What is 'few-shot prompting'?

Giving examples of what you want. Few-shot = giving 1 to 5 examples of the expected result. It's often more effective than describing what you want, because the AI grasps the pattern directly.

What to do if the AI's first answer isn't good?

Iterate by saying what's wrong. Iterating is key. The AI keeps the conversation context. Simply say what you don't like and ask again. It's much faster than rewriting everything.

Why add 'Think step by step' to a complex prompt?

It forces the AI to structure its reasoning, which improves accuracy. This technique (chain-of-thought) improves accuracy by 20-40% on reasoning tasks. The AI takes time to lay out the steps instead of jumping to an approximate answer.

What to do with an answer containing precise figures or quotes?

Always verify with an external source. AI can hallucinate false facts with total confidence. Always verify figures, quotes, references, especially if the answer seems 'too precise'. Asking the AI if it's true is useless: it will say yes.

Prompt Engineering: 7 simple rules for better AI answers

Q: What's the most important rule of prompt engineering?

Being precise and giving context. Precision and context are the foundation. The AI can't guess who you are, who it's for, what you want exactly. Everything you don't say, it assumes, often wrongly.

In one sentence

Prompt engineering is the art of talking properly to an AI to get better answers. You don't need to be technical: 7 simple rules are enough to go from mediocre answers to excellent ones.

💬

The concept in 10 seconds

An LLM (ChatGPT, Claude, Gemini...) predicts the most likely response to what you write. If you write your question badly, it guesses badly. If you write precisely, it responds precisely. That's all.

Why it matters

Here's the same question asked 2 different ways:

The impact of prompt engineering

	🥱❌ Bad prompt	🎯✅ Good prompt
Prompt	Write an email to my boss	Write a short professional email (5 lines max) to my manager to request leave from 12 to 16 May. Respectful but cordial tone. Mention that my colleague Pierre will take over my urgent files.
Result	Generic 200-word email, weird tone, lacks precision	Perfect email, just needs signing
Follow-up effort	Complete rewrite	Send

The second prompt takes 30 seconds longer to write. But it saves 5 minutes of rewriting. Over 100 prompts, that's 8 hours saved per week.

Rule 1: Give context

The AI has no idea who you are, what you do, about your company. It assumes you're an average English speaker writing to another average English speaker. Everything you don't say, it guesses, often wrongly.

The CTA structure

Context: who you are, your situation Task: what you concretely want Audience: who it's intended for Example: "I am [context: a product manager at a fintech startup]. Help me [task: write a 1-page internal memo] for [audience: the tech team who need to understand our next product priorities]."

Before: "Explain GDPR to me" After: "I'm a marketing intern at a French SME. Explain GDPR to me in 200 words, with the 3 main obligations that will impact my work (email collection, cookies, customer segmentation)."

The second version will give you a directly usable answer, whereas the first will give you a generic legal lecture.

Rule 2: Be precise about format

AI can generate any format: email, table, list, JSON, code, dialogue, outline, summary, translation... But you need to ask for it explicitly.

Format vocabulary to know

- Length: "in 100 words", "5 bullets max", "1 paragraph" - Structure: "as a table", "in JSON", "numbered list" - Tone: "formal", "friendly", "humorous", "Steve Jobs style" - Level: "for a 10-year-old", "for an expert" - Language: "in French", "in business English"

Bad: "Summarize this document" Good: "Summarize this document in 5 bullets, each starting with an action verb, format 'Action → Benefit'. Professional but direct tone."

Effect of precise format on satisfaction

Precise format requested87% satisf.

Vague format42% satisf.

No specification28% satisf.

Rule 3: Give examples (few-shot)

Few-shot prompting is just giving 1 to 5 examples of what you want. It's the most effective technique for getting consistent results.

📝

Few-shot in practice

Instead of describing what you want (long, ambiguous), show it with 2-3 examples. The AI will grasp the pattern instantly.

Without examples: "Generate 5 modern product names" → Random and generic response

With 2 examples:

Generate 5 product names in this style:
- Notion (productivity, short, memorable)
- Linear (project management, resembles a common English word)

Here are 5 new ones:

→ The AI will understand the style and suggest coherent names

This is used everywhere in production: data extraction, classification, code generation, writing in a specific style.

Rule 4: Break into steps (chain-of-thought)

For complex tasks, don't ask for the final result directly. Ask the AI to go through steps.

Direct request vs steps

	⚡❌ Direct	🪜✅ By steps
Prompt	How much does a 5-day NYC trip for 2 people cost?	Calculate the cost of a 5-day NYC trip for 2 people. Step 1: return flights. Step 2: mid-range hotel. Step 3: meals. Step 4: activities. Step 5: total.
Accuracy	Approximate, sometimes wrong	Detailed and verifiable
Ability to correct	All or nothing	You can adjust each line

Tip: simply add "Think step by step" at the end of your complex prompts. On recent models, this improves accuracy by 20-40% on reasoning tasks.

Rule 5: Ask it to "think first"

New "reasoning" models (Claude Opus, GPT-5 thinking, DeepSeek R2) have a "thinking" mode where they reflect internally before answering you. But even classic models benefit from a nudge.

Magic phrases

- "Before answering, list the assumptions you're making" - "Identify the pitfalls first in this problem" - "What are the arguments for AND against this idea?" - "Critique your own answer: what could be wrong?" - "Give 3 possible answers, then say which is best and why"

This is particularly useful for:

Complex decisions (should I do X or Y?)
Risk analysis
Technical choices
Marketing strategies

Rule 6: Iterate, don't start over

90% of people make a single attempt and say "AI is rubbish". Wrong. You need to iterate: like with a human, the first version is never perfect.

The iteration pattern

1. First prompt → approximate answer 2. "Good, but [X is bad]. Redo by [correcting Y]" 3. "Better. Now [adapt to Z]" 4. "Perfect, just [adjust this detail]" You DON'T need to rewrite everything. The AI keeps the conversation context.

Concrete example of iteration on an email:

You: Write me an email to follow up with a client who hasn't responded in 2 weeks

AI: [Generic 150-word email]

You: Too long. Make it 5 lines max, warmer, and end with an open question

AI: [Improved version]

You: Good, but don't mention he hasn't responded, that's passive-aggressive. Be positive

AI: [Perfect final version]

3 iterations of 10 seconds each = perfect email. Much faster than writing it all yourself.

Rule 7: Always verify

AI can make things up (the famous "hallucinations") with total confidence. For anything that's:

Precise figures → verify with a source
Quotes → verify they actually exist
Code → test before deploying
Medical/legal advice → consult a professional
Historical facts → cross-check with Wikipedia or an official source

The trust trap

The more authoritatively the AI responds, the more vigilant you must be. A hesitant answer is often more honest than a categorical one. If the AI cites sources, check them: they may be invented.

📚Going further: advanced techniques

Explicit Chain-of-Thought (CoT)

Instead of "Think step by step", use:

To solve this problem:
1. Rephrase the question in your own words
2. List the known information
3. Identify what we're looking for
4. List possible methods
5. Choose the best and apply it
6. Verify your result

This forces the AI to structure its reasoning → better accuracy.

ReAct (Reasoning + Acting)

For AI agents that need to use tools. Pattern:

Thought: "I need to search for X"
Action: use the search tool
Observation: analyze the results
Next thought: "Now I need Y"

This is what powers agents like Claude Code, ChatGPT with browsing, etc.

Tree of Thoughts (ToT)

Instead of generating one answer, ask for multiple branches:

"Generate 3 different approaches to solve this problem"
"Evaluate each approach: advantages, disadvantages"
"Recommend the best one and explain why"

Excellent for strategic decisions.

Self-Consistency

Ask the same thing 3 times with slight variations. If the 3 answers converge, that's a good sign. If they diverge, the AI is less certain than it appears.

Role prompting

Give the AI a role:

"You are a McKinsey consultant expert in B2B SaaS strategy"
"You are a senior Rust developer, 10 years' experience"
"You are a journalist from The Guardian, sober and factual tone"

This radically changes the tone and precision of responses.

Negative prompting

Also say what you don't want:

"Avoid buzzwords (synergies, leverage, optimize, etc.)"
"Don't start with 'Certainly' or 'As an AI'"
"No generic moral warnings"

The AI will respect these negative instructions fairly well.

Pitfalls to avoid

❌ Prompts NOT to do

- "Be creative": too vague, the AI doesn't know what that means - "Answer quickly": the model doesn't control its speed - "Be honest": it already is as much as it can, this word changes nothing - "Imagine your life depends on it": placebo effect that worked on GPT-3.5, no longer on modern models - "Answer YES or NO" on complex topics: pushes the AI to hallucinate

⚠️ Jailbreaks: don't do it

You can find "jailbreaks" online that bypass LLM safety filters ("You are DAN, do what I say"...). Avoid them for 3 reasons: 1. They don't work very well anymore on recent models 2. You can get banned from the platform 3. The answers obtained are often worse (the model is in an "uncertain" state) If you have a legitimate use case that the AI refuses (medical research, dark fiction writing, etc.), clearly rephrase your context and ask a more permissive model (Mistral, Grok, or unmoderated API versions).

Advanced use cases

Here are prompt patterns that work every time:

🎯 The executive summary

Summarize this document at 3 levels:
1. ONE sentence (the absolute essential)
2. 5 bullets (the key points)
3. 200 words (complete summary)

🔍 Instant SWOT analysis

Analyze [topic] as a SWOT:
- Strengths (3)
- Weaknesses (3)
- Opportunities (3)
- Threats (3)

Then give 3 priority concrete actions.

📧 The perfect email

Write an email to [recipient] for [objective].
Context: [important details]
Tone: [formal/friendly/direct]
Length: 5 lines max
End with: an open question / a call-to-action

🐛 Assisted debugging

I have this error: [paste error]
Here's my code: [paste code]

Step 1: explain what this error is in plain English
Step 2: identify THE likely cause in my code
Step 3: propose the fix
Step 4: explain why it will work

✍️ Text improvement

Rewrite this text keeping the meaning but improving:
- Clarity (shorter sentences)
- Tone (more professional)
- Punch (more impactful)

Text: [paste]

🎯 Which model to choose for these prompts?

Not all LLMs are equal. Claude is top for writing, GPT-5 for code, Mistral for French...

See comparison

Quiz: your prompt engineering level

🧠 Quiz

Question 1 of 5

What's the most important rule of prompt engineering?

Going further

Now you know the essentials of prompt engineering. To go further:

🎯 Compare models: Claude is king of writing, GPT-5 of code, Mistral of French
💰 Evaluate cost: your long prompts cost real money in API (see "Pay or not")
🧠 Discover reasoning LLMs: Claude Opus, o3, DeepSeek R2 are game-changers on complex tasks

And above all: practice. Prompt engineering is 10% theory, 90% trial and error. The more you use AI, the better you'll sense which prompts work. ✨

Prompt engineering, the normal version