When should you use Few-Shot Prompting?

Any task where the desired output has a specific shape, voice, or format. Anything generative — writing, code, JSON, summaries. Where the failure mode is 'the answer is correct but in the wrong format'. When a single descriptive instruction isn't getting the structure you want.

When NOT to use Few-Shot Prompting?

Simple factual questions — examples add noise without value. When token budget is tight and the task is easy. When the examples themselves bias the model into one of N answers you don't want.

How does Few-Shot Prompting work?

Provide 2–5 examples directly in the prompt, each showing INPUT → OUTPUT. The model treats these as the implicit pattern to follow. More examples = better adherence, but diminishing returns past ~5 examples for most tasks. Example variance matters: examples should cover edge cases, not all be near-duplicates.

All techniques

Glossary · Technique

Few-Shot Prompting

Also known as: In-context learning, k-shot

Show the model 2–5 examples of input → desired output before asking. Consistently the highest-ROI prompt move in real-world AI work.

When to use it

Any task where the desired output has a specific shape, voice, or format.
Anything generative — writing, code, JSON, summaries.
Where the failure mode is 'the answer is correct but in the wrong format'.
When a single descriptive instruction isn't getting the structure you want.

When not to use it

Simple factual questions — examples add noise without value.
When token budget is tight and the task is easy.
When the examples themselves bias the model into one of N answers you don't want.

How it works

1Provide 2–5 examples directly in the prompt, each showing INPUT → OUTPUT.
2The model treats these as the implicit pattern to follow.
3More examples = better adherence, but diminishing returns past ~5 examples for most tasks.
4Example variance matters: examples should cover edge cases, not all be near-duplicates.

Example

Lazy prompt

Convert this sentence to formal English: i wanna go home

Using the technique

Convert sentences to formal English.

Informal: i'm gonna grab lunch
Formal: I'm going to have lunch shortly.

Informal: u busy rn?
Formal: Are you available at the moment?

Informal: nah it ain't working
Formal: Unfortunately, it is not functioning.

Informal: i wanna go home
Formal:

Common pitfalls

Bad examples = bad output. Quality > quantity.
If examples are too similar, the model rigidly follows the pattern even when it shouldn't.
Long examples push the actual question to the back of context and can be ignored.

Where this came from

Brown et al., 2020 — GPT-3 paper formalized 'in-context learning' as the dominant zero/few-shot paradigm.

Related techniques

Zero-Shot Prompting

No examples — just the instruction. The baseline every prompt is implicitly trying to beat.

Chain-of-Thought (CoT) Prompting

Force the model to think step-by-step before answering. Dramatically improves accuracy on multi-step problems.

Persona / Role Prompting

Tell the model who it is. "Act as a senior X" changes accuracy, vocabulary, and tone more than any other single trick.