Few shot prompting unreliable

Few-shot prompting can feel unreliable because it depends heavily on example quality, order, and similarity to the real request. If the examples are too narrow or too polished, the model may not generalize well when the input changes shape.

The technique works best when the examples reflect actual edge cases, not just ideal ones. It also helps when the prompt is short, the instructions are clear, and the target behavior is tightly defined.

When it fails, it often means the prompt is carrying too much responsibility. In that case, retrieval, better task framing, or fine-tuning may produce more stable results than example-based prompting alone.