You have used a chatbot a few times and want to actually get good at this.
Reviewed
Four weeks, about an hour a day
By the end you will have automated at least one recurring task, know why the model fails when it fails, and have shipped one small thing other people can use.
Most people who bounce off AI do it in the same way: they try a few impressive demos, get an underwhelming result on their own work, and conclude the tools are overhyped. The gap is almost never the model. It is that nobody told them what the thing is actually doing, so they cannot tell a badly phrased request from a genuine limitation.
This path fixes that in the order that keeps you motivated: use it first, understand it second, build third. Every week ends in something you can point at.
Pick three tasks you already do every week and that you find tedious: drafting the same kind of message, summarising notes, cleaning up a spreadsheet, rewriting something for a different audience. Do them with a model every time they come up, for a week. Do not pick anything clever — the point is volume on work you can judge, because you already know what good looks like.
Keep a note of every time the output was wrong or useless, in one line. That note is the single most valuable thing you will produce this month; it becomes the syllabus for week 2 and the test set for week 4.
Go back to your list. Almost every entry will fall into one of four buckets, and each has a name. It invented a fact — that is hallucination. It forgot what you said earlier — that is the context window. It miscounted or mangled a number — that is tokenisation showing through. It gave you a vague answer — that is an under-specified prompt, which is the one you control most directly.
Read the glossary entries for the failures you actually hit rather than all of them. Then rerun the same tasks now that you know what went wrong. The improvement in week 2 over week 1 is usually larger than anything a better model would have given you.
A prompt you rewrite from scratch each time is a conversation. A prompt you save, reuse and improve is a tool. Take the three tasks from week 1 and write a proper instruction for each: what the task is, who the output is for, the format you want, and two examples of a good answer. Save them somewhere you can find them.
Then use them unchanged for a week and edit them only when they fail. You will notice the edits getting rarer, which is what a working tool feels like. This is the whole of prompt engineering in practice — specificity, examples, format, iteration.
Pick the prompt that saved you the most time and make it usable by someone who is not you. That can be as simple as a shared document explaining what to paste in and what comes out, or a small no-code automation that runs it on a schedule. It does not need to be software.
Then give it to one other person and watch them use it without helping. Everything you learn in that ten minutes — the input they gave that you never anticipated, the output format that confused them — is the difference between something that works for you and something that works. Fix one of those things and stop.