Zero-Shot Prompting — AI Glossary

Reference details

Related terms

• Few-Shot Prompting
• Chain of Thought (CoT)
• Prompt Engineering

How it works

1. The prompt states the task and supplies no exemplars, so the entire specification is the instruction.
2. Instruction tuning is what makes this work: the model was fitted to request-response pairs, so an imperative is already a recognised shape.
3. The output format must be described rather than demonstrated, which is where zero-shot most often fails — a described format is less constraining than a shown one.
4. Because no examples narrow the task, the model's own interpretation governs, which is an advantage on common tasks and a liability on idiosyncratic ones.
5. The whole context budget stays available for the actual input, rather than being spent on demonstrations.

Example: What the examples were buying

Classifying tickets into four categories, a zero-shot instruction naming the categories costs about 60 tokens of prompt. A five-shot version adds five ticket-and-label pairs at roughly 80 tokens each, so 460 tokens — nearly eight times the prompt for the same task. If accuracy is equal, the examples bought nothing but cost and latency. If the categories have boundaries the names do not convey, the examples bought the boundary — and that is the question to test rather than assume.

Commonly confused with

The contrast with few-shot, and the default has flipped. Instruction-tuned models frequently do better with a clear zero-shot instruction than with examples, because badly chosen examples teach format quirks and narrow the model's interpretation of the task. Few-shot still wins when the output format is hard to describe but easy to demonstrate. Test both rather than assuming examples help.

When to use it

Reach for it when:

• The task is common and the instruction can state it unambiguously.
• Context is tight and demonstrations would crowd out the input.
• Establishing a baseline before deciding whether examples earn their tokens.

Reach for something else when:

• The output format is idiosyncratic, where one shown example beats a paragraph describing it.
• Category boundaries are conventions the names do not convey.
• The domain is niche enough that the model needs context, not just instruction.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Zero-Shot Prompting — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path