Talking to AI

Why the prompt changes everything

You saw in Module 1 that a generative AI is a probabilistic machine. Direct consequence: the more precise your prompt, the more you frame the probabilities, and the better the answer.

This is the golden rule you'll hear everywhere:

Why you should prompt well

1. To provide context. Otherwise you'll get a generic answer that's not useful.
2. To industrialise. A reliable prompt reused 100 times saves you hours per month.
3. To control format. Never let the AI decide the presentation. Ask for JSON, Markdown, a table, whatever, but ask for it.

The RACE method

RACE is the simplest framework to structure a professional prompt. It's an acronym: four letters that match the four questions to ask yourself before writing your instruction. You memorise it in under a minute, and you'll use it for life.

Let's break down each letter

R, the Role. You tell the AI what "hat" to put on. Instead of asking a question in a vacuum, you give it a perspective. "You are a brand strategy consultant" or "You are an employment lawyer" radically changes the quality of the response. The AI adapts to the register, vocabulary and reflexes expected of that role.

A, the Action. The central verb, what you want it to do. "Analyse", "summarise", "compare", "draft", "structure". Be precise, and stick to one action at a time. A prompt that mixes "Analyse, then draft, then translate" produces a confused output. Three sequential prompts work better.

C, the Context. This is where you slip in everything the AI can't guess: the target audience, business constraints, the brand, the deadline, the history. The more useful context you give, the more the response fits your real situation. It's also how you avoid generic "here are 10 tips to manage well" answers.

E, the Expectation. You tell it how you want the output: shape, length, structure. "Reply in 200 words" or "Format: Markdown table with 4 columns". Without this last brick, the AI picks for you, and it's rarely what you wanted.

🛠 Experience: build your own RACE prompt live

Fill the four fields below (or click a preset), and watch your prompt build in real time. When it looks right, copy it and paste it into Claude or ChatGPT.

Few-shot prompting

The idea is very simple. Before asking the AI to perform a task, you show it a few concrete examples of the result you expect. It's like calibrating a compass before setting off.

The term comes from the machine learning vocabulary: "few-shot" means "a few examples". By contrast, "zero-shot" means no examples at all, and "one-shot" a single example. The theory is well known: the more you calibrate, the more stable the output. Let's see it in practice.

🛠 Experience: compare 0, 1 and 3 examples on the same task

Our task: extract the company name from an email address. Click the three positions and watch the prompt change, the result, and the estimated quality.

This technique turns an ambiguous task into a reliable routine. It's what makes industrialised workflows possible: every row of your spreadsheet runs through the same prompt, and you get the same quality of output everywhere.

Chain of Thought (CoT)

Translated literally, "chain of thought". For complex tasks (analysis, reasoning, computation, strategy), you explicitly ask the AI to detail its reasoning step by step before concluding.

This dramatically improves quality, because it forces the model to "slow down" and build its reasoning token by token, instead of jumping to the conclusion. Remember Module 1: the AI predicts the next word from the previous ones. The more "reasoning beacons" it lays before the answer, the better the answer fits the problem.

🛠 Experience: same question, with and without CoT

Question put to the AI: a trader buys 7 shares at €23 each, sells 4 of them at €31, and keeps the rest. What is their net profit if the 3 remaining shares are now worth €19 each?

Useful variant: "Ask me questions"

Instead of demanding a result right away, ask the AI to put questions to you before producing its output. You enrich the context without having to anticipate everything.

I want you to draft an email to a prospect I'm going to follow up with.

Before writing it, ask me the 5 most important questions
to nail the tone, the context and the angle.

Only start drafting once I've answered.

Common pitfalls to avoid

🚫 Politeness and anthropomorphism

"Hello, could you please…" Pointless. The AI doesn't take offence. Politeness eats tokens for no semantic benefit. Go straight to the point: imperative or infinitive verbs.

🚫 The wall-of-text prompt

An 800-word prompt in one block, with no clear sections, is less effective than a 200-word structured one. Use sections (Role / Action / Context / Format) or Markdown with ### to frame the structure.

🚫 Leaving the format to chance

"Give me an analysis" → you get four unusable paragraphs. "Give me 3 bullets with a bold title and one key figure per bullet" → you get exactly what you need. The difference comes from you, not from the model.

🚫 Asking everything in a single prompt

AI progresses better through iteration. First prompt = framing. Second = tone polish. Third = final format. That's faster than a monster prompt that fails.

The parameters to know (without being technical)

These are the "knobs" behind the interface. You don't need to tune them by hand, but understanding what they do helps you interpret behaviour.

🌡 Temperature

From 0 to 1 (sometimes 2). Controls the randomness of the output.

• Temperature at 0: the model is deterministic. Ideal for factual tasks (code, analysis, data structuring).
• Temperature at 1: the model is creative. Ideal for brainstorming, design, free writing.

📚 RAG (Retrieval-Augmented Generation)

The term you hear everywhere right now, for good reason. RAG is the bridge between the AI and your private data.

Instead of letting the AI answer from its general knowledge (and hallucinate), a RAG system will:

1. Search your document base (Notion, Confluence, SharePoint) for the relevant passages.
2. Inject them into the prompt context.
3. Force the AI to formulate its answer strictly from those sources.

The result: an AI that "knows" your company without needing to be retrained. That's the foundation of most serious business chatbots.

🎯 Fine-tuning

Training a base model on your specific data. Heavier than RAG, more powerful, more expensive. Reserve it for cases where business tone or vocabulary are truly critical.

🧪 Sandbox: try it now

Three ready-to-use prompts. Click "Copy", then "Open in Claude" to test them live.

You are my admin assistant.

Here is an email I received:
"""
[Paste the email to summarise here]
"""

Give me:
1. A summary in 3 bullets max (what's being asked, deadline, stake).
2. A draft reply in 5 lines, professional but warm tone.
3. The urgency level in your view (low / medium / high)
   with a 1-line justification.

You are a brand strategy consultant at a digital agency.

I have a meeting tomorrow with [COMPANY NAME], sector [SECTOR].
The topic: they want [STATED NEED].

Before the meeting, I want a prep brief structured as follows:
- 3 key facts about the company (positioning, size, recent news
  if you know any)
- 3 likely challenges they face in their sector right now
- 5 smart questions to ask them to frame their need
- 2 angles I could propose

If you don't precisely know the company, say so clearly
rather than inventing.

Creative brainstorm. Mode: no self-censorship, we want volume
first, we'll sort later.

Context: we're launching [PRODUCT/SERVICE] for [TARGET AUDIENCE].

Give me 20 editorial angles for LinkedIn posts, organised
into 4 categories of 5:
- 5 "data / shock figure" angles
- 5 "counter-intuitive / you've been lied to" angles
- 5 "behind the scenes" angles
- 5 "opinion piece / strong position" angles

For each angle: a punchy title (max 10 words) plus the opening
hook (1 sentence).