I Discovered Requirements Before Building

The Old Way

My old way was simple: get an idea, open the AI, type a prompt, start building. No requirements. No planning. No breaking the task apart. Just the idea, the model, and the hope that something good would come out.

Sometimes it worked. Most of the time it did not. When it failed, I spent more time fixing and re-prompting than I would have spent writing requirements from the start.

The Discovery

I did not learn about requirements from a book. I learned from failing over and over. After enough broken outputs and wasted hours, a pattern became clear. Every good prompt had one thing in common: I knew what I needed before I asked. Every bad prompt had one thing in common: I was figuring out what I needed while I was asking for it.

Finding requirements and generating output are two separate jobs. They use different kinds of thinking. They need different information. Trying to do both at the same time is like designing a building while you are building it. You end up with something that shows your changing mind, not a solid plan.

Requirements First

Now I do requirements first. Before I write one prompt, I answer these questions: What roles does this task need. What comes first, second, third. What goes in. What comes out. What rules apply. What format does the output need. What unusual cases might come up.

I write them down. Not in my head. On paper or in a file. Writing forces me to be exact. It shows holes in my thinking before those holes show up in my prompts.

The Impact

With requirements ready, the prompts almost write themselves. Each prompt is a direct translation of one requirement into a model instruction. There is no guessing. No hoping. No going back and forth. The requirement says what I need. The prompt tells the model what to do. The output matches because nothing is missing in the chain from requirement to prompt to output.

I have to know what the project needs before I build it. This was always true in software engineering. It is just as true in AI prompting. The tool changed. The principle did not.