Cook Your Own Data

So far you've cooked pages — things meant to be looked at. Today the ingredient changes: your own data. Card statements, a workout log, household spending, that spreadsheet of half-tracked something — every adult owns messy tables they keep meaning to look at properly. You now employ a cook who reads tables tirelessly.

This is also the lesson where the course quietly turns from "fun" to "useful at life." A page is a thing you show people. A data tool is a thing that answers you.

What you'll plate today

A report.html built from a real file of yours: totals, biggest items, a breakdown — and one insight you didn't know was in there.

Ingredients

Any real CSV file. Best sources: your bank/card site (statements export as CSV or Excel), or any spreadsheet via File → Save as → CSV. Aim for a few dozen rows minimum.
Your first-kitchen folder — drop the file in as data.csv
About 30 minutes

Cook

1. The privacy rules (read before cooking)

Two honest rules this kitchen lives by:

Local file ≠ private conversation. What the agent reads gets processed by the AI provider, like a chat message would. So the rule of thumb: don't feed the agent data you wouldn't paste into a chat window.
Mask first, analyze second. Account numbers, card numbers, other people's names — strip them before the analysis, not after.

The nice twist: the agent does the stripping too. Peek safely first — column names only:

Open data.csv and show me ONLY the column names and the first two
rows. Don't display anything else yet.

Then, if anything sensitive exists:

Make a copy called data-clean.csv with the [account number / name]
column removed. We'll work from the copy only.

2. Order the report

Read data-clean.csv — it's [say what it is: my card spending for
June / my running log / …].

Build report.html showing:
1. The total, big at the top
2. The five biggest items
3. A breakdown by [category / month / whatever fits], as a simple
   bar chart — build it with plain HTML/CSS, no external libraries

Keep it one file, match the style of my index.html, and open it
when done. Work locally — do not send this data anywhere.

Watch your unread mess turn into a report you'd forward to yourself. Taste, then adjust (Lesson 3 muscle): "sort the categories biggest first," "make the total impossible to miss."

3. Ask the data a question

Reports show what you asked for. The agent can also show what you didn't:

Look at the data again. What's one thing in here I probably haven't
noticed? Answer in two sentences.

This little prompt is worth the whole lesson. "Your Tuesday spending is double any other day" is the moment data stops being homework and starts being a conversation.

4. Save the pattern

Make a save point: "first data report". Then add a rule to CLAUDE.md:
when I bring a new data file, always start by showing only its
column names and asking me what to mask.

That rule is your seatbelt for every future dataset.

When it burns

Korean text shows as gibberish (ê°€ or ???) — encoding. Say: "the file might be EUC-KR or CP949, not UTF-8 — handle the encoding." Excel exports from Korean banks do this constantly; the agent fixes it in one pass.
Numbers came out as text ("₩1,200" won't sum) — "the amount column has currency symbols and commas — clean it to plain numbers before calculating."
The real headers are on row 3 — bank exports love decorative headers. "The actual column names are on row [N]; ignore everything above."
The chart looks wrong, not the data — split the taste: "is the data right and the chart broken, or is the data itself wrong? Show me the totals as plain text first."

Order up

□ You previewed columns before showing the agent everything
□ Sensitive columns got masked into a working copy
□ report.html answers something you actually wondered about
□ The agent told you one thing you hadn't noticed
□ CLAUDE.md has your mask-first rule

Next up — Lesson 08: The Daily Special. Today's report needed you to order it. Next lesson, something gets cooked without you in the room — your first automation.

Stuck on a step? Question box below. Data files are the wildest ingredient so far — real questions about weird CSVs make this lesson better for everyone.