Don’t prompt yet.

We’ve been exploring the latest AI prototyping and code generation tools like Cursor, V0, and others. And something keeps coming up. Many teams jump in, start typing, and hope to muddle through.

But if you haven’t structured your thinking first, the results are usually disappointing.

And for those who don’t come from a strong engineering background, it can feel even harder. Some experience actual anxiety; fearing failure. 😱

You know what you want to build, but not how to start. Or how to talk to the model. And that’s where things break down.

Through trial and error, here’s the approach I’ve found most helpful:
Work with AI before you ask it to generate anything.
Start by describing your idea to AI, and get it to ask you questions to clarify and deepen your thinking.

Use those questions to build out a more detailed picture: your functional requirements, user types, use cases, and constraints.
Review and refine that dialogue until it naturally forms the bones of a spec or PRD that you can be proud of.

Then, ask AI to compile this into a clear and structured specification. This is not as simple as it sounds, so it might take some good old copy and paste.

From that, ask AI to create a high-level development plan: frameworks, architecture, tooling, stages.

And only then—ask it to write targeted prompts to generate code for each specific chunk.

It’s a process of co-development.

You’re tuning the model as you go, by giving it more and more context.
And in return, you get more accurate, more useful output.

The speed of AI is real.

But the quality comes from the craft of defining, structuring, and curating the problems we’re trying to solve.

It reminds me of the classic 1980s film Short Circuit. Johnny 5 had an insatiable appetite for input and contextualization, but without Stephanie's direction, he quickly lost the plot (and destroyed the house).
Just like our favourite AI models!