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
Prompt is the instruction. Context is what's laid out on the table. RAG goes and looks things up in the book. An agent picks up the tools. An eval just asks the question back: 'did it actually work?' It's like a chef cooking a dish and having three people taste it — one person saying 'tastes good' doesn't mean you can open the restaurant. An eval is a card made of a question + an expectation + a note. Don't chase dashboards, leaderboards, or LLM-as-judge setups yet. Start with three cards.
Let's connect it to everyday life
Ask a study bot 'can I use my phone?' The eval card requires the answer to say, per the handbook section, that it's not allowed — and to cite the source. Writing a pretty sentence alone doesn't make it pass. If you ask it a joke instead, it can answer like any chatbot — that's a different card entirely. This course doesn't score 'sounds nice.' It only checks 'does this behavior show up.'
Let's try it together
Eval card
Ask: ဖုန်းသုံးလို့ရလား
Expect: No, quote handbook / phones
Pass if: says no + shows the line
Fail if: guesses “modern schools allow it”You'll be able to explain an eval as a short test card.5-Minute Try-It
Write 3 draft eval cards for one of your own AI tasks. For each one, include the question and what counts as a pass.
A Quick Word of Caution
If you call it done just because it looked good one time, that's not an eval. That's a vibe check.