Guide
Zero-shot prompting
Zero-shot prompting is an instruction with no worked examples. The model has to apply the task from the wording alone. It is the default way people chat, and it is the right default when the task is common and the output shape is simple.
Updated 2026-09-23
Before and after
Before
Make this nicer: the report is late because the data was wrong.
After
Rewrite the sentence as a status update to a manager. One sentence. State the fact, the cause, and the next step. Do not add an apology. Sentence: the report is late because the data was wrong.
How to do it
1. Name the task with a concrete verb
Summarize, classify, translate, extract, or rewrite. “Help me with” leaves the task open.
2. State the output limits
Length, language, and what to leave out. A zero-shot prompt with no bounds invites a generic essay.
3. Define custom words
If “urgent” means a specific rule in your team, write the rule. The model does not know your glossary.
4. Add examples only after a miss
If two zero-shot tries fail in the same way, that is the signal to move to one-shot or few-shot, not to write a longer essay of instructions.
When to use this
Use zero-shot for summaries, translations, straightforward rewrites, and questions the model already knows how to answer. Switch to few-shot when the label set is custom, the format is strict, or the first zero-shot answer keeps drifting. Switch to chain of thought when the task needs checkable steps.
Why zero-shot fails
Zero-shot fails when the instruction uses private terms, when several answers would all be “correct,” or when the format is unusual. The model then guesses the average answer from training. An example removes that guess.
Anthropic’s prompting docs describe the same split: start with a direct instruction, then add examples if the model does not follow the format. That is zero-shot first, few-shot second.
Sources
Questions
What is zero-shot prompting?
Zero-shot prompting gives the model an instruction and no examples. The model completes the task from the instruction alone.
Is zero-shot better than few-shot?
Zero-shot is better when the task is standard and you want a short prompt. Few-shot is better when the format or the labels are specific to you.
Can a zero-shot prompt still include context?
Yes. Background facts are context, not examples. Zero-shot means you did not show a completed input-output pair.