Few-shot prompting — Learning by example

Section 10.2.3 : Few-shot prompting — Learning by example

🎯 Learning objective

Mastering few-shot prompting: guiding the model by providing concrete examples to obtain results that follow exactly the desired format, tone, and logic. This is the most powerful technique for consistency and customization.

Few-shot: teaching by example

Few-shot prompting involves providing a few examples (typically 2 to 5) before asking the actual question. The model learns the pattern from the examples and applies it to the new input. It's a form of in-context learning — the model isn't retrained, it understands the pattern from the examples.

Imagine you're onboarding an intern. Rather than explaining for 30 minutes how to write a meeting summary, you show them 3 exemplary summaries and say "Do the same". That's exactly what few-shot prompting does.

Structure of a few-shot prompt

The structure is always the same: instruction → examples → actual question.

[General instruction describing the task]

Example 1:
Input: [example input 1]
Output: [expected output 1]

Example 2:
Input: [example input 2]
Output: [expected output 2]

Example 3:
Input: [example input 3]
Output: [expected output 3]

Now:
Input: [your actual question]
Output:

Concrete few-shot examples by use case

1. Support ticket classification (format + categorization)

Classify each ticket into a category and urgency level.
Strict format: Category: [X] | Urgency: [High/Medium/Low]

Ticket: "I can't log in since this morning"
Category: Authentication | Urgency: High

Ticket: "Is it possible to export my data as CSV?"
Category: Feature | Urgency: Low

Ticket: "The app crashes when I open the reports tab"
Category: Bug | Urgency: High

Ticket: "How do I change the interface language?"
Category: User support | Urgency: Low

Ticket: "Your competitor offers a built-in chat feature, 
is this planned for your product?"
Category:

2. Product description generation (consistent tone + style)

Write an e-commerce product description in this exact style.
Each description: emoji + hook sentence + 3 features + 
closing sentence. Less than 50 words total.

Product: Bluetooth headphones
Description: 🎧 Dive into crystal-clear sound. 30h battery life, 
active noise cancellation, and ultra-light 180g design. 
Your everyday audio companion, from the office to the commute.

Product: Urban backpack
Description: 🎒 The city-dweller's ally. 15" laptop compartment, 
IPX4-certified waterproof fabric, and 12 organized pockets. 
From the coffee shop to the boardroom, it follows you everywhere.

Product: Smart thermos
Description: 🌡️ Your drink, always at the right temperature. 
LED temperature display, keeps hot 12h/cold 24h, 
500ml stainless steel. The companion for busy mornings.

Product: Sport smartwatch
Description:

3. Structured extraction (precise JSON format)

Extract named entities in the following JSON format.
Required fields: company, product, location, date.
If a field is absent, use null.

Text: "Apple announced the iPhone 16 in Cupertino on September 9."
JSON: {"company": "Apple", "product": "iPhone 16", 
"location": "Cupertino", "date": "September 9"}

Text: "Tesla will deliver the Cybertruck to European customers 
from Berlin in March 2025."
JSON: {"company": "Tesla", "product": "Cybertruck", 
"location": "Berlin", "date": "March 2025"}

Text: "Sanofi's new medication will be available 
in pharmacies in the first quarter."
JSON: {"company": "Sanofi", "product": "new medication", 
"location": null, "date": "first quarter"}

Text: "Google presented Gemini 3.1 at I/O in Mountain View 
on May 14."
JSON:

4. LinkedIn hook writing (personal style)

Rewrite this raw information as an engaging LinkedIn hook.
Style: punchy first line (hook), then 3-4 lines 
of context, then conclusion with a question.

Raw info: "I automated 40% of my tasks with AI"
LinkedIn post: 
I eliminated 3 hours of repetitive work per day.

Not by hiring someone.
Not by working less.
By using ChatGPT + Zapier on 5 key workflows.

The result: 40% of my tasks run on autopilot.
What's the next task you're going to automate?

Raw info: "Our startup went from 0 to €100K MRR in 8 months"
LinkedIn post:
€100K in monthly recurring revenue.

8 months ago, it was €0.
No fundraising. No paid ads.
Just a product that solves a real problem + word of mouth.

The 3 decisions that changed everything ⬇️
What's YOUR strategy to reach PMF?

Raw info: "I changed careers at 40 to join tech"
LinkedIn post:

How many examples to provide?

Few-shot pitfalls to avoid

Zero-shot vs Few-shot: the decision tree

Practical exercise: building your own few-shots

Duration: 20 minutes

1. →Choose a recurring task in your work (email writing, classification, meeting summary)
2. →Write 3 high-quality examples (real input + ideal output)
3. →Test the few-shot prompt on a new real case
4. →Compare with the same case in zero-shot — note the difference
5. →Save this prompt in your template library (cf. section 10.2.8)