1 · Structure and tidy thinking

Take a messy spreadsheet, have an agent tidy it, verify the result by hand, and document your supervision. This is your Stage 1 Structure Audit — your capstone launches on the habits it builds.

At a glance


Time	~80 minutes, in class
Work with	A partner (or your capstone team)
You need	An agent (Claude or Manus) + a spreadsheet (Google Sheets or Excel)
You hand in	A tidied CSV + a 150–200 word verification note
Graded on	Traceability — did you actually check the agent’s work?

The scenario

Kopi Kembali’s finance team tracks outlet revenue in a spreadsheet built for humans to read, not machines to analyse. Head office wants outlet totals for the first half of 2026, by category.

You have two files:

kk-outlet-revenue-2026H1.xlsx — the workbook the finance team actually uses.
kk-outlet-revenue-2026H1.csv — what they got when they clicked Save As → CSV.

Open both. Part of this lab is discovering what the CSV export quietly destroyed.

Step 1 — Look before you touch · 10 min

Open the workbook and, in writing, before any agent sees the file, list at least three structural problems. The checklist to run against it:

Merged cells
Totals or subtotals mixed in with the data
Values that live in formatting (a colour, bold) rather than in a cell
Multiple facts crammed into one cell
Repeated values left blank because they seem “obvious”
Inconsistent or unclear labels
Blank or duplicate rows
Text notes where a value should be (e.g. “Updated by Sarah”)

Step 2 — Compare the CSV · 5 min

Put the workbook next to the CSV export and write down anything the export lost. If a meaning survived only as a colour, where did it go?

Step 3 — Direct the tidy · 20 min

Give the agent the file and your list — not the other way round. Do not let it decide what is wrong; you already did that.

This spreadsheet is structured for reading, not analysis. Convert it to tidy
long format with columns: outlet, region, category, month, revenue_sgd.

Before you transform anything, list every structural problem you find and how
you propose to handle each one. The problems I have already spotted are:
  - [your problem 1]
  - [your problem 2]
  - [your problem 3]

Wait for my approval, then proceed — reporting the row count before and after.

Then approve, amend, or reject each proposed handling. That decision log is the heart of your verification note. Take a screenshot of the agent’s profile and its response for your orchestration log.

Step 4 — Verify by hand · 15 min

The agent returned something clean-looking. Now prove it is right.

Pick one outlet and one month. Trace that tidied figure back to the workbook, cell by cell.
Check the grand total against what the workbook’s own TOTAL row claims.
Explain any difference in one sentence — do not hide it.

Step 5 — Write the note · 10 min

150–200 words covering: what was structurally wrong, what you told the agent to do about each problem, what you checked by hand, and what remains uncertain.

Verify & submit

Before you submit, tick every box:

Your tidy file has one observation per row and one variable per column
No total, subtotal, or footnote row survives as data
Every value that existed only as formatting is either recovered or declared lost
One figure traced end to end by hand, and named in the note
The note explains differences from the workbook’s own totals rather than hiding them
Column names are clear and consistent (lower_case_with_underscores)
Verification note is 150–200 words

Then submit:

The tidied dataset — save it in your team’s Google Sheets Analysis Workbook or linked Drive evidence folder; link the exact tab from your portfolio.
The verification note — add it to your individual portfolio and link the relevant Project Hub row.
The checkpoint record — use the Checkpoint Record Template (Assessment Pack folder in Google Drive). Submit your individual copy through xSiTe, not the living team files.

If this were your capstone

This verification habit is exactly what you will do for each stage of the capstone. The agent profiles, cleans, analyses, visualises, and recommends. You check, question, and decide — and your orchestration log follows this same template, stage by stage.

Before Stage 1, frame the stakeholder’s decision with SCQA:

Situation — the current state (“We post daily but don’t know what works.”)
Complication — the gap (“We waste effort on underperforming platforms.”)
Question — what you must answer (“Which platforms and formats drive engagement?”)
Answer — what you will recommend (“Shift 60% of effort to video on Instagram.”)

Your analysis is designed to test the Answer. Start documenting your supervision from day one.