Show, Don't Tell: Prompting with Examples

Ask Dev to describe how his review replies sound and he'll shrug: "Friendly, but real? Not corporate?" That description is true, useless, and the reason his first attempts at automating replies went nowhere.

Dev's Westgate pizza location sits at a 3.9-star average on Google, and he answers every review himself. Over four years his replies have developed a genuine voice — he thanks people by name, owns mistakes without weasel words, offers a concrete fix, and signs off like a person instead of a brand. Customers mention the replies in other reviews. That's marketing he can't buy.

When he asked Claude to write replies "in a friendly but genuine tone," he got mall-kiosk cheerfulness. He tried more adjectives: authentic, warm, down-to-earth. Same syrup, different label. The problem wasn't effort. It's that voice lives in details no adjective can carry — sentence rhythm, the sign-off, the words you'd never use. You can't describe your voice. You can only demonstrate it.

Paste the evidence

So Dev stopped describing and started showing. He dug up two past replies he was proud of, labeled them, and put them in the prompt ahead of the ask. The whole thing took three minutes, most of it spent scrolling his own review history:

The draft that came back opened with the customer's situation instead of a canned apology, owned the miss in plain words, and ended with a specific make-good — recognizably Dev, on the first try. The technique has a name in the trade: few-shot prompting, the "shots" being examples. You're welcome to forget the term the moment you close this tab; the showing is what matters.

Three rules for choosing examples

Two or three beat ten. Each example teaches the model something, but the returns fall off fast. Past the third, you're mostly burying the actual task under a pile of homework — and crowding the desk with material it no longer needs.

Show the range. Notice Dev picked one happy reply and one angry-customer reply. If both examples had been sunny thank-yous, the model would have learned his happy voice and cheerfully applied it to a furious customer. Pick examples from different situations so it learns the voice, not one mood.

Label everything. "Example 1 — my reply to a 5-star review" costs six words and removes all ambiguity about where one thing ends and the next begins. Unlabeled walls of pasted text force the model to guess at the boundaries, and you already know how guessing goes.

Where this fits in your toolkit

If you're mapping this onto The Five-Part Prompt, examples slot into the Context section — they're just context so concentrated that two of them outwork a hundred adjectives. Keep the rest of the structure: the task, the format, the constraints all still apply.

And the trick travels well beyond reviews. Paste two job postings that attracted good applicants before asking for a third. Paste your three best Instagram captions before requesting next week's batch. Paste the proposal that closed your biggest client before drafting a new one. Anywhere your business has a track record of writing that worked, you're holding training material.

One honest caveat: examples teach voice and shape, not facts. The model will mimic how you write while still needing the details of what you're writing about. Dev's examples taught it his sign-off, not tonight's delivery delay — he still had to paste the new review itself. Voice comes from the past, details come from today, and a good prompt carries both.

Try it now

Spend two minutes finding two pieces of writing that genuinely sound like you — sent emails, review replies, social posts, even texts to customers. Pick two from different situations, label each with one line about what it was, and ask for a new piece in the same voice. Then run the same request with adjectives only ("friendly, professional, warm") and put the results side by side. One of them will sound like you. The other will sound like everyone.