Concept Contrast Fable

Few-shot prompting and zero-shot prompting

Conclusion first:Few-shot prompting stuffs a few examples into the prompt so the model can understand the task by imitation. Zero-shot prompting gives no examples, only instructions. Which is better depends on the task—tasks with clear examples use few-shot to save debugging time, while general reasoning tasks use zero-shot for more flexibility. But the choice itself is a cost decision, not just an effectiveness decision.

Few-shot prompting stuffs a few examples into the prompt so the model can understand the task by imitation. Zero-shot prompting gives no examples, only instructions. Which is better depends on the task—tasks with clear examples use few-shot to save debugging time, while general reasoning tasks use zero-shot for more flexibility. But the choice itself is a cost decision, not just an effectiveness decision.

Fable

Bring drawings or bring instructions

Two new apprentices, Xiao Tang and Xiao Song, joined the workshop. Supervisor Lao Tan assigned each of them a task: Xiao Tang was responsible for making standard wooden stools, and Xiao Song was responsible for handling custom customer requests—some wanted to change dimensions, some wanted to switch wood, some wanted to add armrests.

Old Tan sent Xiao Tang a file: photos of three finished standard wooden stools, each labeled underneath with materials and dimensions. Xiao Tang followed them for two weeks, and the finished products looked almost identical to the photos. One photo showed an oak stool, so Xiao Tang used oak, with dimensions according to the numbers below the photo. Occasionally, when a photo's angle was unclear, Xiao Tang referred to the other two to fill in the missing details.

Xiao Song's side was different. Old Tan gave Xiao Song only one sentence: according to customer requirements, make wooden furniture that can be sat on, is stable, and looks good. In the first week Xiao Song delivered three pieces, two of which passed, and one had chair legs of unequal length. In the second week, all five pieces delivered passed. By the third week, Xiao Song began proactively adjusting designs based on the customer's floor plan, no longer merely executing instructions—he was understanding what "can be sat on, stable, and good-looking" meant in different floor plans.

One day Old Tan gave both of them a new task at the same time: the client asked to "make a chair suitable for reading on a balcony, with a place to put a teacup." Xiao Tang flipped through the three photos he had on hand, and none of them was a balcony chair. He found the closest one, a dining chair photo, scaled it down and added a cup holder, and submitted it. It was usable, but the seating feel was poor—the dining chair's backrest angle was not suitable for reclining reading.

Xiao Tang delivered, and Xiao Song delivered too. Xiao Tang's chair was sent back and revised twice; Xiao Song's passed on the first try. During the retrospective, Old Tan said: Xiao Tang's strength is when your example library has many samples; Xiao Song's strength is when your task requires understanding constraints from scratch. The way the two were trained determines the boundaries of their abilities.

Old Tan later adjusted task assignments. Standard products went first to Xiao Tang—many examples, small deviations, high efficiency. Tasks requiring exploration and adaptive judgment went to Xiao Song—giving him no examples instead forced him to understand the essence of the task. Each worked to their strengths, and overall output doubled.

In his year-end summary, Old Tan wrote: how many examples you give equals how much you fix the answer space. The more examples, the smaller the room for exploration. For tasks where the answer space is known, this is a good thing. For tasks where the answer space needs to be reconstructed, examples will trap your output in existing shapes.

When a new apprentice joins, Old Tan asks a test question: if you were given an unfamiliar manual task, would you rather receive three photos of the finished product, or a one-sentence task description? There is no right or wrong answer, but the answer determines which kind of task Old Tan assigns to him later.

Few-shot prompting

Provide several input-output examples in the prompt so the model completes similar tasks by imitating the pattern of the examples. Suitable for scenarios where format, style or judgment criteria are relatively fixed.

Zero-shot prompting

Provide no examples, only describe the task requirements through instructions, and let the model rely on its own knowledge and reasoning ability to complete it. Suitable for scenarios requiring flexible understanding, general reasoning, or novel tasks.

Story Mapping

Xiao Tang works from three photos: few-shot prompting—completing a task by imitating examples, where output quality depends on how well the examples cover the space.
Xiao Song received only one task instruction: zero-shot prompting—completing tasks by understanding instructions, with higher flexibility.
In the balcony chair task, Xiao Tang couldn't find an example: the boundary of few-shot—tasks outside the example library tend to get matched to unsuitable patterns.
Old Tan assigns apprentices by task type: the choice of prompting strategy depends on task characteristics—some tasks suit imitation, others suit reasoning. Neither is absolutely better; there's only a degree of fit.
The number of examples determines the size of the answer space: few-shot locks down the output distribution, while zero-shot preserves a larger space for exploration.

In the project

Choosing a prompting strategy is a two-step process. First, determine the task type: are the output format and judgment criteria fixed? If so, few-shot examples are your cheapest tuning method. If the task requires understanding different constraints each time, the combination of zero-shot plus refined instructions is often more stable. Second, calculate the cost: the extra tokens consumed by few-shot translate into the labor cost of example maintenance and version management, while the tokens saved by zero-shot may be spent on more iterative debugging rounds. Calculate both dimensions before choosing, and do not rely on habit alone.

Few-shot prompting Zero-shot prompting