What next?
Listen to this article
Read by Anchor
Before we continue... where did we leave off?
In the first article, we talked about the limits of using AI, and about keeping responsibility and decision-making in human hands.
In the second article, we went a step further and asked how a model can influence a decision even without making it outright -- through what it selects, what it summarises, and what it places at the top of our priorities.
In both cases the idea was the same: using AI does not start from trusting it or fearing it, but fromunderstanding its role.
But if it is going to become part of our daily work, there is a simpler and more practical question:
Do we actually know how to talk to it?
Because the quality of what we get out of it often starts with the quality of what we put in.
AI has become the language of the age -- not just another piece of software
I remember when fluency in English, or any foreign language, was a genuine advantage in the job market.
It was never just about knowing more words.
You had to know how to express your idea, how to ask, how to explain the problem, and how to choose the right term for the context.
I think we are entering a similar phase today with AI.
I do not mean it is a "language" in the literal sense, but it has become "a new way of engaging with knowledge and work".
Almost anyone can open ChatGPT or a similar tool and type a question.
But owning the tool does not necessarily mean using it well, just as owning an English dictionary does not mean you speak English.
There is a significant difference between someone who "uses AI" and someone who knows "how to have a real conversation with it to get what they actually want".
And I believe this skill will, over time, become a natural part of professional competencies -- like writing a business email, preparing a presentation, or explaining a task to a team.
Before we talk to it... what exactly are we talking to?
When we work with generative AI tools, we are mostly working with what is known as a "large language model," or what is called an L
"Large Language Model – LLM".
In very simple terms, it is a model capable of processing the language and information we give it, then producing text, summarizing, comparing, classifying, organizing, and helping with analysis.
But there is something very important we need to remember:
The model cannot see the image inside your head.
It does not know why you are asking.
It does not know your project.
It does not know your client.
And it does not know what you consider important... unless you tell it.
This is where three core concepts come in:
Input:The information we give the model.
Context:The background that allows it to understand the meaning of that information and why we are using it.
Prompt:What we are asking it to do with that information.
And what we get at the end is:
Output.
So the equation is not:
Good Prompt = Good Answer.
The matter is slightly broader:
Relevant information + clear context + specific request + clear boundaries = output that is closer to what you actually need.
"Review this report"... let's start there
Suppose I have an inspection report and I want to use AI to help me review it.
I write:
"Review this report."
What do I expect to happen?
The model will try to interpret the word "review" in its own way.
It might correct the language.
It might summarize the report.
It might comment on the content.
Or it might look for errors it considers significant.
Then I look at the result and say:
This is not the review I wanted.
But the truth is, I never told itwhat review I wanted..
So we start by improving the prompt.
> Act as a QA/QC reviewer and review the attached inspection report.
Better.
Now it has a role.
But there are still questions.
Do I want it to review the language?
Or the completeness of the data?
Or technical compliance?
So we add the task:
> Act as a QA/QC reviewer. Review the attached inspection report to verify data completeness and internal consistency.
Now the task is much clearer.
But we can go one step further.
Context... is the part we usually forget.
Let us say this report is going out to the client, and what is needed right now is a Document Review only, not an accept-or-reject decision.
That is information I have.
But the model does not.
So I add:
> This report is intended for final issue to the client. What is required is a review of data completeness and internal consistency only, not a technical accept-or-reject decision.
Here I did something important.
I did not just give it more information.
I transferred to ita piece of the picture that exists in my head.
and that is the context.
And in my view, context is sometimes more important than the wording of the Prompt itself.
Because the request:
> Summarize this document.
is completely different from:
> Summarize this document for a non-technical manager who needs to understand the main risks before a client meeting in one hour.
The same document.
Butthe context changed what was needed entirely.
Now give it the reference
We go back to our report.
I also have a Procedure and a project specification.
So instead of letting the model rely on its general knowledge, I tell it:
> Use only the attached report, procedure, and specification as sources for review.
This is where the inputs got better.
I no longer just tell it:
"Know the answer."
Instead, I say:
Here is the information I want you to work within.
And that is a very important distinction.
If I want to review a contract, I give it the contract.
If I want to analyze a policy, I give it the approved policy.
If I want to compare two versions, I give it both versions.
And if I want to analyze actual data, I give it the data instead of asking it to imagine it.
The AI does not know there is a new Revision sitting in your inbox.
It does not know the client changed the Requirements yesterday.
And it cannot read a file you never gave it.
Sometimes the problem is not that the AI did not understand.
The problem is that we assumed it knew something we never told it in the first place.
Then set the boundaries.
Something is still missing.
I do not want the model to see incomplete information and then invent what it thinks is logical.
So I added:
> Do not assume any information that is not present, and do not add acceptance criteria or clause numbers from your general knowledge. If any information is unavailable, flag it explicitly as requiring human review.
Here I did not only say what I wanted.
I also saidwhat I did not want.
And this is one of the most important skills in working with AI.
Because the model, if left with a gap, may try to be helpful and fill it in.
But "helpfulness" is not always what we need.
Sometimes the best answer the model can give is:
The information is not there.
And now... define the output format
We have almost reached the complete request.
But I can still define the shape of the result:
> Present the findings in a four-column table:
> The finding - its location in the report - the reference linked to it - the required action.
> Classify each finding as: missing data, conflicting information, or a point requiring human verification.
Now compare that with where we started:
> Review this report.
And the conclusion:
> Act as a QA/QC reviewer. The attached report is intended for the client's final release. Review it for data completeness and information consistency only, not to make an acceptance or rejection decision. Use only the attached report, procedure, and specification. Do not assume information, acceptance criteria, or item numbers that are not present. Place any uncertain point under the item "Requires Human Verification". Present findings in a table showing the observation, its location, the related reference, and the required action.
The difference between the two is not that the second is a "genius Prompt".
The difference is that Iknew what I wanted, then explained it clearly.
And that, to me, is where the real skill of the future lies.
This is not only about QA/QC
The same logic can be applied to any field.
An HR manager, instead of saying:
> Review these CVs.
Can define the role, its requirements, what they want compared, and what they do not want the model to infer.
And customer service, instead of:
> Analyse the complaints.
Can specify the time period, the classification method, the desired output, and whether they want root causes or just recurring patterns.
And the CFO, instead of:
> Analyse the numbers.
Defines the data, the period, the indicators that matter, and the format in which they want to see the result.
The tool is the same, but the quality of its use varies drastically depending on the user's ability toTranslating what is in his head into clear context and instructions.
And sometimes... let it ask you
There is a method I use and find very useful, especially when the task is complex.
Instead of assuming I have given the model everything it needs, I say:
> Before you begin, ask me about any missing information you need to complete the task better.
Suddenly the relationship shifts from question and answer toa dialogue.
It might ask me:
What is the approved version?
Who is the target audience?
Do you want a technical review or a language review?
What form should the output take?
And sometimes the value of the question lies in the fact that it makes me realise that what was needed was not clear in my own head to begin with.
And here AI becomes more than a tool for answering.
It becomes a tool that helps youarticulate the problem itself more clearly.
Do we all need to become Prompt Engineers?
In my opinion, no.
I don't think every employee will need the title "Prompt Engineer."
We don't call everyone who writes a good email an "Email Engineer."
What we need is a simpler and more important skill:
clarity of thought before clarity of request.
Before you talk to AI, ask yourself:
What exactly do I want from it?
What information does it need?
What context does it not have?
What source do I want it to rely on?
What boundaries do I not want it to cross?
And in what form will I use the output?
If you know the answers, you are already talking to AI in a better language.
What this means for you
I believe that working with artificial intelligence will become one of the new languages of work.
Not because it has words and grammar rules to memorize.
But because some people will know how to express what they need precisely, while others will keep writing a vague question and then wait for a perfect result.
The tools will be available to almost everyone.
But the real difference will not lie only inWhoever owns AI.
will bewhoever knows how to use context, information, and instructions to turn it into a real working tool.
So perhaps the next question will not be:
Do you know how to use artificial intelligence?
But rather:
Do you know what you want from it in the first place... and can you transfer the picture in your head to it?
Because the model may be very intelligent.
But there is still one piece of information it cannot reach on its own:
What you actually meant.
......to be continued