Prompt Engineering: How to Talk to AI Effectively

Two people who ask a generative AI the same question can get radically different results. The difference does not come from the tool, but from the way the request is phrased. At Adelya, this observation directly shaped the design of the sales copilot deployed by Qolaig: faced with insufficient adoption when salespeople had to invent their own wording, adding a library of pre-recorded prompts unlocked adoption. Here is what prompt engineering is, the principles that really change the quality of answers, the most common phrasing mistakes, and how to train your teams without turning it into a theoretical course that nobody will remember.

What is prompt engineering

Prompt engineering refers to the art of phrasing an instruction to a generative AI to get the most relevant result possible. A “prompt” is simply the request you address to the tool: a question, an instruction, a text to rephrase.

It is not a technical skill reserved for developers. It is a way of structuring your request, accessible to any employee, and closer to clear communication than to computing. A good prompt resembles a good brief given to a colleague: you explain the context, the objective and the expected result.

The Adelya case: from free phrasing to a prompt library

The first deployment of Adelya's sales copilot relied on questions asked freely in natural language in Microsoft Teams: “Generate a sales pitch for product X aimed at a high school”, “What is the difference between products X and Y?”. Technically, the agent worked well. But adoption remained insufficient: faced with a blank page, many salespeople did not know how to phrase their request effectively.

Qolaig then developed an interface dedicated to field teams, with a library of pre-recorded prompts matching the team's recurring questions: similar products, customer history, comparisons, segmented sales pitches. The salesperson clicks, fills in a variable, and gets the answer, without having to invent any wording.

Result: immediate adoption, without heavy training. For the full context of the project, our article on the AI copilot in the enterprise details the Adelya case.

The principles that really change the quality of answers

Give context. An AI does not guess your situation. Specifying the industry, the target audience or the constraint to respect immediately steers the answer. Adelya's pre-recorded prompts include precisely this context by default, with the type of customer and the type of product.

Be precise about the expected format. Explicitly asking for a list, a table, a formal or informal tone, or a length avoids having to rephrase several times.

Give an example. Showing an example of the desired result guides the AI far more effectively than an abstract description. It is often the fastest way to get the right style.

Proceed step by step. For a complex task, splitting the request into several successive exchanges often gives a better result than a single instruction that is too broad.

These principles are reflected in Anthropic's documentation, which offers an official prompt engineering guide with concrete examples on clarity, examples and breaking tasks into steps.

A before and after example

Take a very common request: “Write me a product pitch”. The answer will be generic, with no link to your customer or your sector. Rephrased as “Write a pitch for product X aimed at a high school, with an educational tone and three arguments maximum, as a list”, the same AI produces a result that can be used directly.

The difference comes down to one sentence of context and two format instructions. This is exactly the logic applied in Adelya's pre-recorded prompts: the context and the format are already included, and the user only has to fill in a variable.

A simple prompt structure to remember

Many good prompts follow the same five-part skeleton. The role: “You are a sales advisor in B2B distribution”. The context: “My customer is a high school renewing its cleaning products”. The task: “Write a sales pitch”. The constraints: “Three arguments maximum, educational tone”. The format: “As a list, followed by a concluding sentence”. You do not need to fill in everything every time, but running through this list before sending avoids most disappointing answers.

A few standard prompts by job

For a salesperson: “Prepare three discovery questions for a meeting with a customer in this sector.” For an HR manager: “Summarize this CV in five lines according to the job criteria, without interpreting.” For an accountant: “List the points to check in this invoice before payment.” For a lawyer: “Rephrase this clause in plain language and flag any ambiguities.” These templates are starting points: each team adapts them to its own vocabulary and enriches them over time.

The most common phrasing mistakes

Staying too vague. A prompt without context gives a generic result. The more precise the request is about the audience, the objective and the format, the more useful the answer.

Trying to get everything at once. A request overloaded with different instructions often produces an answer that handles each of them poorly.

Never rephrasing after a disappointing result. The first answer is never final: adjusting the request according to what is missing is the fastest way to reach the right result.

Forgetting to review. Even with an excellent prompt, the AI can make mistakes. Any factual information must be checked before being used.

Putting sensitive data in the prompt. A well-written prompt does not need names, amounts or confidential information: they can almost always be replaced by general descriptions.

How to train your teams in prompt engineering

Start from real everyday cases. Abstract training on prompt principles remains theoretical. Adelya's success comes from the fact that the prompt library was built from the questions the team actually asked, not from generic cases.

Build an internal prompt library. The best wording found by one team can be reused by others, which speeds up collective adoption, as shown by Adelya's move from free use to guided use.

Accept iteration. Prompt engineering is learned by trial, not by theory alone. Setting aside practice time on real tasks is worth more than a long presentation.

How to measure that prompts are improving

To know whether a prompt is good, compare its result with that of the previous version on the same task. Note three things: the time saved, the number of corrections needed before the text can be used, and the user's satisfaction. When a phrasing gives good results repeatedly, add it to the shared library with a clear name and an example. A library of ten to twenty well-tested prompts already covers a large share of a team's needs.

According to the Liferay 2026 Agentic AI Adoption and Governance Report, IT assistance and support functions are among the growing uses of agentic AI, an area where the quality of request phrasing makes a measurable difference to adoption.

FAQ

Do you need to be technical to do prompt engineering?
No, it is above all a skill of clear communication. The Adelya case even shows that a good ready-to-use prompt library can entirely remove the need to phrase a request yourself. Our article on generative AI in the enterprise shows where these skills pay off first.

Will prompt engineering remain useful as AI becomes more capable?
Models are improving in their ability to interpret imprecise requests, but the clarity of the context and the objective remains a decisive factor in the quality of the answer. Thinking clearly about what you want before asking remains the most valuable skill in the age of AI.

At Qolaig, our Workshop → Build → Run methodology frames this type of project before any technical investment. Let's talk about your project.