File: 'Q1-2026-revenue-by-segment.csv'\nRows: 18 · Columns: month, segment, mrr, customer_count, churn_rate\n\nFull data:\n2025-11, SMB, $42,300, 312, 4.1%\n2025-11, Mid-Market, $128,500, 87, 2.2%\n2025-11, Enterprise, $89,000, 12, 1.0%\n2025-12, SMB, $40,800, 298, 4.8%\n2025-12, Mid-Market, $134,200, 91, 2.0%\n2025-12, Enterprise, $94,500, 13, 0.8%\n2026-01, SMB, $38,100, 281, 5.4%\n2026-01, Mid-Market, $141,800, 95, 1.9%\n2026-01, Enterprise, $98,200, 14, 0.7%\n2026-02, SMB, $35,200, 263, 6.1%\n2026-02, Mid-Market, $148,900, 99, 1.7%\n2026-02, Enterprise, $108,000, 16, 0.5%\n2026-03, SMB, $33,400, 245, 6.8%\n2026-03, Mid-Market, $156,200, 104, 1.5%\n2026-03, Enterprise, $115,500, 18, 0.4%\n2026-04, SMB, $31,200, 228, 7.2%\n2026-04, Mid-Market, $164,400, 109, 1.4%\n2026-04, Enterprise, $122,800, 20, 0.3%\n\nAuto-detected: SMB MRR is declining month-over-month while Mid-Market and Enterprise are growing. Total MRR is growing despite SMB decline. SMB churn rate has nearly doubled over the period.

Summary  
This chapter covers prompt engineering for structured data interpretation, showing how to design system prompts that force a language model to produce specific insights—top findings, anomalies, and concrete recommendations—instead of generic restatements.

General domain of usage  
Business spreadsheet analysis (e.g., revenue and churn reporting)

You can throw a CSV at Claude and get back **"MRR is up overall."** Useless. You can also get a sharp three-line read that names the segment in trouble, the month it started, and what to do about it. Same data – different system prompt.

## The Shift from Notes to Numbers

A notes-reading tool returns prose. A spreadsheet tool returns rows – but that changes the **failure mode** rather than removing it.

| Input Type | Typical Failure |
|---|---|
| Prose (notes, docs) | **Missing** things |
| Numbers (spreadsheets) | **Restating** things |

Ask **"summarize this spreadsheet"** and you'll get: **"Revenue grew, churn varied, customers fluctuated."** All true, all worthless. The model has to **interpret** the numbers, not echo them.



## Forcing Interpretation

The system prompt in this lab forces three moves:

1. **Top 3 findings** – with specific numbers cited (months, segments, percentages);
2. **One notable anomaly** – something most people would miss on a first read;
3. **One recommendation** – a concrete action, not a platitude.

**Takeaway:** structured data needs structured demands. Otherwise the model defaults to politely describing the spreadsheet back to you.

## Task

>Analyse this revenue spreadsheet. What are the top trends I should bring to next week's leadership meeting?

Then push it:

>Drill into SMB. What's driving the churn jump?

>What would happen to total MRR if SMB churn keeps climbing at this rate?

>Which segment should we double down on for Q3?

Each follow-up uses the same loaded data. Same numbers, different lens.

A hands-on, beginner-friendly course designed for non-technical business users to understand and leverage AI tools and concepts in their daily work. The course covers foundational AI concepts, effective communication with AI, practical use of AI-powered tools, and building simple AI assistants.

Introduction to artificial intelligence, large language models, and essential AI vocabulary for business users.

Learn how to communicate effectively with AI systems through prompt engineering and tone selection.

Explore how AI can be integrated with business tools for research, note-taking, and data analysis.

Learn the basics of agent personas and how to build and tune simple AI assistants for work.