, a value silently changed, a date read as month-day where the source meant day-month. The user is not going to diff your output against their input, so anything you change without saying is a change nobody will ever find.

Step 1: scope

Take the range the user named, or the full extent of the sheet. Then profile each column before touching it: what is the dominant type, and which cells disagree with it. That profile is the whole basis for everything below, and it is also the most useful thing you can say back if the user only wanted an opinion.

Step 2: what to look for

Two of those are worse than they look in a Thai file. 3/8/26 is ambiguous between 3 August and 8 March and the source usually does not say which — do not guess, ask, or keep both readings visible. And a year may arrive as พ.ศ.: 2569 is 2026, and a column where 2569 and 2026 both appear has two calendars in it, not one bad row.

Step 3: propose before you change

Show this table and stop:

Anything destructive — removing duplicate rows, filling blanks, collapsing two columns into one — is asked about here, not reported afterwards. A duplicate row is sometimes two real transactions of the same amount on the same day.

Step 4: clean it

You do not edit a workbook in place: sheet_write produces a new file, and the original stays exactly as the user left it. That is the safety net for this whole job, and it is worth saying in your reply so the user knows their source is untouched.

Prefer a formula over a value you computed. Where the repair can be written as one, put it in an adjacent helper column — =TRIM(A2), =VALUE(SUBSTITUTE(B2,"฿","")), =UPPER(C2), =DATEVALUE(D2) — rather than working the answer out yourself and writing the result. The user can then see what was done to every row, and undo it by deleting a column. A cleaned value with no formula behind it is a number they have to take on faith.

Write computed values in place of formulas only when the user asks for it, or when no formula expresses the fix — mojibake repair is the honest example.

The repair that matters most is not a formula at all: send the value with its real JSON type. A cleaned amount goes out as 1234.5, a cleaned date as "2026-03-08". Handing back "1,234.50" with the spaces trimmed is a column that still sums to zero.

Work in the order the categories are listed above — whitespace, casing, numbers, dates, duplicates — and show a sample after each one. A user who sees the casing pass go wrong stops you before the dedupe pass eats the evidence.

Step 5: hand it back

Say what changed, per column, with counts. Say what you did not change and why: the ambiguous dates you could not resolve, the three text entries in a number column that turned out to be "n/a", the near-duplicates you left alone because they might be real.

Never fill a blank with something plausible. An empty cell the user can see is a question they can answer in a second; a filled one is a fact they will never check. If the data is too broken to clean without deciding something only they can decide, say which decision it is and stop there.

If what they actually needed was to know whether an existing workbook computes correctly rather than whether its data is tidy, that is aetox-audit-xls.