Most advice written for people starting with AI is produced to sell something. A course, a tool subscription, a newsletter. That shapes the advice in a specific direction: it makes the subject sound larger and more urgent than it is, because a small, finishable task is harder to sell than a transformation.
We sell courses too, so treat that as a disclosure. This page is the version I would give someone who asked me at a dinner, and it is deliberately short, because the honest answer is smaller than the genre suggests.
Your first week
Five things. They take a few hours in total, spread across the week.
1. Pick one tool and stop comparing
ChatGPT, Claude or Gemini. At this stage the differences between them matter far less than the hours you put in. Comparison shopping is a way of feeling productive while deciding nothing, and the time is better spent using whichever one you already have access to.
2. Use it on work you already know how to judge
This is the most important item on the page, and it is the one most guides skip.
Start with tasks where you can tell immediately whether the output is good: an email in your own area, a summary of a document you have read, a first pass at something you have written a hundred times. Your expertise is what makes the tool useful, because it lets you catch what is wrong. Starting in a domain you know nothing about teaches you to trust fluent writing, which is the opposite of the habit you want.
3. Do the same task three times, three ways
Take one real task. Ask once, casually. Ask again with the context you would have given a new colleague. Ask a third time with an example of what good looks like.
The difference between those three outputs teaches you more about working with these tools than any amount of reading about prompting. It also shows you where the effort actually pays, which is in supplying context and standards.
4. Catch it being wrong, on purpose
Ask it something you know cold. A detail from your own field, a figure you can verify, a process you run. Push until it produces something confidently incorrect, which will take less time than you expect.
That moment is the most valuable half hour of your first week. Everything it produces arrives in the same assured tone whether the model is certain or guessing, and seeing that once, in your own domain, gives you a calibration that no warning paragraph can.
5. Write down what it did badly
Keep a note through the week. Three columns is enough: the task, what went wrong, what you did about it. By Friday that note is a more useful guide to your own work than anything you could have read on Monday, and it tells you which parts of your job are actually worth delegating.
What to ignore for six months
The genre will push all of this at you early. It can wait, and waiting costs you nothing.
- Prompt engineering courses. Models got considerably better at interpreting ordinary requests. Being clear about what you want still matters, and that is thinking, which you practise by doing the work.
- Tool roundups and "top 50" lists. Most of those products will be repriced, acquired or shut down within the year. Learn the craft on one tool and the rest transfer in an afternoon.
- Building agents. Genuinely interesting and genuinely premature. Agents are worth building once you can describe a process precisely, and describing processes precisely is a skill you do not have yet if you are in week one.
- Running models on your own machine. A real topic with real reasons behind it, none of which apply to someone learning the basics.
- The discourse. Whether this is a bubble, whether it will take your job, whether it is intelligent. Interesting to argue about, useless for deciding what to do on Monday.
A reasonable rule: if the advice would make sense to someone who has never done the work, it is probably written for engagement.
How to tell whether it is working
Three signals, in order of how much they tell you.
You catch it being wrong more often than you did a month ago. This is the one that matters. Rising correction rate means your judgment is engaging, and people who never catch anything are either working below their expertise or approving without reading.
You have started declining to use it. Choosing to write something yourself, because the checking would cost more than the drafting, is a sign of calibration. Reaching for it every time is a habit wearing the costume of fluency.
A specific task is measurably faster. Named, with a before and after. "It saves me time" is the thing people say when nothing in particular has changed.
Notice that none of these is about how many tools you have tried.
What to learn next
Once the first weeks are behind you, three things carry most of the value. Each has its own piece here.
- Judging work you did not produce. The daily reality quite quickly becomes reviewing output rather than making it, and that is a distinct skill with its own technique. The AI skills that actually separate people covers it, along with the three others that held their value.
- Checking what a tool gives you. Research tools return confident reports with real citations that do not always support the claims attached to them. Best AI for deep research has a ten-minute pass for exactly this, plus which tool suits which job.
- Describing a process well enough to automate it. The point where this stops being a personal productivity tool and starts changing how work gets done. Automating one workflow end to end walks a real one from trigger to measure.
That order is deliberate. Judgment first, verification second, automation third. Reversing it is the single most common way these projects go wrong.
Where people get stuck
Waiting to feel ready. There is no threshold of understanding to cross before starting. The understanding comes from the hours, and the first ones are supposed to be clumsy.
Starting with the most impressive use case. People reach for something ambitious, cannot evaluate the output, and conclude the tool is either magic or useless. Both conclusions come from the same mistake, which is starting outside your own expertise.
Treating a bad result as a verdict. One disappointing attempt usually means the request lacked context, and the natural response is to conclude the tool cannot do it. Give it what you would give a new colleague and try again before deciding.
Collecting instead of practising. Twenty bookmarked tools and four hours of actual use is a common and comfortable place to be, and it feels like progress. Depth beats breadth here by a wide margin.
Common questions
How do I start with AI if I am not technical?
Pick one assistant, use it on work you can already judge, and run the same real task three ways to see what changes the output. Deliberately catch it being wrong in your own field during the first week. No technical background is needed for any of this, and the expertise you already have is what makes the tool useful.
Do I need to learn prompt engineering?
Not as a course. Models improved substantially at interpreting ordinary requests, so the gap between an engineered prompt and a clear one narrowed. Supplying context and an example of what good looks like will cover most of the benefit, and you learn that by doing the work.
Which AI tool should a beginner use?
ChatGPT, Claude or Gemini, and the choice matters less than the hours. Pick whichever you have access to and stay with it long enough to learn where it fails. You can evaluate alternatives sensibly once you know what you are comparing.
How long does it take to get good with AI?
A few weeks for the habits. Longer for the judgment, because that comes from being wrong repeatedly and noticing. A useful marker of progress is catching the tool's mistakes more often than you did a month ago.
What should I not do when starting with AI?
Avoid starting in a domain you cannot evaluate, which teaches you to trust fluent writing. Skip tool roundups, agent building and local models for the first six months. And avoid treating one poor result as proof the tool cannot do something, since it usually means the request lacked context.
Is AI going to replace my job?
The more useful question is which parts of your work can be written down completely, because that is the exposed part. Work that needs judging quality, holding context nobody wrote down, or deciding what should happen is considerably safer, and those are the capabilities worth building.
How do I know if it is actually helping?
Name one task that is measurably faster, with a before and after. Check whether you catch errors more often than you used to. And notice whether you sometimes decide against using it, which is a sign of calibration rather than reluctance.
What I would do
Give it four hours this week, on work you already know how to judge, and spend thirty minutes of that deliberately pushing it until it is confidently wrong. Keep the note of what went badly. That is the entire first step, and it is smaller than the genre wants you to believe.
After that, the order is judgment, then verification, then automation. Our curriculum lays that ladder out if you want a structured route through it, and the bootcamp is the version where you build something real with your own work as the project. Neither is necessary for the first week. The first week only needs you to start.
Charafeddine Mouzouni
