Starting experience

Even AI. Same subject. Two questions:

A: “Tell me about volcanoes.”

B: “You are a museum guide who talks to 10 year old children. Explains a blowout in 5 numbered steps, with a fun comparison per step. Maximum 10 lines.”

What has changed? Not the AI. The question.

On the program today

1. What is a prompt?
2. The recipe: the 5 ingredients of a good prompt
3. Magical techniques: lists, tables, roles, constraints
4. Iterate: improving your request is normal!
5. Chain: cut out big requests
6. The big prompt competition

What is a prompt?

A prompt = the message you send to the AI.
A question, an instruction, an instruction.

• English word which means “invite” / “instruction”
• This is all the AI knows about what you want
• She doesn't read your thoughts

The restaurant analogy

Customer 1: “I want to eat.”
→ The waiter brings… anything.

Customer 2: “A margherita pizza, thin crust, no olives, please.”
→ Exactly what he wanted.

The AI is an ultra-talented server… but not a soothsayer.

“Garbage in, garbage out”

(“Waste in, waste out”)

• Question fuzzy → answer fuzzy
• Question precise → answer precise

A bad result is not always the fault of the AI.
It's often the fault... of the prompt!

Good news: it can be learned. And it's today.

Ingredient 1: CONTEXT

“This is my situation”

What the AI needs to know about you and your needs.

“I'm in 5th grade and I have a presentation on bees on Friday.”

Without context, the AI responds “on average,” for no one in particular.

Ingredient 2: the ROLE

“You are…”

The character that the AI should play.

“You are a SVT teacher passionate about bees.”

• Changes the tone, the vocabulary, the examples
• One of the most powerful levers of prompting!

Demo: the power of the role

Same question: “Explain fractions to me”

Role	Result
(none)	Standard response, neutral tone
“6th grade math teacher”	Simple, educational, school examples
“Pastry chef”	Fractions explained with cakes

Three radically different answers!

Ingredient 3: TASK

“Do this” — with a precise action verb:

• Explain to me…
• Summarize for me…
• Compare…
• Correct…
• Invent…
• List…

“Explain how bees make honey.”

Ingredient 4: FORMAT

“Put it like this”- Bulleted list

• Table with columns
• Steps numbered
• Short text / dialogue / poem…

“In the form of 5 numbered steps.”

Ingredient 5: CONSTRAINTS

“With these limits”

• “In 3 sentences maximum”
• “For a 10 year old child”
• “With emojis”
• “Without complicated words”
• “In a funny / serious tone”

Constraint sculpts the response.

All assembled!

“ I’m in 5th grade and I’m preparing a presentation on bees.
You are a passionate SVT teacher.
Explain to me how bees make honey,
in 5 numbered steps,
with simple sentences and one emoji per step.”

You don’t need all 5 every time — but the more, the better!

Magic words

To ask without moderation:

• “Make me a list of…”
• “Present this in a table with the columns…”
• “Explained in numbered steps”
• “Summary in 3 sentences”
• “Give an example for each point”

Structure requested = clear answer.

Iterate: it’s NORMAL to start again

The first answer is not perfect? No one gets it right the first time — even the pros!

• “Too long → summary in 5 lines.”
• “Too complicated → explains as if I were 10 years old.”
• “Good, but gives concrete examples.”
• “Start again with a funnier tone.”

The AI remembers the conversation: no need to repeat everything.

The sculptor analogy

• 1st prompt → a rough block
• Each raise → a scissors blow
• At the end → exactly the desired shape

Talking to AI is not a single question.
It's a conversation.

The chaining of prompts

Cut a big request into small steps, like links in a chain.

“Give me my entire presentation on dolphins.”

Link by link:

1. “Give me 5 plan ideas.”
2. “Develop plan #2.”
3. “Write the introduction, 5 sentences max.”
4. “Suggest 3 questions that the class could ask.”

Why chain?

All of a sudden	Step by step
We suffer the result	We keep control
Average response everywhere	Every step is careful
We copy without understanding	We understand what we are building

The tools are improving quickly, but chaining remains the most reliable method to get started.

A good prompt isn't everything!

Even with a perfect prompt:

• The AI can make mistakes (hallucinations, remember!)
• She can invent details with confidence
• We always check important information

Good prompt = more chance of a good answer.
Not a guarantee.

Safety reflex

In a prompt, we never put:

• His full name + address
• Their passwords
• Private information about yourself or others

A prompt is a message sent over the Internet.
We think before we write.

Contest Rules

1. Teams of 2-3
2. A common challenge is announced
3. Each team writes its prompt on paper (5 min)
4. We test all prompts live
5. The class votes for the best answer
6. We analyze the winning prompt: how many ingredients out of 5?

May the quickest win!

To remember (the essentials)

1. Prompt = the instruction given to the AI
2. Fuzzy question → fuzzy answer (garbage in, garbage out)
3. The recipe: Context Role Task Format Constraints
4. Magic words: list, table, steps, “in X sentences”
5. Iterate, that’s normal — we specify, we refine
6. Big requests → one step at a time
7. Good prompt or not: we always check