Key answer
There is no universal average improvement from adding one example. Results may improve considerably, stay unchanged or worsen. First identify what it should demonstrate: format, style, classification criteria or a solution method.
Full explanation
Min et al. found in classification and multiple-choice experiments with 12 models that shuffling demonstration labels did not significantly reduce performance in some settings. Input distributions, available labels and formatting mattered. Demonstration benefits may therefore reflect recognition of task form rather than entirely learning correct solutions. This does not license careless incorrect examples. Paper
Models also differ. Gemini's guide encourages examples of desired patterns; the initial DeepSeek-R1 report found few-shot prompting harmed some evaluations and recommended directly stating the problem and format first. This does not prohibit providing R1 with task material or show examples are ineffective on every R1 task. Gemini guide, R1 report, Sections 3 and 5
Find what is hard to express. If showing a layout is clearer than a lengthy description, provide a representative example and label what to imitate. If the task is already clear and the example only repeats an obvious answer, omit it initially. One positive example rarely establishes complex rules with exceptions; add discriminating cases.
The example demonstrates only the “conclusion–evidence–next step” structure. Answer from the current material; do not reuse its facts or conclusion.
Measure your own benefit by comparing the same model on the same new questions with and without examples over repeated generations. Separate format compliance, factual accuracy and satisfaction. One better output establishes help in that instance only.
References
3. Google, Prompt design strategies, official Gemini developer documentation