Artificial Idea | AI careers · practical prompts · no hype Thursday, December 18, 2025 · Issue #38c · Prompt Tutorial
The example effect
Few-shot prompting: how to train AI to match your style with just 3 examples
You have been telling AI what you want. The professionals getting the best outputs have been showing it. The difference in output quality is immediate and significant.
Issue #39 made the case that the management challenge of AI adoption is fundamentally a psychological safety challenge rather than a technical one. This issue addresses a technical challenge that is also fundamentally a communication one: how to get AI tools to produce outputs that match your specific style, your specific standards, and your specific professional context rather than the style, standards, and context of the average professional.
The technique is called few-shot prompting. It has been in the academic AI literature since 2020 and in practitioner use since the first large language models became publicly accessible. It is also, in the observation of this newsletter across five months of reader interactions, one of the least used techniques among non-technical professionals despite being one of the most immediately impactful.
The gap between awareness and use is partly explained by the technique's name, which sounds more technical than the underlying concept warrants. Few-shot prompting means showing the model examples of what you want before asking it to produce something. That is the entire technique. The sophistication is in understanding why it works and how to choose the right examples.
Why examples work better than descriptions
The intuitive approach to getting AI to produce outputs that match your style is to describe your style. Write in a professional but direct tone. Use short paragraphs. Avoid jargon. Lead with the key point. These instructions are not useless. They produce modest improvements over no instruction at all. They are also significantly less effective than showing the model three examples of the style you want and asking it to match them.
The reason is grounded in how language models learn. They learn from examples rather than from rules. The training process that produced them involved exposure to enormous quantities of text, from which they learned patterns of style, structure, and tone through statistical association rather than through explicit rule application. When you show a model an example of what you want, you are activating the patterns it learned from similar text in training. When you describe what you want in abstract terms, you are asking the model to translate that description into those same patterns, which introduces an additional inference step that reduces the precision of the match.
A 2023 study by researchers at Stanford's Center for Research on Foundation Models found that few-shot prompts, defined as prompts containing three or more examples of the desired output type, produced style-matched outputs that independent raters judged as more consistent with the target style than description-based prompts in 78% of comparison pairs. The effect was largest for distinctive professional styles, specifically those that deviated from the generic professional writing that description-based prompts tend to produce.
How to choose the right examples
The quality of few-shot prompting output is directly determined by the quality and appropriateness of the examples provided. Three guidelines for selecting them.
First, choose examples that represent your best work rather than your typical work. The model will learn from and attempt to replicate the style of whatever you provide. Providing examples of work you were proud of rather than work that was merely adequate calibrates the model toward the standard you want to achieve rather than the standard you usually achieve.
Second, choose examples that are specific to the task type you are prompting for. Examples of your email writing style will not transfer reliably to your report writing style, because the stylistic patterns relevant to each are different. Maintain a small library of examples organised by task type, adding to it when you produce work you consider particularly good.
Third, choose examples of appropriate length relative to the output you are requesting. Examples that are significantly longer or shorter than the desired output will produce outputs calibrated to the length of the examples rather than to the length you need.
The five few-shot prompts
These five prompts apply the few-shot technique to the professional writing tasks most commonly encountered across the readership of this newsletter.
Prompt 1: Style matching for professional writing
I am going to show you three examples
of my professional writing style.
Study them carefully before
completing the task that follows.
Example 1: [paste a piece of your
professional writing you consider
representative of your best work]
Example 2: [paste a second example,
ideally of a different format
or topic to show the style
is consistent across contexts]
Example 3: [paste a third example]
Now that you have studied these examples,
please complete the following task
in a style that closely matches
the examples I have provided:
Task: [describe the writing task specifically]
Before writing, briefly describe
the key stylistic characteristics
you have identified in my examples,
so I can correct any misreadings
before you produce the full output.The instruction to describe the identified stylistic characteristics before producing the output is the quality control step that makes this prompt more reliable than simply providing examples and requesting output. Misreadings of example style are common and catching them before the full output is produced saves the iteration time of correcting a long piece that missed the style in a systematic way.
Prompt 2: Tone calibration for a specific audience
I am going to show you three examples
of how I communicate with
[describe the audience: senior leadership,
clients, technical peers, external partners].
Example 1: [paste example]
Example 2: [paste example]
Example 3: [paste example]
These examples demonstrate the tone,
level of detail, and framing
I use with this specific audience.
Please write the following communication
in a style consistent with these examples:
[describe the communication task]
If anything about the task requires
a different approach than the examples
suggest, flag it rather than
defaulting to the examples.
The examples are the standard
for routine communications with
this audience, not for every situation.The caveat at the end, that the examples are the standard for routine communications rather than every situation, prevents the model from mechanically applying a style that may be inappropriate for a communication that differs meaningfully from the examples in content or stakes.
Prompt 3: Analytical output calibration
I am going to show you three examples
of how I structure analytical outputs:
the way I present findings,
frame recommendations, and
communicate uncertainty.
Example 1: [paste analytical output]
Example 2: [paste analytical output]
Example 3: [paste analytical output]
Please conduct the following analysis
and present the output in a structure
and style consistent with my examples:
Analysis task: [describe what
you need analysed and what
information is available]
Pay particular attention to:
how I handle uncertainty in the examples,
how I structure the relationship
between findings and recommendations,
and the level of detail I include
versus what I leave for the reader
to infer.Prompt 4: Feedback style calibration
I am going to show you three examples
of feedback I have given to colleagues
or direct reports that I consider
representative of how I communicate
constructive criticism.
Example 1: [paste feedback example]
Example 2: [paste feedback example]
Example 3: [paste feedback example]
Please draft feedback on the
following situation in a style
consistent with these examples:
Situation: [describe the work,
behaviour, or output you need
to give feedback on, and what
the key messages are]
The feedback should maintain
my voice rather than defaulting
to generic feedback frameworks.
If the situation is significantly
different from the examples in
its sensitivity or stakes,
flag that before drafting.Prompt 5: The style extractor
This prompt is for professionals who want to build a few-shot library but are not sure what their distinctive stylistic characteristics are. It uses the model to analyse examples and extract the patterns, which can then be used to inform future prompts.
You are a professional writing analyst
helping me understand my own writing style.
I am going to provide five examples
of my professional writing across
different contexts and formats.
Example 1: [paste]
Example 2: [paste]
Example 3: [paste]
Example 4: [paste]
Example 5: [paste]
Please analyse these examples and identify:
1. The three most distinctive
characteristics of my writing style,
the ones that consistently appear
across all five examples and
that distinguish my writing
from generic professional prose
2. The structural patterns I use
consistently: how I open,
how I transition between ideas,
how I close
3. My consistent approach to
hedging and certainty:
when I qualify claims and
when I state them directly
4. The vocabulary choices that
recur across examples and
that seem characteristic
rather than coincidental
5. The one thing about my writing
that is most distinctive and
that I should explicitly include
in any prompt asking a model
to write in my style
This analysis will form the basis
of a few-shot library I will use
to configure AI tools for
my professional writing.
Precision matters more than completeness.The style extractor is the foundational prompt for professionals who want to build a few-shot practice systematically. Understanding what your style actually is, rather than what you believe it to be, is the prerequisite for choosing examples that represent it accurately and for evaluating whether model outputs are genuinely matching it.
Building the few-shot library
The few-shot technique reaches its full value when examples are collected and organised as a working library rather than assembled fresh for each prompt. A professional who maintains a library of ten to fifteen examples organised by task type has a resource that improves every AI interaction it is applied to, requires no additional development effort once built, and improves passively as new examples are added from work the professional considers particularly good.
The library takes approximately ninety minutes to build from existing work and fifteen minutes per month to maintain. It is one of the lowest-effort, highest-return investments in AI capability available to a professional who is already producing quality work and wants AI tools to produce outputs consistent with that quality.
The examples already exist. They are in your sent folder, your document library, your past presentations. The few-shot library is built from work you have already done. Its value is in organising that work into a form that makes every future interaction with AI tools more calibrated to who you actually are as a professional rather than to who the average user is.
The average user is, as this newsletter has said from the beginning, nobody's target.
Monday we return to the year-end analysis with the piece on what 2025 actually settled in the AI and careers story. Three issues caught up, calendar restored, and the final weeks of the year back on track.
See you Monday.
— The Artificial Idea team
