← AI 101

Intermediate training

Claude with Excel

Spreadsheet work where Excel does the math and Claude does the explaining, structuring, and editing. Clean messy data, build formulas you actually understand, turn numbers into narratives — without ever asking Claude to be a calculator.

18 minutes Builds on Module 3.3 Includes formula translator

Start here

What you'll be able to do after this lesson

Clean a messy export

Turn an LOS export, a rate sheet, or a partner-supplied spreadsheet from chaos into a real table you can analyze — without doing the structure work by hand.

Build formulas you understand

Write the formula in plain English first, get the Excel syntax back from Claude, and understand what every piece does before you trust the result.

Numbers to narrative

Turn a sheet of data into a chart with full labels and one paragraph of plain-English meaning — the difference between data on a screen and a decision in a room.

Three Excel moves Claude does well

Structure, formula, narrative

Structure

Turn a messy export — merged headers, blank rows, dollar signs stored as text, dates in three formats — into a clean table. The single highest-leverage Excel pattern Claude does well.

Formula work

Write the formula in plain English first, ask for the syntax, then ask Claude to explain every piece before you trust it. Trial-and-error formula building is a tax you don't have to pay.

Narrative

Numbers don't sell themselves. Claude is excellent at turning a sheet of data into the one-paragraph "what does this actually mean" that goes in the email above the attachment.

Trainer note: Excel computes, Claude communicates. Don't ask Claude to be a calculator — that's where math errors hide. Use Excel for the arithmetic; use Claude for the structure, the formula text, and the explanation.

Where the work happens

Three surfaces, no live data

Cowork with the .xlsx

The main surface. Cowork opens the file, modifies cells, writes formulas in place, saves the .xlsx back. The actual math runs in the file, not in Claude.

Main surface

Chat for "what formula"

For one-off "what's the right formula for X" questions — paste a small sample, ask the formula in plain English, get the syntax + explanation. Don't open Cowork for a single cell.

Quick formulas

The exported file from the source

The LOS export, the rate sheet CSV, the partner spreadsheet. Claude doesn't pull live data — the data has to be in a file you attach. Live numbers come from the source system.

No live data

Spreadsheet work

Five rules that keep the math honest

Show Claude the actual data structure

Paste five real rows — including the header — into the conversation. Describing "I have a pipeline export with various columns" makes Claude guess; pasting the rows shows it.

Paste a sample, not the whole sheet

For chat-based formula work, five to ten representative rows is plenty. Whole-sheet paste is wasteful and lossy — Claude is reasoning about structure, not summing the column.

Write the formula in plain English first

"Count rows where state is FL and loan_type is VA" beats "build a SUMPRODUCT with conditional logic." Translate after the question is clear, not before.

Ask Claude to explain any formula it gives you

If you can't explain every piece of the formula, you can't verify it. Always ask for the breakdown. The breakdown is the documentation.

Verify on a row you can compute by hand

Pick one row in the data, compute the expected output mentally, and confirm the formula matches. One row of verification catches most formula errors before they spread.

Prompt upgrade

Weak prompt

Build me a pipeline model.

Spreadsheet-ready prompt

Work-ready prompt

(In Cowork.) Attached is gmfs-pipeline-export.csv. Columns: borrower, loan_type, state, close_date, amount, fico. In a new cell, give me total committed dollars for purchase loans in Florida closing in October 2025. Show the formula in F2 and a one-line plain-English explanation of each piece. Verify by manually summing the first three matching rows.

Best patterns

Four spreadsheet workflows that earn the time

Data cleanup

Messy export in, clean table out. The single biggest time-saver for anyone who works with LOS data, rate sheets, or partner spreadsheets that arrive un-modelable.

Formula construction & explanation

Plain English → Excel syntax → breakdown of each piece. Use the breakdown to verify the formula does what you actually meant.

Chart selection & labeling

"Given this dataset, which chart type tells the story? Build it with axis labels, units, and a source line." Claude picks the chart type more reliably than most people.

Number-to-narrative summaries

"Here's the Q3 data. Write the one paragraph I'd put above the attachment in an email to leadership." Translate the sheet into the decision it should drive.

Sheet checks

Five things to verify on every sheet

Spot-check the math on at least one row you can compute by hand Confirm formulas reference the right cells after any sort, filter, or row insertion Verify chart axis labels, units, and date ranges — no decorative charts Confirm totals match the source — if the export says $14M committed, your model should too Never publish a number Claude generated without an independent verification step

Employee rule: A wrong formula that sums to a confident-looking number is the single most expensive Excel mistake. The verification step is where that gets caught — and it takes thirty seconds, every time.

Drive the sheet

Six exercises to build spreadsheet fluency

The point isn't to memorize Excel functions. It's to build the habit of writing the formula in English first and verifying the math afterward.

Open the formula translator in this lesson. Click each plain-English request and read the breakdown — the breakdown is the actual skill.
Clean a messy export you actually have. Time the manual cleanup against the Claude-in-Cowork cleanup. Note where Claude needed your verification.
Write three formulas in plain English first, then ask for the syntax. Compare to how you'd have built them by trial and error.
Build one chart with full labels, units, and a source line — then write the one-paragraph plain-English summary that goes with it.
Debug a broken formula. Paste the formula and the data, ask Claude to find the bug and explain it. Note what kind of bug it was.
Convert a pivot table into a narrative summary. The same numbers, in prose. Watch for what got lost in translation — that's usually the most important data point.

Completion standard

You've finished this module when you can clean a messy export in under five minutes, write any formula in plain English before touching Excel syntax, and turn a sheet of numbers into a one-paragraph narrative you'd put in front of a leader.