Zero-shot prompting

Zero-shot prompting means giving a model a task described in plain language, with no examples of finished answers. Every time you type "summarize this email in three bullets" into a chat window, you are writing a zero-shot prompt. This chapter covers where the term comes from, what a good zero-shot prompt contains, and how to tell when the task needs examples instead.

What zero-shot means

The term was set out in the 2020 GPT-3 paper by Brown et al., "Language Models are Few-Shot Learners". The authors compared several ways of getting a model to do a task. In the zero-shot setting, "no demonstrations are allowed" and the model receives only a natural-language instruction describing the task. One-shot adds a single demonstration. Few-shot adds several. None of the three change the model's weights: the task is specified purely through the text you send.

The paper also names the trade-off. Zero-shot offers "maximum convenience" and avoids the model copying quirks of particular examples, but it is "the most challenging setting". The authors give an example: a request to "make a table of world records for the 200m dash" can be ambiguous, because the format and contents of the table are not obvious from the instruction alone.

So the word "zero" counts examples, nothing else. A zero-shot prompt can still be long, detailed and full of context. It simply does not show the model a sample of the answer you want.

Zero-shot prompt examples

Each prompt below is zero-shot: it states the task, the input and the shape of the output, and shows no sample answer.

A reusable version with blanks to fill:

You are helping [who the output is for].
Task: [one sentence describing the job].
Input:
[paste the text, data or question]
Output format: [bullets / table with columns ... / one paragraph / JSON with keys ...].
Constraints: [length limit, tone, words or claims to avoid].
If something in the input is unclear, say what is missing instead of guessing.

Why zero-shot works as well as it does

Modern chat models are trained to follow instructions, and that training is the reason a bare instruction gets a usable answer. In "Finetuned Language Models Are Zero-Shot Learners" (Wei et al., 2021), Google researchers fine-tuned a 137-billion-parameter model on more than 60 tasks phrased as natural-language instructions. The tuned model, FLAN, beat zero-shot 175B GPT-3 on 20 of the 25 tasks they evaluated.

A second result shows how much a single added sentence can change. Kojima et al. (2022), "Large Language Models are Zero-Shot Reasoners", added "Let's think step by step" before each answer, with no examples. On the InstructGPT model text-davinci-002, accuracy rose from 17.7% to 78.7% on the MultiArith benchmark and from 10.4% to 40.7% on GSM8K. That is still zero-shot prompting, because the added text was an instruction. Reasoning prompts get their own chapter, chain of thought prompting.

How to write a strong zero-shot prompt

With no examples to lean on, the instruction carries all the weight. The vendor guides agree on a short list.

1. Name the task in one sentence. OpenAI's ChatGPT help center advises prompts that are "clear, specific" and give enough context to avoid ambiguity.
2. Describe the output. Length, structure, columns, reading level. Anthropic's guide says to be specific about "the desired output format and constraints".
3. Say why. Anthropic recommends giving the reason behind an instruction. Its example: tell Claude the response will be read aloud by a text-to-speech engine, rather than only writing "NEVER use ellipses", and the model generalizes from the explanation.
4. Run the colleague test. Anthropic's "golden rule" is to show the prompt to a colleague with little context on the task. If they would be confused, the model will be too.
5. Iterate. OpenAI suggests starting with a first prompt, reading the response and refining the wording or context from there.

For a longer walk through these steps, see how to write prompts. If you would rather start from a draft, paste a rough request into our ChatGPT prompt optimizer and edit what comes back.

When zero-shot is not enough

Zero-shot breaks down in predictable places. Add examples when you see one of these:

• The format drifts. You asked for a table and got a table most of the time, but column names or order change between runs.
• The labels are your own. Your team's definition of "urgent" or "qualified lead" is not something the model can know from a word.
• Tone is hard to describe. "Friendly but not chatty" means different things to different readers. One sample paragraph settles it.
• Edge cases keep slipping. If the model mishandles the same kind of input twice, show it that input with the right answer.

The vendors weigh this differently. Google's Gemini prompting guide says: "We recommend to always include few-shot examples in your prompts." Anthropic calls examples "one of the most reliable ways" to steer output format, tone and structure. OpenAI, for its reasoning models, advises trying a prompt without examples first and adding them only if needed. In practice, start zero-shot because it is faster to write, then add examples the moment the output is inconsistent. The next chapter, few-shot prompting, covers how to choose them.

Zero-shot, one-shot, few-shot and fine-tuning compared

The four settings from the GPT-3 paper, side by side:

Brown et al. note that fine-tuning typically uses thousands to hundreds of thousands of labeled examples. For a person working in a chat window, the real choice is between the first three rows. All of them fit inside the broader skill covered in prompt engineering, and the full chapter order is on the prompt engineering guide contents page.

FAQ

What is zero-shot prompting in simple terms?

You describe the task in words and let the model answer, without showing it any sample answers. The GPT-3 paper defines the zero-shot setting as one where the model gets only a natural-language instruction and no demonstrations.

Is zero-shot prompting the same as zero-shot learning?

They are related. In the GPT-3 paper, "zero-shot", "one-shot" and "few-shot" describe how many demonstrations the model sees in its prompt at inference time, with no weight updates. Zero-shot prompting is that setting applied to a prompt you write.

Does adding "Let's think step by step" make a prompt few-shot?

No. The phrase is an instruction, and the prompt still contains no examples. Kojima et al. (2022) call this zero-shot reasoning, and it raised accuracy sharply on arithmetic benchmarks with InstructGPT.

When should I switch from zero-shot to few-shot?

When the output format, labels or tone are inconsistent across runs. Google's Gemini guide recommends including examples as a rule, and Anthropic suggests 3 to 5 for best results.