What it is
Three short parts: a good example, a bad example, and a sentence that names the difference. The bad one is usually the answer you keep getting, which is why it is worth showing. The sentence matters as much as the samples, because it tells the model which of the many differences between two pieces of text is the one you care about.
This builds on showing one example. On the ten checks, the good sample is what passes Reference; the counterexample and the sentence sharpen it.
When it helps
It helps when the line between good and weak is subtle. Specific feedback against generic feedback, confident against overclaiming, plain against condescending, warm against gushing. Each is hard to pin down in a description and obvious in a pair.
It helps when you have already seen the failure. If every draft of a customer reply opens with three apologies, paste one as the bad example and you have shown the model exactly what to move away from.
And it helps at boundaries. When sorting or grading, a pair of similar cases on either side of a line (a billing question that belongs in Billing, and one that is really a login problem) teaches the boundary faster than a definition.
When it does not
The bad sample is still text in the prompt, and models copy what they see. Keep it short, label it plainly as the bad example, and always pair it with a good one and the reason. A bad example on its own gives the model something to imitate and no target to aim at.
The vendors do not ask for this. The prompting guides from Anthropic, OpenAI and Google describe examples as samples of the output you want; as read on October 29, 2026, none of the three recommends adding bad ones. That does not mean it fails, but it does mean you should check it: run the prompt with and without the counterexample on a few real inputs in Prompt Tester and keep it only if it helps.
Write the bad sample yourself. A real colleague’s weak review or a real customer’s angry email is not yours to paste as a cautionary example, and a made-up one works as well.
If you cannot write a good sample, a counterexample will not fill the gap. Start by describing a good result in words, and add the pair once you have seen an answer you like.
Before and after
Two requests where the after version adds a good and a bad sample with the difference named. Each score comes from the Prompt Grader’s free ten-check estimate, written for any chat model.
Write three sentences of praise for each person on my team, from my notes below, for their year-end reviews. """ [your notes on each person] """
Write three sentences of praise for each person on my team, from my notes below, for their year-end reviews. Good example: "Your handover notes for the March release let the night shift fix two outages on their own." Bad example: "Great job on the release this year. You are a real team player." The good one names what the person did and what it changed. The bad one could be about anyone. Write every line like the good example. """ [your notes on each person] """
Asked for praise, a model can reach for the phrases praise usually contains, which is how reviews end up calling everyone a team player. The pair, and the sentence under it, push each line toward a specific action and its effect, drawn from your notes.
Write a reply apologizing to the customer below for the double charge on their account. """ [the customer's email] """
Write a reply apologizing to the customer below for the double charge on their account. Good example: "I'm sorry you were charged twice on May 3. I've refunded the second charge today, and it will show on your card within five business days." Bad example: "We are so incredibly sorry for this terrible experience and completely understand your frustration." The good one says what happened and what we did about it. The bad one apologizes at length and fixes nothing. Match the good example: one apology, then the facts. """ [the customer's email] """
“Apologize” on its own invites an answer that is all apology. With the pair, the reply apologizes once and spends the rest on what happened and when the money comes back. This pair also gains a point by accident: the estimate reads “five business days” in the good sample as a size.
The line to add
Good example: "{{a_sample_that_hits_the_mark}}"
Bad example: "{{a_sample_that_misses}}"
The good one {{what_it_does}}. The bad one {{what_it_gets_wrong}}. Write like the good example.
Give the two samples the same shape and length, so the only difference is the one you are pointing at. If the good one is a sentence and the bad one is a paragraph, the model may learn that shorter is better, which was not the lesson.
In ChatGPT, Claude, Gemini and Copilot
The pattern works the same way in all four, and the vendors’ advice on examples applies to it. Anthropic’s guide for Claude recommends wrapping examples in tags so Claude can tell them apart from instructions, and using consistent, descriptive tag names, so for Claude you might put the pair in <good_example> and <bad_example> tags. Google’s guide for Gemini asks for the same structure and formatting across examples, which is the reason to give the two samples the same shape.
OpenAI’s prompt engineering guide, for the models behind ChatGPT, describes examples as inputs paired with the desired outputs, so lead with the good sample and treat the bad one as an addition. Microsoft’s guide to Copilot prompts counts samples as part of the source, and a labeled pair fits there.
Key takeaways
- A good sample, a bad sample and one sentence naming the difference.
- Use it for lines that are easier to show than to describe, such as specific against generic.
- Make the bad sample the answer you keep getting, written by you.
- Never show a bad example on its own, and keep it short and labeled.
- Vendors recommend examples of what you want, so test the counterexample before you keep it.
Questions
Should I show AI examples of what not to do?
Sometimes, and never alone. A bad sample next to a good one, with a sentence naming the difference, can teach a line that is hard to describe, such as specific praise against generic praise. On its own, a bad example gives the model text to imitate and no target. Test the prompt with and without it on a few real inputs.
What is a counterexample in a prompt?
A sample of the output you do not want, placed next to one you do, with the difference named. It is most useful when the bad version is the one the model keeps producing, because it shows exactly what to move away from. Give both samples the same shape so the difference is the only thing that stands out.
How do I get AI feedback that is specific instead of generic?
Show the difference. Give one line of feedback that names what the person did and what it changed, one generic line such as “great job this quarter”, and say every line must be like the first. Then give it your own notes, because specific feedback needs specific facts. The guide to performance reviews with AI covers the rest.