Zero-shot prompting — Direct instructions

Section 10.2.2 : Zero-shot prompting — Direct instructions

🎯 Learning objective

Mastering zero-shot prompting: obtaining quality results without providing any example, relying solely on clear instructions and the model's pre-existing knowledge. This is the most commonly used technique in daily practice.

Zero-shot: the power of pure instruction

Zero-shot prompting involves asking the model to execute a task without providing any example. The model relies solely on knowledge acquired during training — and with 2025-2026 models (GPT-5, Claude Opus 4.6, Gemini 3.1), that knowledge is massive.

It's the most natural form of interaction with an LLM — you ask a question or give an instruction, and the model responds directly. 90% of your daily interactions with AI will be zero-shot. Mastering it is therefore essential.

Why zero-shot works so well in 2026

Zero-shot quality has dramatically improved between 2023 and 2026. Three factors explain this progression:

1. →

   Training data scaling: GPT-5 was trained on an estimated corpus of 15-20 trillion tokens, covering virtually every imaginable professional format — emails, reports, analyses, code, presentations, contracts. The model has "seen" millions of examples of each document type.

2. →

   RLHF and alignment: Reinforcement Learning from Human Feedback (RLHF) has taught models to follow instructions much more faithfully. A clear prompt in 2026 gets results that even a very elaborate prompt wouldn't have produced in 2023.

3. →

   Extended context window: With 200K tokens (Claude Opus 4.6) or 1M tokens (Gemini 3.1), you can provide all necessary context directly in your zero-shot prompt — no need for examples when you can include the complete document, marketing brief, or conversation history.

When to use zero-shot?

Effective zero-shot examples by use case

Sentiment classification:

Classify the sentiment of this tweet as POSITIVE, NEGATIVE, or NEUTRAL.
Justify in 1 sentence.

Tweet: "The new MacBook Air is incredible, best 
battery life ever!"

→ Sentiment: POSITIVE
→ Justification: The user expresses enthusiasm 
  with "incredible" and "best ever."

Structured information extraction:

Extract the following information from this text and present them 
as JSON: name, company, position, email, phone (null if absent).

Text: "Hello, I'm Sophie Durand, marketing director 
at Databox (sophie@databox.fr). Call me at +33 6 12 34 56 78."

→ {"name": "Sophie Durand", "company": "Databox", 
   "position": "Marketing Director", 
   "email": "sophie@databox.fr", 
   "phone": "+33 6 12 34 56 78"}

Audience-based reformulation:

Reformulate this technical paragraph for 3 different audiences.
Keep the essential information but adapt the vocabulary.

Original text: "The microservices architecture decouples 
application components into autonomous services communicating via 
REST APIs, enabling independent horizontal scaling 
and continuous deployment via CI/CD pipelines."

1. For a CEO (business focus):
2. For a non-technical recruiter (issues focus):
3. For a computer science student (pedagogical focus):

Summary with constraints:

Summarize this document in exactly 5 bullet points.
Each point begins with an action verb and is < 20 words.
The first point is the most important.

[document content]

Contextual translation:

Translate this text into professional business English.
Don't translate word-for-word: adapt idiomatic expressions 
to the English-speaking context.

"Suite à notre dernier point de contact, je me permets de 
revenir vers vous concernant notre proposition commerciale."

Test data generation:

Generate 10 fictional but realistic entries for a French 
business contacts database. CSV format with columns:
last_name, first_name, company, position, email, city, annual_revenue (K€).
Data should be varied (sectors, sizes, cities).

The 5 principles of an effective zero-shot prompt

Principle 1: Specify the output format

Principle 2: Break down complex tasks

Principle 3: Specify the audience

Principle 4: Add negative constraints

"Do not" constraints are just as important as "do" instructions:

Write a pitch for our product.
DO NOT use superlatives ("the best", "revolutionary").
DO NOT exceed 50 words.
DO NOT mention competitors.

Principle 5: Ask for a confidence level

Answer this question and indicate your confidence level 
(1-10) for each part of your response.
If your confidence is < 7, state it explicitly and suggest 
how I could verify the information.

The 7 fatal errors in zero-shot (and how to fix them)

Even with a good prompt, certain structural errors systematically sabotage response quality. Here are the most frequent ones observed in enterprise settings:

Advanced zero-shot patterns for professionals

Beyond the 5 basic principles, experienced professionals use more sophisticated patterns:

The meta-prompt (asking the model to improve your prompt):

I want to achieve [objective]. Here's my current prompt: 
"[your prompt]"

Analyze this prompt and propose an improved version that:
1. Has more precise instructions
2. Produces better structured output
3. Avoids ambiguities

Give me the improved version, ready to use.

This pattern is particularly powerful for beginners: it turns the model into a prompting coach. As of March 2026, GPT-5 and Claude Opus 4.6 can diagnose your prompt weaknesses with remarkable accuracy.

The chain-ready prompt (structured output for chaining):

Analyze this customer feedback and return ONLY a JSON with:
{
  "sentiment": "positive|negative|neutral",
  "urgency": 1-5,
  "category": "technical|billing|delivery|other",
  "summary": "1 sentence max",
  "suggested_action": "1 sentence max"
}
No introduction, no explanation. Just the JSON.

This pattern is essential when connecting AI to an automated workflow (Make, n8n, API). The output must be parsable by a machine — hence the "just the JSON" constraint.

Practical exercise: mastering zero-shot

Duration: 20 minutes

1. →Classification: take 5 emails from your inbox and ask the AI to classify them into categories (urgent/important/informational/action required) with justification
2. →Extraction: copy a web page and ask for structured extraction of key information as JSON
3. →Reformulation: take a technical text you've written and request 3 reformulations for 3 different audiences
4. →Compare the results with and without the 5 principles above