Two people get very different results from the same AI model, depending on how they phrase their request. "Prompt engineering" isn't esoteric — it's the art of giving an instruction that a competent human could also follow without ambiguity.
It builds on understanding how the models work — see our feature on artificial intelligence.
01 — The framing
Role, goal, audience
Tell the model what role to play, what it should produce, and for whom. "You are a lawyer; draft a confidentiality clause for a small business, in plain English" is a thousand times better than "write me a confidentiality clause."
02 — Context and examples
Show rather than describe
- Provide the information the answer depends on (data, constraints, expected tone).
- Give one or two examples of the desired result: this is the single most effective lever.
- State what not to do, when that's useful.
03 — The output format
Framing it so it can be reused
Explicitly ask for the shape: a bullet list, a table, JSON with specific fields, a maximum length. A constrained format is easier to check and to integrate.
04 — Complex tasks
Break it down and chain it
Instead of one giant prompt, chain steps together: first an outline, then drafting section by section, then a critical review. Each step can be checked.
05 — Iterate, then switch tools
When the prompt is no longer enough
You refine through successive attempts. But if the answer depends on up-to-date or organization-specific data, no prompt will be enough: you need retrieval-augmented generation (RAG) — see RAG explained simply.
06 — Frequently asked questions
2 minutes to understand AI and prompt engineering (in French) — Wild Code School
Frequently Asked Questions
Is there a perfect prompt?
No. There are clear, well-contextualized prompts tested for a given task. The same prompt doesn't transfer as-is to another case.
Do phrases like "act as an expert" help?
Giving a precise role helps; hollow formulas don't. Context and examples carry far more weight than incantations.
How do you verify an AI response?
By asking for sources, cross-checking the facts, and keeping a human review for anything consequential.
07 — Resources & links
Where to go, concretely
Understand
- NEWTIV — Artificial intelligence
- How the models work, so you can steer them better.
- newtiv.com/article/les-intelligences-artificielles.html
- NEWTIV — AI for small businesses
- Concrete use cases to put to work.
- newtiv.com/article/intelligence-artificielle-petites-entreprises.html
Réagissez
Commentaires