What zero shot means
Zero shot prompting asks a language model to do a task using only an instruction, with no example pairs to copy. You describe what you want in plain language and the model relies on patterns learned during training to respond.
When it works well
- Common tasks the model has seen many times, like summarizing or translating.
- Clear instructions that name the input, the action, and the desired output.
- Simple formats where there is little ambiguity about how the answer should look.
Where it struggles
- Niche formats the model rarely encountered may come out inconsistent.
- Subtle rules about edge cases are easy to miss without a demonstration.
- Strict structure like exact field names often drifts unless you show a sample.
How to strengthen it
Be specific. State the role, the constraints, and the format you expect. A sentence such as classify the sentiment as positive or negative beats a vague ask. If zero shot output is shaky, adding one or two examples turns it into few shot, which usually lifts reliability.
Key idea
Zero shot prompting gives only an instruction and trusts the model to generalize, which works for familiar tasks but weakens when the format or rules are unusual.