Zero-Shot Prompting: 5 Recipes That Ship

Quick Answer: Zero-shot prompting is giving a model a task with instructions only, no examples. As of 2026 it is the right default for most tasks: a tight output contract usually matches few-shot accuracy and costs fewer tokens on every call. Below are five zero-shot recipes, what each returns, and the exact point where you should stop and add one example instead.

Everyone reaches for examples too fast. Start here.

You describe the task. The model does it. No sample inputs, no sample outputs, no "here are three examples." That is zero-shot. It is the leanest prompt you can write, and for most tasks it is enough.

The brands referenced here: Anthropic (Claude) and OpenAI both document zero-shot as the baseline, linked below.

What is zero-shot prompting?

Zero-shot prompting is a prompt that contains instructions and no examples. The model relies on what it already learned in pretraining to complete the task. Contrast few-shot prompting, where you paste example input-output pairs into the prompt to steer the format.

One line to remember: zero-shot describes the target, few-shot demonstrates it.

The default nobody admits: reach for zero-shot first

Here is the counter-take. Few-shot is overused.

Every example you paste rides along on every single call, forever. Three examples at roughly 200 tokens each add about 600 tokens to each request. Run that a million times and it is a real line on your bill, not a rounding error. See what those tokens actually cost in BudgetForge's Claude API pricing teardown.

A zero-shot prompt with a named output contract often matches few-shot accuracy at a fraction of the tokens. So the rule: start zero-shot, measure, and add examples only when a specific failure forces it. Do not pay for examples you never needed.

5 zero-shot recipes that ship

Recipe 1: Name the format instead of showing it

Most people add examples just to lock the output format. You can describe the format in words for free.

Extract every company name from the text below.
Return a JSON array of strings, deduplicated, sorted A-Z.
No prose, no code fences, just the array.

TEXT:
"""
{paste text}
"""

Outputs: a clean JSON array, no examples required.

Why it works: the model already knows JSON. You only had to name the shape, the dedup rule, and the sort order.

Fails when: the schema is deeply nested or idiosyncratic. Then one example is worth 100 words of description. That is your signal to upgrade to one-shot.

Recipe 2: Give it a role, not examples

A role is compressed context. It replaces a stack of examples with one noun.

You are a senior copy editor. Rewrite the sentence below to be
clearer and shorter without changing its meaning. Return only
the rewritten sentence.

SENTENCE: {paste}

Outputs: an edit in the voice of the role, no sample edits needed.

Why it works: the role pulls a whole distribution of behavior from pretraining, which is why a named role pairs so well with zero-shot.

Fails when: the role is generic ("you are a helpful assistant"). Name a specific role or skip it.

Recipe 3: Enumerate the label set for classification

Classification is where zero-shot quietly wins. You do not need labeled examples, you need a closed list.

Classify the support message below as exactly one of:
BILLING, BUG, FEATURE_REQUEST, OTHER.
Return only the label.

MESSAGE: {paste}

Outputs: one label from your set, every time.

Why it works: constraining the output to an enumerated set removes the model's freedom to invent categories.

Fails when: you leave the set open ("classify this message"). You get a new label every run. Always enumerate.

Recipe 4: Add an escape hatch

Zero-shot hallucinates most when the answer is not in the input. Give the model a way out.

Answer the question using only the passage below. If the passage
does not contain the answer, return exactly: UNKNOWN.

PASSAGE: {paste}
QUESTION: {paste}

Outputs: a grounded answer, or a clean UNKNOWN instead of a confident guess.

Why it works: naming an explicit "I do not know" token gives the model a lower-risk path than fabricating.

Fails when: you omit the escape hatch. The model will fill the gap with plausible fiction. For the fabrications that still slip through, keep a small eval that flags any answer not grounded in the passage.

Recipe 5: The one-shot upgrade path

Zero-shot has a ceiling. Know when to add exactly one example, not ten.

Convert the input to our changelog format.

EXAMPLE
input: fixed the login redirect bug
output: - Fixed: login redirect no longer drops the return URL

Now convert:
input: {paste}

Outputs: the format locked by a single demonstration.

Why it works: one well-chosen example teaches a house-specific format that words struggle to pin down. One example, not five, is usually the whole jump in accuracy.

Fails when: you keep piling on examples. Diminishing returns start fast. If one example does not fix it, the task probably needs few-shot prompting done deliberately, or a different approach entirely.

When zero-shot is the wrong call

• House-specific formats that words cannot capture. Show one example.
• Style or voice mimicry. You have to show the voice, not describe it.
• Edge-case-heavy extraction where the rules only emerge from samples.

Everywhere else, zero-shot first. It is faster to write, cheaper to run, and easier to debug. For the full technique index, see the prompt engineering cheat sheet. The canonical reference for the term lives in the Prompt Engineering Guide, and both Anthropic and OpenAI treat it as the baseline you tune from.

Write the instruction. Skip the examples. Add them back only when the data tells you to.

Cost to test: $0.00 (every recipe runs on a free-tier chat).