A Quick Warm-Up Before the Scene

Opening Scene

A director sometimes runs a quick warm-up exercise with an improv troupe right before a specific scene — a couple of quick example rounds establishing the exact tone, format, or style wanted for what’s about to happen. The performers don’t need retraining; a few well-chosen examples, given right before the scene, are often enough to lock in exactly the right pattern for that specific performance. Language models can be guided in a strikingly similar way.

In Plain English

Few-shot prompting (or in-context learning) provides a model with a small number of example input-output pairs directly within the prompt itself, before asking it to handle a new, similar case. Remarkably, this lets a model adapt its behavior to match the demonstrated pattern — format, style, task type — without any actual change to its underlying parameters. This is genuinely different from fine-tuning, covered in Article 12, which does permanently adjust a model’s weights; few-shot prompting’s effect lasts only within that specific context window.

The Old Way

Before few-shot prompting was well understood as a distinct, reliable technique, getting a model to follow a specific format or pattern required different approaches:

Earlier machine learning approaches generally required retraining or fine-tuning a model to adapt it to any new task, a genuinely more expensive and slower process than adjusting a prompt.
Zero-shot prompting — simply asking a model to do a task with no examples — was often the only option before few-shot techniques were well understood, and it sometimes produced less reliable, less precisely formatted output.
Rule-based systems, covered in Article 3, required entirely new, explicit rules for each new task, with no equivalent of “just show it a couple of examples” available at all.

Few-shot prompting’s discovery — that a model could learn a task pattern purely from context, with no weight updates — was itself a genuinely surprising and significant finding.

What’s Changing (and Why AI Is the Reason)

Larger, more capable models have generally shown stronger in-context learning ability, connecting directly to the emergent capabilities covered in Article 14 — this specific capability scales meaningfully with model size and capability.
Few-shot prompting has become a standard, practical technique for quickly adapting a general-purpose model to a specific format or task, without the cost and complexity of fine-tuning.
This has shifted a meaningful amount of practical AI application development away from model retraining and toward careful prompt design, covered in far more depth in this content library’s dedicated prompt engineering series.

The Metaphor, Fully Extended

The Improv Scene	Few-Shot Prompting Concept
A quick warm-up establishing a scene’s specific tone before it starts	A few examples establishing a task’s specific pattern within the prompt
Performers adapting on the spot without retraining	A model adapting its output without any change to its underlying weights
The warm-up’s effect lasting for that one specific scene	Few-shot prompting’s effect lasting only within that specific context window
A director’s choice of which quick examples to demonstrate	A practitioner’s choice of which examples to include in a prompt

For Beginners: What to Actually Do

Practice writing a few-shot prompt for a simple task, comparing its output quality and consistency against a zero-shot version of the same task.
Learn the basic distinction between few-shot prompting (temporary, within-context adaptation) and fine-tuning (permanent weight adjustment), covered in Article 12.
Get comfortable choosing representative, well-formatted examples, since example quality directly affects how well the model picks up the intended pattern.

For Practitioners and Leaders: The Deeper Layer

Default to few-shot prompting before considering the more expensive fine-tuning route, covered in Article 12, for many practical adaptation needs.
Invest in prompt engineering capability as a genuinely valuable, distinct skill, connecting directly to this content library’s dedicated series on the topic.
Recognize in-context learning’s dependence on model scale and capability when choosing which model to deploy for a task requiring reliable few-shot performance.

Quick Recap

Few-shot prompting provides example input-output pairs directly in the prompt to guide a model’s behavior on a new task.
This adapts model behavior without any permanent change to its underlying parameters, unlike fine-tuning.
Larger, more capable models generally show stronger in-context learning ability.
This has shifted meaningful practical AI development effort toward careful prompt design.

Where This Fits in the Series

Article 10 covered a lightweight way to guide a model’s behavior. Article 11 covers what happens when a model’s confident-sounding output diverges from the truth entirely.