A team is building an AI assistant to categorize customer feedback into eight categories: pricing, product quality, shipping, support, ease of use, features, bugs, and "other." The categories are reasonably well-defined in writing, but several have overlap — e.g., "the app crashes whenever I try to upgrade my plan" could plausibly be bugs, pricing, or features. The team is debating whether to use zero-shot prompting (just describe the categories) or few-shot prompting (also include 2–3 example feedback items per category with their correct labels).
When does few-shot beat zero-shot for this task?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Few-shot examples are most valuable where category boundaries are ambiguous. A "the app crashes during upgrade" example labeled as bugs (with reasoning: "the crash is the user's primary issue; the upgrade is the trigger but not the category") teaches the model exactly the kind of judgment call it would otherwise guess at. For crisp, non-overlapping categories, zero-shot is fine and cheaper. The decision rule: use few-shot when judgment between alternatives matters; use zero-shot when categories are unambiguous.
Few-shot adds token cost on every call. For tasks with crisp categorization, that cost buys nothing. Universal best-ness is a non-criterion; match the technique to the task.
Some category boundaries are genuinely fuzzy in the real world — "is this a bug or a feature gap?" is a judgment call humans disagree on. No description, however clear, eliminates that fuzziness; examples encode the actual boundary the team wants drawn.
Larger models do better at zero-shot in many tasks, but they still benefit from few-shot when the task involves judgment that descriptions can't fully convey. Capability isn't the right axis for this decision.