Skip to content

AR Digital / Daily AI

Give the AI examples: the habit that sets the tone and format of its answers

Describing the tone you want is not enough: show the AI three to five anonymised examples. The habit, a step-by-step case and the trap to avoid.

Radu UleaFounder of AR Digital

7 min read

  • Written with AI
  • 2 sources cited
Give the AI examples: the habit that sets the tone and format of its answers

“Professional but warm”, “short and clear”: instructions like these produce generic texts that everyone then rewrites. Showing the AI a few pieces of work you consider successful, three to five according to Anthropic, is one of the most reliable ways to shape its answers, and Google considers prompts without examples likely to be less effective. Here is how to do it without exposing client data.

Daily AI. You ask the AI for a professional but warm tone, and you get a text that everyone rewrites? The problem is the description. Today’s habit: show rather than describe. Paste into your request three to five examples of answers you consider successful, anonymised, varied and clearly separated from the instruction. Anthropic presents examples as one of the most reliable ways to set tone and format, and Google recommends always including them. Watch out for the trap: with a single example, the tool may copy its length, its phrases, sometimes its figures. Vary the situations, and say clearly what must not be reused. Kept in a shared space, these prompts help the whole team speak with one voice. The detail is just below.

In short: instead of describing the tone or format you expect, paste into your request three to five real, anonymised and varied examples of what you consider successful, clearly separated from the instruction. Anthropic presents them as one of the most reliable ways to shape an answer, and Google recommends always including them.

The problem at the office

In many teams, the typical request looks like this: “Write a reply to this client, professional but warm tone, not too long.” The result is fine, but it does not sound like your company. The phrases are stock phrases, the structure changes from one time to the next, and everyone then spends several minutes reworking the text so that it sounds “like us”.

The reason is simple: adjectives like “warm” or “clear” do not mean the same thing to you, to your colleague and to the tool. You have a precise idea of the right answer in your head, but you describe it instead of showing it. The AI then fills the gap with what it has seen most often, which means average text.

Adding instructions does not solve the problem. A list of ten style rules often ends up contradicting itself, and nobody on the team writes it the same way.

The habit: show rather than describe

The habit is to place in your request a few examples of what you expect: replies already sent, minutes, product sheets that you consider successful. Specialists call it “multishot prompting”, in other words an instruction accompanied by a few examples.

The vendors document it in black and white. In its guide to prompting best practices, Anthropic writes that examples are one of the most reliable ways to steer the format, tone and structure of Claude’s answers, and recommends providing three to five. The guide specifies that they should be relevant (close to your real case), varied (so the tool does not pick up a detail by mistake) and clearly separated from the instructions.

Google says the same for Gemini: its prompt-writing guide recommends always including examples, considers prompts without them likely to be less effective, and notes that you can even remove instructions when the examples are enough to show the task.

So the principle is not specific to one tool. It is documented for Claude and for Gemini, and should hold for other assistants: try it on your own requests.

A step-by-step example: replying to quote requests

Take a services SME that receives quote requests by email every week. Here is the approach, a theoretical example to adapt to your business.

  1. Choose three replies you sent and are happy with, in different situations: a client in a hurry, an incomplete request, a polite refusal.
  2. Anonymise each reply. Remove names, addresses, amounts and any detail that would let someone recognise a client. Also check that you are working in a suitable work account, as we explained in our edition on choosing between a personal and a work account.
  3. Separate the examples clearly from the instruction. Introduce each example with a heading (Example 1, Example 2, Example 3). For Claude, Anthropic recommends wrapping them in <example> tags. Whatever separator you choose, use the same one for every example.
  4. Write a short instruction saying what to reuse and what not to reuse.
  5. Compare the result with your examples. If it does not fit, change an example rather than adding adjectives.

Sample prompt: “Here are three replies we sent to quote requests. They show our tone, our structure and the way we ask questions. [Example 1] [Example 2] [Example 3]. Write the reply to the following request in the same spirit, without reusing their sentences or their figures. If any information is missing to draw up the quote, ask the question instead of making it up. [Client request]”

Once the prompt works, save it and its examples in the team’s shared space, next to your context sheet. The whole team then starts from the same base, and the replies resemble one another without everyone having to reinvent the request.

The trap to avoid: the AI that copies

The classic trap is excessive fidelity. With a single example, the tool may reuse the same length, the same opening phrase, sometimes a figure or a name that has no place in the new reply. The risk also runs the other way: Google points out in its guide that with too many examples, the model may stick too closely to them. Anthropic, for its part, insists on the variety of examples, so that the tool does not pick up a pattern you did not intend.

Three habits limit this risk:

  • vary the situations and the lengths, so that the tool picks up the tone and not a single template;
  • keep the same layout from one example to the next, as Google recommends, otherwise the answer may not take the format you want;
  • state explicitly what must not be reused: figures, names, terms specific to one client.

One last point, often forgotten: a bad example is imitated as faithfully as a good one. A reply sent on a rush day, with a mistake or a promise you would no longer keep, has no place in the selection. And every factual claim produced still needs checking, as with any AI answer (see our advice on the exact quotation).

Going further with your team

On its own, the habit improves quality. At team level, it changes more: the examples become a company asset. A trained team builds a small library of validated examples for its recurring tasks (replies to quotes, follow-ups, minutes, product sheets), names one person to keep it up to date, and removes an example as soon as an offer or a rule changes.

This is also where the difference between individual use and controlled use is made: knowing which documents can serve as examples, how to anonymise them, and how to check that a new example really improves the answers. These habits are best learned on the team’s real documents rather than in a general tutorial.

Our take

Describing the style you expect is a very common and not very effective way of working with AI. The vendors themselves recommend showing rather than describing. For an SME, the real gain is not an occasionally better-phrased reply, it is a consistent voice across all exchanges, built from what the team already does well.

Frequently asked questions

How many examples should I give the AI?

Anthropic recommends three to five examples for Claude. A single example risks being copied, and Google warns that too many can also lead the model to stick to its examples. Three varied examples are a good starting point.

Does this habit also work with Gemini or Copilot?

The principle does not depend on the tool. Google recommends including examples in prompts addressed to Gemini, Anthropic does the same for Claude, and the method holds in principle for any assistant that receives a written instruction.

Can client emails be used as examples?

Yes, provided you anonymise them: remove names, contact details, amounts and any detail that would let someone recognise the client, and work in a work account whose data-use terms suit you.

Where should I keep the validated examples?

In the shared space where the team keeps its prompts, for example a shared project in your AI tool or a common document, next to your context sheet. Name one person to update them when an offer or a rule changes.

ShareLinkedInBy email

Radu UleaFounder of AR Digital

Founder of AR Digital, a web and digital marketing agency in Luxembourg. He helps SME owners with their website, visibility, advertising and move to artificial intelligence, and writes the Daily AI edition every working day. LinkedIn

Every working dayAR Digital / Daily AI

Get Daily AI every working day.

One piece of AI news, what it changes for your business and what to do, in your inbox in the morning. Free, no commitment.

Free, one email per working day, unsubscribe in one click. The email edition is currently sent in French. An email asks you to confirm your address.

AR Digital / Contact

Let’s talk about your project.

Thirty minutes, by video call or at the agency, to understand what you need and tell you frankly what we would do, at what price and with what funding. No commitment.

Radu Ulea, founder of AR Digital

Radu Ulea

Founder, AR Digital

“We tell you what works, what doesn’t, and what it costs. Then we do it.”

hello@ardigital.eu