Prompt Engineering in Depth

What this pack asks

The free questions in this pack. Answer choices and explanations appear as you play.

• A logistics company runs an unattended nightly batch job where Claude extracts all rows matching a filter from roughly 400 supplier data files. The team validated the prompt against ten hand-picked, well-formed files and it succeeded on all of them, but in production the job intermittently emits invented rows or free-text apologies that break the downstream loader. Constraints: no analyst is available to review outputs at night, the downstream loader cannot be changed, and the prompt must produce a predictable output for every file it receives. Which approach BEST addresses the failures?

  Answer this one →

• A SaaS company embeds Claude as the support chat widget inside its own invoicing product. The current system prompt opens with "You are an experienced call-centre agent with 15 years of phone support experience," and answers come back with phone-call conventions, generic advice, and refund guidance that contradicts the company's actual policy. Constraints: the team cannot fine-tune and must stay prompt-only, answers must reflect this specific product and its documented policies, and the assistant must not claim to be a human. Which approach BEST meets these requirements?

  Answer this one →

• A platform team maintains a shared prompt library used by 60 engineers. Roughly half of it consists of workarounds added for older models — explicit "think step by step" directives on arithmetic, output-shape nudges, and similar patches. After each model upgrade, evaluation runs show these patches now add little or nothing, while genuinely new model capabilities go unexploited. Constraints: the team must adopt each new model release within a quarter, cannot fine-tune, and has the budget to actively maintain only a small core of the library. Which approach BEST meets these requirements?

  Answer this one →