Prompt Example Generator: What Makes an Example Actually Useful

Adding more examples to a prompt isn't automatically better. Three examples that all look alike teach the model less than two examples that cover genuinely different cases — the value of an example comes from what pattern it demonstrates, not just its presence in the list.

Cover the Range, Not Just the Middle

Three sentiment examples that are all clearly positive don't show the model how to handle an ambiguous or negative case. A useful set typically spans the range of what's actually expected:

Input: The battery life is incredible, lasts two full days.
Output: Positive

Input: Arrived broken and customer service never replied.
Output: Negative

Input: It does what it says. Nothing more, nothing less.
Output: Neutral

One clearly positive, one clearly negative, one genuinely neutral — three examples, but they demonstrate the full range of the classification, not three variations on the same case.

Include the Edge Case That Actually Trips Things Up

If a task has a known tricky case — a review that's mixed ("great product, terrible shipping"), or a format that's easy to misparse — including one example specifically for that case is often worth more than adding a fourth or fifth straightforward one. The examples that prevent the most errors are usually the ones addressing where the model would otherwise guess wrong, not additional confirmations of the easy cases.

Match the Output Format Exactly

Whatever format the output examples use — a single word, a JSON object, a specific sentence structure — the model will tend to mirror that format closely. An inconsistency between examples (one plain "Positive", another wrapped in quotes as "Positive.") teaches an inconsistent pattern, which often shows up as a similarly inconsistent format in the actual output.

How Many Is Enough

Two to five examples is a reasonable range for most tasks — one example establishes a pattern loosely, but doesn't give the model much to generalize from beyond that single case. Beyond five, additional examples usually add prompt length without adding much new signal, unless the task genuinely has that many distinct cases worth demonstrating separately.