Show, don't tell

Some things are hard to describe and easy to show. Tone is one. “Friendly but not cheesy” means something different to everyone. One good example settles it.

Putting examples in your prompt is called few-shot or multishot prompting. Anthropic calls it one of the most reliable ways to steer the format, tone and structure of Claude's replies.

Say your app shows raw errors like ECONNRESET and you want them in plain English, in your house style. You could describe the style. Or you could show it:

Notice three things Anthropic recommends. The examples are relevant: real errors from the same app. They're diverse: a server problem, a sign-in problem and a user mistake. And they're wrapped in tags, so Claude can tell the examples apart from your instructions. Three to five is a good number.

There's a catch. Claude copies examples closely, including the things you didn't mean to teach. If every example starts with “Oops!”, every reply will too.

On the left, all three examples opened with “Oops!” and ended with “Please try again!”, so the reply did too. On the right, the examples varied their openings and each said what to do next, so Claude picked up the real pattern: plain words, then a next step.

If you're stuck, ask Claude to check your examples for variety, or to write a few more like them.

Your turn. The draft has one example. Add a second example for a sign-in or password error, keep both in <example> tags, and keep the new error at the end. Run it, then try changing your examples' style and watch the reply follow.