The Spreadsheet Station

Every office has one: the spreadsheet ritual. Export from one system, export from another, paste into the template, fix the dates that came in three formats, hunt the row that breaks the total, send. Monthly, quarterly, or — condolences — weekly. It's Cook Your Own Data all over again, except this time the numbers matter to other people, which changes one thing completely: being wrong is expensive.

So this lesson teaches the two-part discipline that separates office data automation from hobby data automation: the cooking (which you already half-know) and the cross-checking (which is the actual job).

What you'll plate today

Your real spreadsheet grind, cooked end to end — with a checksum ritual that proves the output right before anyone else sees it.

Ingredients

• Your prep-station kitchen
• The real files from your grind: the exports, and one finished past edition (last month's completed sheet — it's your answer key)
• About 35 minutes

Cook

1. The seatbelt, then the shape

Work numbers are work data — Lesson 1's house rule runs first (columns only, mask decision). Then teach the grind:

In grind/ you'll find this month's raw exports plus last-month-DONE
— a finished edition I made by hand. That file is the answer key.

Step one: read the answer key and describe my ritual back to me —
what got merged, what got cleaned, what the final shape is, what
the totals are. Write it as GRIND.md, a recipe in plain words.
I'll correct it.

Correcting GRIND.md surfaces the folklore — "column F is manually overridden for the Kim account, always" is exactly the rule that lives in your fingers and nowhere else. Say it out loud once; it's automated forever.

2. Cook against the answer key

Here's the move that makes this trustworthy, and it's borrowed straight from real engineering:

Now run GRIND.md against LAST month's raw exports — the ones that
should produce the answer key. Then compare your output to
last-month-DONE, cell by cell where it counts: row counts, totals,
each section sum. Report every mismatch. Don't touch this month yet.

If it matches: your recipe is proven on known ground. If it doesn't, this is the good part — each mismatch is either a bug in the recipe (fix it now, on data where you know the truth) or, occasionally, a mistake in your old manual work. Yes, really: hand-made spreadsheets carry hand-made errors, and the machine just found one. Either way you win, and either way you learned it on last month's numbers instead of this month's meeting.

3. Now this month — with the checksum ritual

Run the proven GRIND.md on this month's exports. Then, before
showing me the result, run the checks: row counts in vs out
(explain any drop), every section total vs the grand total,
dates all in range, and flag any value that moved more than 30%
from last month with a ⚠ — big moves are sometimes real and
sometimes a broken export, and I decide which.

Those ⚠ flags are the station's soul. A 40% jump in one account is either the quarter's biggest story or a duplicated export — and both deserve a human eyebrow before the sheet ships. The machine's job is to make the eyebrow-raising unmissable.

4. Station it

Create a skill "grind": seatbelt ritual on new exports, run
GRIND.md, run the checksums, ⚠ the big movers, save to
[month]-draft.xlsx, and say "checks passed" or list what didn't.
Save point "the spreadsheet station", CHORES.md done, and log it.

The output is named -draft on purpose. Same law since Lesson 3: it ships when you've tasted it, under your name, never before.

When it burns

• The exports are .xlsx with merged cells and decorative headers — the classic corporate lasagna: "the real data starts at row [N]; unmerge and flatten before anything else." Say it once, into GRIND.md it goes.
• Step 2 mismatches everywhere — usually one systematic cause, not many: date formats, a filter you apply by hand without thinking, or the answer key including something the exports don't cover. Ask: "what single difference would explain most of these mismatches?"
• Korean/currency encoding gibberish — old friend, same fix: "handle EUC-KR/CP949 and strip currency symbols before summing." (It's in your First Kitchen scar tissue already.)
• The numbers feed a system you paste into, not a file — fine: "give me the final table as paste-ready TSV." The station cooks; you plate into whatever dish the company owns.
• You're tempted to skip the checksums "just this once" — that sentence is how spreadsheet horror stories start. The checks cost nine seconds. The meeting where a wrong total gets caught costs a quarter of trust.

Order up

□ GRIND.md holds the ritual — folklore rules included
□ The recipe reproduced last month's DONE sheet (or taught you why not)
□ This month cooked with checksums passed and ⚠ movers reviewed
□ Skill "grind" standing by; output ships as -draft, under your name
□ CHORES.md: sixth DONE — and it was probably the biggest one

Next up — Lesson 08: The Presentation Plater. The numbers are cooked; now they have to face a conference room. Messy doc in, presentable outline out — without surrendering your judgment about what matters.

Stuck on a step? Question box below.