How the cases work
Every lesson ends with an Apply it exercise set in one of these companies. The same four companies appear across the whole guide, so the design for each one builds up as you move through the domains: a feasibility brief in Domain 1, an evaluation plan in Domain 4, a control register in Domain 5, and so on.
The companies, people and numbers are invented for study. Any resemblance to a real organisation is coincidental.
| Case | Industry | The problem in one line |
|---|---|---|
| Larkfield Mutual | Insurance | Claims intake is slow and error-prone, and the acknowledgement SLA is missed |
| Corvane Home | Online retail | A high-volume support centre answers the same questions all day, with live order data |
| St. Aldric Hospital | Healthcare | Physicians spend too long writing discharge summaries, and coders keep sending them back |
| Quillstone Software | B2B software | A 420-engineer company wants to roll out Claude Code safely and use it for large code changes |
Assumptions used in every case
Exercises that ask for a cost estimate use these assumed prices, so the arithmetic is the same for everyone. They are round numbers for practice, not current Anthropic pricing; always check the pricing page for real work.
| Model tier | Input, per million tokens | Output, per million tokens |
|---|---|---|
| Tier S (small, fastest) | $1 | $5 |
| Tier M (balanced) | $3 | $15 |
| Tier L (most capable) | $15 | $75 |
- Prompt caching: a cache read costs 10% of the normal input price; writing to the cache costs 125% of it. Treat a stable prefix as written once and read on every other request.
- Batch processing: 50% off both input and output, for work that can wait up to a day.
- Rough token sizes: one page of business text is about 600 tokens; a one-line answer is about 30.