Provide Examples to Guide the Output

Examples can demonstrate a distinction, structure, tone, or level of detail that would be cumbersome to describe. This is often called few-shot prompting. It guides the current response; it does not retrain the model.¹

When examples help

Use examples when:

• categories depend on course-specific definitions;
• a response must follow a particular format;
• acceptable and unacceptable outputs differ subtly; or
• the required tone is easier to show than describe.

An example is unnecessary when the task and output are already unambiguous. One representative example may be enough; complex boundaries may need several. Quality and coverage matter more than a fixed number.

Build a useful example set

• Keep the input and output labels consistent.
• Use examples that resemble the real task.
• Cover important boundaries, not only easy cases.
• Check every example: the model may reproduce its errors and biases.
• Separate examples clearly from the new input.
• State which features should transfer and which are incidental.

Example: classifying feedback comments

Suppose an educator wants to distinguish feedback grounded in observable work from unsupported judgments.

Classify each comment as EVIDENCE-LINKED or UNSUPPORTED. Then give one sentence
explaining the classification. Follow the distinctions in the examples.

<example>
Comment: "Your second paragraph states the conclusion but does not cite or
analyse evidence for it."
Classification: EVIDENCE-LINKED
Reason: The comment points to a location and an observable omission.
</example>

<example>
Comment: "You did not put enough effort into this assignment."
Classification: UNSUPPORTED
Reason: The comment infers motivation rather than describing evidence in the
work.
</example>

<new_input>
Comment: "The graph labels revenue in euros, but the explanation discusses the
values as percentages."
</new_input>

Return exactly:
Classification: [label]
Reason: [one sentence]

The examples teach both the labels and the output pattern. The final format instruction prevents incidental features, such as quotation marks, from being mistaken for the important distinction.

Specify format and style directly

Examples work best alongside explicit requirements. State the intended length, structure, language, tone, and audience. Do not rely on the model to infer which features of an example matter.

References & Footnotes