Lesson 10.7 · 45 minBeyond code: Claude for the rest of the business

What you will have

The student has three real work products for their own business, made with Claude and each checked before being relied on: an analysis of an exported CSV with three findings, one draft document (a proposal or a policy), and one sourced research note or slide outline. Each file ends with a note of what was verified by hand.

Video

The video for this lesson is not recorded yet.

Prompts used in this lesson

Part 3 — Numbers: analysing an exported CSV

Purpose: I am deciding where to put my effort next quarter: which lead
sources are worth more attention and which are not. Help me see what my own
data says, and how sure I can be.

Context: data/leads-export.csv is an export of my leads for the last three
months. Contact details have been removed. Columns include created_at,
status, utm_source, utm_campaign and landing_page. Statuses are new,
contacted, qualified, won, lost.

Do this:
1. Before any analysis, describe the file: number of rows, date range, each
   column, how many empty values, anything odd. Stop and show me.
2. Then write a small script to count, run it, and save it in the folder so
   I can run it again next quarter. Do not calculate in your head.
3. Give me at most three findings. For each: the numbers as counts ("4 of
   19", never a bare percentage), what it suggests, and a label: solid,
   tentative or too early to say. If moving one or two rows would change
   the conclusion, it is not solid.
4. For each finding, tell me how to recount it by hand: which column to
   filter and the number I should get.
5. List what this file cannot tell me.

Done looks like: a file analysis.md with the description, the three
findings, the recount instructions and the limits.

Constraints: never invent or estimate a number; if a figure is not in the
file, say so. No advice beyond what the data supports.

Part 4 — Documents: a proposal or a policy

Purpose: a one-page proposal I can send to the manager of a small office who
asked whether we clean offices. I want it to get a reply, not to win a prize.

Context: facts-office-cleaning.md contains everything that is true about
this offer. The reader is busy, not technical, and will read it on a phone.

Done looks like: proposal.md with: what we will do and how often, what is
included and what is not, how to start, and one clear next step. One page.
Plain words, our own voice (short sentences, no sales language).

Constraints: use only facts from the file. Where a fact a reader would
expect is missing (price, insurance, notice period, start date), do not
invent it: write [OWNER TO SUPPLY: what is needed]. No promises, guarantees
or figures that are not in the file. After the draft, list every sentence
that is not directly supported by the file, even if it seems harmless.

Part 5 — Research, with the sources opened

Purpose: I want to understand what a small business in [my country] must do
before sending marketing emails to past customers, well enough to prepare
questions for an adviser. I am not asking for legal advice.

Context: we are a home-cleaning company with a list of past customers'
email addresses collected through our own quote form.

Do this: search the web and read the pages; do not answer from memory.
Prefer official sources (the regulator, the government) over blogs and
vendors. For every claim give the link, the name of the page, its date if
shown, and a short exact quote that supports the claim.

Done looks like: research-email-rules.md with: (1) what the official sources
say, point by point, each with link and quote; (2) where sources disagree or
are unclear; (3) what you could not find or could not open; (4) five
questions to ask an adviser.

Constraints: if you cannot find an official source for a point, say so
rather than filling the gap. Keep "what the source says" separate from
"what I infer". Name the country and date each source applies to.

Part 6 — Slides: an outline from something you trust

Purpose: a ten-minute talk to the office manager and her director, to get a
yes to a one-month trial.

Context: proposal.md (checked by me) and nothing else. The audience has not
read it.

Done looks like: slides-outline.md with six slides at most. For each: one
message as a full sentence in the title, up to three supporting points, what
to show (a photo, a number, nothing), and what I say in two sentences. The
last slide is the one request I am making.

Constraints: nothing that is not in proposal.md. Keep every
[OWNER TO SUPPLY] placeholder visible. No slide without a message.

Do along

Work on your own business. Pause the video where a step says so. Choose research or slides for the third product; the other is homework.

1. Create a folder outside your website project, with a subfolder data. Start Claude Code in it.
2. Pause after Part 2. Export one CSV from your business. Delete contact columns and anything sensitive before saving it into data.
3. Pause after Part 3. Run the analysis prompt with your own decision in the Purpose and your own columns in the Context. Check the row count. Recount one finding by hand. Add your "Checked by hand" line to analysis.md.
4. Pause after Part 4. Write half a page of your own facts for one document you actually need. Run the document prompt. Fill the placeholders with true answers or leave them marked. Add a line at the end: which unsupported sentences you corrected.
5. Pause after Part 5 or Part 6. Run the research prompt with your own question and country and open the two sources that matter most; or run the slides prompt on your checked document and read the titles in a row. Add your check line.
6. Keep the three prompts in a file prompts.md in the same folder.

Check your work

1. Open the folder. Expected: three files for your own business, for example analysis.md, proposal.md and research-<topic>.md or slides-outline.md.
2. Open the analysis. Expected: every rate has its counts, each finding has a label, and the last line names the number you recounted and the date.
3. Open the document. Expected: no price, date or promise that you did not supply; remaining gaps are visible as [OWNER TO SUPPLY: ...].
4. Open the research note or the outline. Expected: for research, every claim has a link and a quote, and two are marked as opened by you; for slides, six titles that make the argument on their own.
5. Look in data. Expected: no names, emails or phone numbers in any file there.

Common problems

• The analysis gives percentages without counts, or findings with no doubt in them → the constraints were shortened → "Rewrite the findings: counts beside every rate, and label each one solid, tentative or too early, with the reason."
• Your recount does not match → the file was read with the wrong separator, date format or decimal mark, or rows were dropped → "My count for this finding is X. Show me the rows you counted and how you read the file." Do not use any finding until it is resolved.
• The document contains a fact you never gave → the facts file is thin and Claude filled the space → add the true fact to the facts file, or tell Claude to replace the sentence with a placeholder; then ask again for the unsupported list.
• A research link does not open, or the quote is not on the page → treat that point as unconfirmed → "This link does not support the claim. Re-read the page and correct the note, or move the point to 'could not find'."
• The research is about another country → the country was missing from the brief → state it in the Purpose and ask for sources that name it.

Homework

About 30 minutes, on your own, after the lesson. Nothing depends on it.

1. The fourth kind. Produce the work product you skipped: the research note or the slide outline, with its check. Deliverable: the file. Done when: its last line says what you verified.
2. Your verification checklist. Write checklist.md: for numbers, documents, research and slides, the one or two checks you will always do, in your own words, plus your personal "never paste" list. Deliverable: the file. Done when: it fits on half a page and you would actually follow it on a busy day.
3. Run it again. Next month, or now with a different export, rerun the saved analysis script and prompt on new data. Deliverable: a second analysis file. Done when: it was produced with no changes to the prompt, or you have noted in prompts.md what had to change and why.

Nothing here touches your website project, so there is nothing to commit.