Learn to Ask Before You Build

After my first days with AI, I thought I needed better prompts. Those long templates you find online, written like small contracts with the model. Role, goal, context, format, constraints, examples. They can help. But they are not the beginning.

The beginning is simpler: I had to learn to ask better questions.

A bad question is not stupid. It is usually just too large. "Explain LLMs to me" is large. "Explain LLMs so I understand the difference between ChatGPT, a local model, and an adapter" is better. "Make software for me" is large. "Show me how a small web app is split into three parts: interface, logic, and data" is better.

LLMs react strongly to context. That sounds technical, but it is easy to understand in human terms. If I tell a person "make this better", I get whatever they think better means. If I explain who it is for, what I want to avoid, what is fixed, and what is still open, the answer improves. AI is similar.

Over time I got used to a small question structure:

• What am I trying to achieve?
• What do I already know?
• What do I not understand yet?
• How will I know whether the answer is good?

The last point is the important one. Without a check, almost every AI answer can sound good. With a check, you quickly see whether it helps. For text, that might mean: understandable, but not salesy. For code: it runs, it is tested, and it contains no sensitive data. For a concept: it shows the first three steps, not the whole universe.

For non-technical readers, this is where AI becomes practical. You do not need to ask like an engineer. You can ask like a person who wants to understand. "Use an example from office work." "Use an example from cooking." "Use an example from a warehouse." "Do not skip steps." These are good prompts because they give the model a route into your world.

I also learned to make the model push back. Not just "do this." Instead: "What assumptions are you making?" "What could be wrong here?" "What information would you need to be more certain?" "Give me the simple version first, then the technical one." That makes the answers less smooth and often much more useful.

The most useful AI conversations often have a rhythm:

1. Ask the first imperfect question.
2. Read the answer and mark what you do not understand.
3. Ask the model to explain only that part.
4. Ask for a concrete example.
5. Ask how to verify the claim.

This rhythm is slower than copy-pasting a big prompt from the internet. It is also much better for learning. You are not trying to trick the model into a perfect response. You are using it to build your own understanding.

For software development, another layer appeared. I do not only ask for a solution. I ask for a working path. What should I read? Which file matters? What change is small enough? How do I test it? What could break? These questions do not turn AI into magic, but they do make it a usable companion.

In normal words

Context means the background information the model sees before it answers. More useful context usually gives a better answer.

Verification means checking whether an answer is actually true or useful. In software, that can be a test. In writing, it can be reading the text as the target audience.

Hallucination means the model gives an answer that sounds confident but is wrong or invented. It is a normal risk in AI work, which is why checking matters.

At some point something shifts. You stop using AI only to get answers. You use it to organize your own thinking. For me, that is still one of the biggest levers.