# Test a reusable prompt with a counterexample

> Published 2026-11-06 · https://www.promptzone.com/aiai_creative/test-a-reusable-prompt-with-a-counterexample-3oad


![Image description](https://promptzone-community.s3.amazonaws.com/uploads/articles/iooih84rhryp19i3mdhl.jpg)

A prompt can look dependable when you try it on a sample that already resembles the answer you want. A better trial includes one case where a tempting inference would be wrong. That case helps you decide whether to save the prompt as a reusable tool or keep it as a one-off instruction.

## Define the boundary

Imagine a prompt that turns workshop notes into a short handoff. It should identify confirmed decisions and open questions. The note “Could we meet on Friday?” is an open question. The note “Friday is confirmed” belongs under decisions. Give the assistant both types of sentence before evaluating the result.

## Keep an expected answer

Write a small answer key in your own words. Include the item that should remain unresolved. If the prompt asks the tool to organize the notes, avoid rewarding a result simply because it fills every section. Empty space can reflect missing evidence.

## Compare runs at the same review point

Use AI workflows as a reference for separating the draft from the review. Give each candidate the same input and record which sentences were moved, rewritten, or added. Then look for the single counterexample. Can you identify it without reading the entire conversation again?

## Make the smallest repair

If the draft upgrades the question to a decision, add a precise rule for language that asks about a possibility. Run the original sample again alongside the counterexample. A repair should preserve the good behavior you already observed. If the prompt now labels every future date as uncertain, it has traded one error for another.

## Look for failures that a tidy output hides

Suppose your test note says that an organizer can reserve either of two rooms, but the booking request has yet to be sent. A summary that places one room under “Location” may sound helpful while hiding the unresolved choice. Compare the statement with the actual note, then ask whether someone could act on it without checking. This is a practical way to distinguish a harmless wording difference from a consequential error.

Try another case where the notes themselves contradict each other. An early entry names the library and a later entry mentions a possible switch to the hall. Without evidence that the change was approved, a reliable result should flag the ambiguity. Avoid using the order of the notes as an automatic substitute for authority. Later text may report a question, while earlier text records a decision.

## Check what happens when the sample changes shape

A prompt written against neatly labeled notes may depend on those labels more than you expect. Try the same confirmed decision in two forms. In one, put it in a bullet under a heading. In the other, include it in a plain paragraph next to an unrelated question. If the result changes its interpretation, inspect which part of the prompt relies on formatting instead of meaning.

Keep the comparison small enough to review by hand. Three examples with explicit expected results will often tell you more than a large collection of outputs you never have time to read closely. Add a new case only when it tests a different failure mode or records a problem that actually occurred.

## Make the review easy to repeat

Save each case with its original input, the expected classification, and the reason for that classification. Record whether the output preserved the distinction and whether the reviewer had to look outside the sample. Use plain labels such as confirmed, open, and insufficient information.

A useful record also tells a colleague when to stop. If an output invents a commitment, mark it for revision instead of passing it on to the next stage. If the prompt preserves uncertainty but uses a slightly different phrasing, that may be acceptable. Write down the acceptance rule so future editors can tell these cases apart.

## Record the limit of the trial

Keep the prompt, the two examples, the accepted output, and one line about what remains untested. The next editor can extend the sample with their own borderline case. A prompt becomes easier to maintain when a reader can see what its instructions protect and when those instructions need another example.