← Tous les articles

Prompt Engineering · October 5, 2026 · 4 min de lecture

Your AI Prompts Are Fine. What Is Missing Is Context.

When AI output disappoints, we rewrite the prompt. The real gap is usually context: what a new hire would ask before starting. Here is how to supply it once and reuse it.

A manager pastes a request into an AI assistant: "Write a follow up email to a client who is late on payment." The result is polite, fluent and useless. Too soft for a client who has already been chased three times, too formal for a relationship built over ten years. The manager concludes the tool is generic and writes the email by hand. (The example is illustrative, but you will recognise the scene.)

What is actually happening

Most prompt training teaches phrasing: assign a role, specify a format, set a tone, ask for step by step reasoning. All of that helps. But the quality of an answer depends mostly on what the model knows about your situation, and it knows nothing. It has no access to this client's history, your payment policy, or the rules nobody ever wrote down.

A new hire would ask five questions before writing that email. The model asks none unless you tell it to. It fills the gaps with the average of everything it has read, which is why the output is correct and lifeless.

The strategic reframe

What many people conclude: we need better prompts, a prompt library, or a better model. What actually matters: context is the bottleneck, and context is a knowledge problem, not a writing trick.

In the MAKIA framework, this is the Knowledge dimension: what does the organisation really know, and in what usable form? The people who get good results are not better writers. They are better at stating what they already know.

Four implications for an SME or leadership team

  1. Write a context block once, reuse it everywhere. Who we are, who we write to, our tone, what we never say. One page is enough.
  2. Let the AI interview you. Instead of anticipating everything, ask it to put its questions to you before it produces anything.
  3. Separate stable context from today's task. The stable part lives in a document, a project or custom instructions. The task fits on one line.
  4. Treat context as a team asset, with clear limits. The unwritten rules of your best people are worth writing down. Also decide which information may be shared with which tool.

A context block template to adapt:

Standing context We are [company], [size], in [industry]. Our clients are [profile]. We address them formally / informally. Our tone: direct, warm, never pushy. We never promise [x]. We never mention [y]. When something is uncertain, say so instead of guessing.

And the instruction that turns the tool into a conversation partner:

Before you write anything, ask me up to five questions about the situation, the recipient and the constraints. Wait for my answers. Only then, write the draft.

A closing thought

If a new hire could not succeed at this task with your instruction alone, why would the AI? Which rules of your team exist today only in the heads of two or three people?

Try this week

Pick one written task you do every week. In 25 minutes: write a 150 word context block using the template above, run the task once without it and once with it, then compare the two results. Keep the block in a shared document, because the whole team can use it.

Partager cet articleLinkedInFil RSS