Testing and improving prompts instead of guessing

Most people write a prompt, are unhappy with the result, rewrite the prompt from scratch, are unhappy again, and at some point give up frustrated. That's guessing, not improving. The difference sounds small but makes all the difference. Whoever guesses goes in circles. Whoever tests gets closer to the goal with every round.

In this lesson you'll learn a loop you can apply to any prompt. It's not complicated, but it demands a bit of discipline: change only one thing at a time and see what happens. That's exactly what the fewest people do, and that's exactly why it works.

Why the first prompt almost never lands

You have a precise picture in your head of what you want. The AI doesn't have that picture. It only has your words. Everything you carry along in your head but don't write down is missing for it. That's why the first prompt is usually too vague, leaves too much open, or assumes knowledge the AI can't pull from your text.

That's completely normal and no sign that you can't do it. Even people who work with AI daily rarely write the perfect prompt on the first try. The only difference is that they don't give up but improve systematically. They treat the first attempt as a rough draft, not as a finished product.

The test loop in four steps

The core is a small round that you repeat until the result fits. Four steps, always the same.

First, formulate a clear goal before you write the prompt. Not "I want a good text", but "I want an ad text for my carpentry shop, three sentences, friendly but not fawning, emphasizing handcraft". The more precise your goal, the better you can judge afterwards whether the result is right. Without a clear goal you don't even know what to improve.

Second, write the prompt and look at the result closely. Not just "good" or "bad", but concretely: What's missing? What's too much? Where does it miss the tone? Write down the one or two biggest problems. That's the most important step, and the one most people skip.

Third, change exactly one thing. Don't throw out the whole prompt, but tackle the biggest weak point specifically. Was the tone wrong, add a tone instruction. Was it too long, give a length specification. Was the content off, sharpen the brief. Only one change per round, so you know what worked.

Fourth, compare the new result with the old one. Better? Keep the change. Worse? Throw it out and try something else. Then start over until it fits. Usually you're there after three to five rounds.

Why only one change at a time

This is the rule people fail at most often. You're unhappy, so you change the tone, the length and the content all at once. The new result is different, but you have no idea which of the three changes helped and which hurt. You learn nothing from it.

If instead you turn only one screw per round, you see the effect immediately. You build up a feel for which words do what. After a few weeks you intuitively know that "in three sentences" tames the length and "explained for a layperson" takes out the jargon. You only accumulate this knowledge if you test cleanly instead of throwing everything out at once.

It feels slower to change one thing after another. In fact you reach the goal faster, because you don't run in circles.

A concrete example

Let's say you want to write a rejection to an applicant. First prompt: "Write a rejection." The result is cold and generic. Problem identified: too impersonal.

Round two, one change: "Write a friendly, appreciative rejection." Better in tone, but too long and vague. New problem: length.

Round three, another change: "...in four short sentences." Now concise and warm, but the concrete reference is missing. Last problem: no reference to the person.

Round four: "...mention that we liked the project experience, but another candidate was closer to the profile." Now it lands. Four rounds, each with a clear step, and along the way you've learned what details make a good rejection.

Save what works

When a prompt lands after a few rounds, don't throw it away. Set up a simple collection, a note will do, with your best prompts for recurring tasks. Rejections, quotes, social media posts, whatever you need often. Next time you don't start from zero but from your proven template and only adjust the specifics.

That's the point where AI really saves time. Not in one-off tinkering, but when you build up a small library of proven prompts that you use again and again.

What you take away now

A good prompt isn't a flash of inspiration but the result of a few clean test rounds. Formulate the goal clearly, assess the result honestly, change one thing, compare. Repeat until it fits. And save what works for next time.

This switch from guessing to testing helped me most when I realized that bad results almost never come from the AI but from my prompt. Once you see that as your lever rather than a matter of luck, you have control. In the capstone of this level you bring it all together and build your first really well-thought-out prompt for a real task.