ChatGPT prompts: how to talk to AI like you know what you're doing

You've been using ChatGPT for months. You type questions as if it were Google, get a vaguely useful answer and think, "It's nice, but it'll never replace a real professional." Except the problem isn't the AI. It's how you talk to it.

I've seen clients get VERY different results with exactly the same tool simply by changing how they phrase a request. That's called prompting. Let's demystify it.

Your prompt is your brief

Imagine hiring a freelancer. Say "make me something marketing-related" and you'll get anything. Say "write a follow-up email for prospects who downloaded our SEO guide five days ago; relaxed tone, maximum 150 words, with a call to action for a free call", and you'll get something useful.

AI works the same way. A prompt is a brief. The more precise it is, the better the result.

The structure that works every time:

Context — who you are and the situation. "I'm a plumber in Angoulême with a five-page business website."

Role — which expert the AI should act as. "You're an SEO consultant specialising in local search."

Instructions — exactly what you want. "Write five Google Posts for my Google Business Profile."

Format — how you want the result. "Each post should be 100–150 words, with an emoji at the start of the title and a call to action at the end."

Job done. With those four elements, you move from generic answers to immediately usable results.

Three techniques that change everything

Chain-of-thought ("think step by step")

The simplest and most effective technique. Instead of asking directly for an answer, ask the AI to reason step by step. It works incredibly well for analysis, strategy and diagnosis.

Weak example: "What SEO strategy should I use for my site?"

Strong example: "Analyse my website semzen.fr step by step: 1) identify current SEO strengths, 2) list technical weaknesses, 3) propose five priority actions ranked by impact, 4) estimate the time needed for each action."

The difference? The first gives you a generic wall of text from any SEO blog. The second gives you a genuinely structured analysis.

Few-shot (providing examples)

Show the AI what you expect with two or three examples and it reproduces the pattern. It's EXTREMELY effective for content.

"Here's the style of my LinkedIn posts:

• Example 1: [your post]
• Example 2: [your post] Now write a post in the same tone about [topic]."

I use this technique for my blog posts. I provide a sample of my writing style — the relaxed tone, humorous asides… — and the AI follows it closely. If you want to see the result, you're reading it right now.

Persona (assigning a role)

"You're a Google Ads expert with ten years' experience, specialising in low-budget e-commerce campaigns (under €1,000 a month)."

A precise role sets the expertise level and vocabulary of the response. A "Google Ads expert" won't answer like a "general assistant". Try it; you'll see the difference.

The mistakes EVERYONE makes

Being too vague. "Help me with my marketing" → useless. Be specific about what you want, who it's for and the context.

Not iterating. The first result is rarely perfect. Say "that's good, but shorten the paragraphs", "add more figures" or "the tone is too formal; make it more relaxed". Each exchange helps improve the result.

Forgetting the output format. If you don't specify, the AI chooses for you, often producing an unreadable wall of text. Ask for a table, numbered list, email or draft with [placeholders].

Putting everything into one prompt. Break complex tasks down. First research, then analysis, then writing. Each step in a separate message. It's exactly the method I use when helping clients adopt AI : break it down, structure it and iterate.

ChatGPT, Claude, Gemini… are they all the same?

Prompting techniques work across ALL major models. But each has its strengths:

ChatGPT (OpenAI) — the most popular, good at creative writing and coding. GPT-4o mode performs best.

Claude (Anthropic) — my favourite for complex tasks and long-form writing. More nuanced, less robotic in its replies. And with Claude Code, it can even code directly in your projects.

Gemini (Google) — interesting for anything involving Google data (Analytics, Ads, Search Console), with direct access to real-time data.

Mistral — the French contender: capable and more privacy-conscious, with servers in Europe.

Honestly, your prompt quality matters more than the tool you choose. A good prompt on any model will always beat a bad prompt on the best model.

The prompt worth its weight in gold (my own template)

Here's the template I use every day for my clients:

Context: [describe your situation in 2–3 lines]
Role: You are [an expert in X with Y years of experience]
Objective: [what you want to achieve]
Constraints: [budget, deadline, style, length, target audience]
Format: [table / email / list / article / etc.]
Example: [optional — show the expected result]

I've refined this template over hundreds of client sessions. It works for emails, articles, strategies, competitor analysis, social posts… pretty much everything.

Now you have the basics to move from "ChatGPT is a gimmick" to "ChatGPT is my best collaborator". Start with the template above, test and iterate. If you'd like to go further, I help small businesses bring AI into their everyday work: let's discuss it whenever you're ready.

Let's get started!

Source: OpenAI — Prompt Engineering Guide, 2026

Translated from the original French article. Publication dates, examples and figures refer to that original version.