The Prompt Engineering article named several core techniques, starting with the simplest possible baseline. This article covers exactly that baseline: zero-shot prompting.
The simple definition
Zero-shot means asking a model to perform a task with no examples provided at all — just a description or instruction, relying entirely on what the model already learned during training. “Translate this sentence into French” with no example translation shown is a zero-shot prompt.
The “zero” refers to zero worked examples given in the prompt itself — the model has to rely purely on its pretrained knowledge and its understanding of the instruction itself.
Why this is genuinely worth naming as its own concept
Recall from the Large Language Model article’s discussion of emergent capabilities: before large-scale language models, getting a system to perform a new task usually required training it specifically on labeled examples of that exact task, as covered throughout the Data Handling phase’s discussion of supervised learning.
Zero-shot performance is notable precisely because it doesn’t require this at all — a model performing a task well from an instruction alone, with zero task-specific examples, was a genuinely surprising, headline-worthy capability when it first appeared at scale, not something taken for granted in earlier, smaller models.
flowchart LR
A["Prompt: 'Translate to French: The cat is sleeping.'"] --> B[No examples provided]
B --> C[Model relies entirely on pretrained knowledge]
C --> D["'Le chat dort.'"]ANALOGY vs. TECHNICAL REALITY
Analogy: Think of asking a well-read, broadly educated person to do something they’ve never specifically practiced — say, write a haiku about traffic.
They’ve never been trained specifically on “writing haikus about traffic,” but their broad general knowledge of what a haiku is, and what traffic looks like, lets them produce a reasonable attempt purely from the instruction alone, with no example haiku shown to them first.
Where this breaks down: A person draws on genuine, flexible understanding built from lived experience.
A model’s zero-shot ability comes entirely from statistical patterns absorbed during training across an enormous amount of text, as covered throughout the Training Mechanics phase — it has, in a real sense, seen countless examples of instructions and their corresponding correct-seeming completions during training, even if it’s never seen this exact instruction before, which is why zero-shot performance depends heavily on how well-represented similar tasks were in the training data.
Why zero-shot performance became a real, tracked benchmark
This is worth grounding concretely, since zero-shot capability is a genuine, historically significant measurement, not just a casual prompting technique.
OpenAI’s own GPT-3 paper, titled “Language Models are Few-Shot Learners,” explicitly reported and compared the model’s zero-shot, one-shot, and few-shot performance across many different tasks — treating zero-shot capability as a real, quantifiable property worth measuring and publishing, precisely because it demonstrated the model could generalize to new tasks without any task-specific training or examples at all, a genuinely different capability profile from the fine-tuned, task-specific models that preceded it.
A concrete example, layered
For a simple beginner example: asking a model “Is the following review positive or negative: ‘This restaurant exceeded all my expectations’” with no example reviews shown is a zero-shot classification task — the model has to infer both the task itself and the correct answer purely from the instruction.
For a production example: a startup building a content-moderation feature might first test whether a general-purpose model like GPT-4 or Claude can zero-shot classify user-submitted text as appropriate or inappropriate, simply by describing the categories clearly in the prompt.
If zero-shot performance is already good enough, it can save the company from having to collect and label a large task-specific training dataset at all, echoing the Annotation and Labeling articles’ discussion of how expensive that process can be.
Where zero-shot genuinely falls short
It’s worth being honest about a real limitation here, not presenting zero-shot as universally sufficient.
For tasks with unusual, highly specific formatting requirements, or tasks that differ significantly from anything well-represented in training data, zero-shot performance can be noticeably weaker and less consistent than approaches that provide concrete examples — precisely the gap the next article, Few-Shot, is built to close, by showing the model exactly what a correct response looks like rather than relying entirely on its ability to correctly infer the task from a description alone.
One task in zero-shot, one-shot, and few-shot form
Zero-shot
Classify as BILLING, TECHNICAL, or ACCOUNT:
“The app charged my card twice.”
One-shot
“I cannot sign in.” → ACCOUNT
Now classify: “The app charged my card twice.”
Few-shot
“I cannot sign in.” → ACCOUNT
“The screen stays blank.” → TECHNICAL
“My refund has not arrived.” → BILLING
Now classify: “The app charged my card twice.”The task is still zero-shot even though the model learned from enormous training data. “Zero” refers to examples of this task supplied in the current prompt, not zero prior training or zero instructions.
Use zero-shot first when categories are clear, the task is common, and you want the shortest prompt. Move to few-shot when labels are unusual, boundaries are subtle, or output style is difficult to describe precisely.
Real-world example: sorting support tickets immediately
A new online store has no prompt examples ready yet. It asks a model:
Classify this message as DELIVERY, PAYMENT, RETURN, or OTHER.
Return only the category.
Message: “The tracking page says delivered, but no parcel is here.”The likely output is DELIVERY. This is useful for a quick prototype because GPT and Gemini already learned broad language patterns during training. Before production, the store should test messages with mixed problems, different languages, sarcasm, and unclear wording, then route low-confidence or high-risk cases to a person.
Common misconception
A frequent beginner assumption: that “zero-shot” means the model has “zero knowledge” of the task, guessing blindly.
As this article has explained, this is backward — zero-shot specifically means zero examples in the prompt, not zero underlying knowledge; the model is still drawing on everything it absorbed during training about similar tasks, instructions, and patterns, which is exactly why zero-shot performance can still be genuinely strong for tasks that resemble things the model saw plenty of during training.
Where this fits in what comes next
You now understand the baseline case — no examples, relying purely on pretrained knowledge and instruction-following.
The next article, Few-Shot, covers what happens when you provide the model with a handful of worked examples directly in the prompt, and why that specific addition, as named in GPT-3’s own paper title, mattered enough to become the defining capability that paper is remembered for.
In one sentence
Zero-shot prompting asks a model to perform a task with no examples at all, relying entirely on pretrained knowledge and instruction-following — a genuinely notable capability that large language models made real, and one that works well for familiar-feeling tasks but often benefits from the concrete examples covered in the next article for anything more unusual.
Related Terms
- Author
- TechByteByByte Editorial Team
- Reviewed by
- TechByteByByte Admin
- Published
- Last reviewed