Relax. We're not going to explain this chapter like a textbook — we'll just talk it through so it's easy to follow.
Let's think about it this way for a second
If your question is too broad, you can't tell pass from fail. 'Is the AI good?' isn't a question. 'Per the handbook, is phone use allowed?' is a question. You need to be able to write down the correct behavior in advance. If you can't, what you have is an exploratory chat, not yet an eval. Don't ask the server 'was the food good?' Ask 'was the fish dish not spicy?'
Let's connect it to everyday life
Good — 'Per the Week 1 notes, why does a plant need sunlight? Two sentences.' Bad — 'Explain photosynthesis fully, with an example, a question, and a poem.' The second one mixes three tasks together — you won't know which part failed. In your first pack, one question = one behavior. Same idea as the Prompt Engineering lesson: one task at a time.
Let's try it together
Too wide: Is my study bot good?
Better: From week-1 notes, why do plants need sun?
Expect: 2 sentences + no exam jargonYou'll be able to pick a short, focused question for an eval.5-Minute Try-It
Write 2 overly broad questions. Split each one into 2 short, focused questions.
A Quick Word of Caution
Don't eval a question you don't know the answer to yourself. That's just you guessing, then grading it as if you knew.