1. An example is an instruction in disguise

Course 1 said it in passing; this section makes it a tool. An example in a prompt is not an illustration. It is a pattern the model will continue, usually more faithfully than it follows a description. 'Reply like this:' plus one real support reply you loved outperforms three sentences of adjectives about warmth and brevity.

The standard shape is a labeled input → output pair: 'Customer message: … / Our reply: …'. One pair teaches the transformation; the labels teach where the pattern starts and stops. The technique has a name, few-shot prompting but the habit matters more than the name: when you can show it, show it.

2. One to three good ones beat ten mediocre ones

Examples are potent, and that cuts both ways: the model imitates everything about them, including the flaws you didn't notice. A mediocre example actively teaches mediocrity. Quality beats quantity, one to three strong examples usually cover a task, and past that you are adding weight, not signal.

Choose examples that show the hard part. If your real inputs range from angry customers to billing confusion, don't paste three easy thank-you notes, show the edge you most need handled well. Good examples vary along the same dimension your real inputs vary.

3. The example IS the format spec

The model treats your example's format as part of the pattern. If the example has a subject line, outputs grow subject lines. If it runs four sentences, expect four-ish sentences. So build the example in exactly the shape you want back, same headers, same length class, same markup.

The trap is instructions that argue with examples. You write 'keep replies under 100 words' and paste a 300-word example; now the rule and the pattern disagree, and outputs drift long or wobble between the two. When you spot that conflict, fix the example. It usually speaks louder than the rule.

4. Label your counter-examples

Sometimes the fastest way to define good is to show the bad thing you keep getting. That is safe only with labels: 'Not like this: "I sincerely apologize for any inconvenience…", why: it opens on apology instead of the fix.' Unlabeled, a bad example is just another pattern, and the model may continue it.

The strongest setup bounds the target from both sides: one 'Like this' example and one labeled 'Not like this' with the reason it fails. The reason matters. It tells the model which feature of the bad example to avoid, instead of leaving it to guess.