A client dashes off a message: "I'd like a kind of Tinder for dogs." In two weeks they're meeting an investor and want to show them something they can touch, not slides.
There are no screens in there, no flow, no style. The obvious move — pasting it into Claude Code and asking for the app — doesn't work, and not because the agent is bad at its job: it executes decisions, it doesn't make them. Whatever you don't tell it, it invents, quietly.
So you make the decisions first, and write them down for it: the screens, the flow, the style, the empty states. Then they become a single file, you hand it over, and the app comes out of it.
This is what comes out of it
The result before the method: PawSoul, ten screens, fake but consistent data.
The briefing that produced it is six hundred lines long, and no one wrote it by hand: Claude Code wrote it, reading two documents decided beforehand. You'll write those documents in the parts that follow, for your own project.
All you need is Claude Code and an empty folder.
FAQ
Do I need to know how to code to follow this guide?
No, but you need to be able to read what comes out. You don't write the components yourself — you have Claude Code generate them. What you do need is to recognize when the result doesn't match the brief, and that's a matter of eye, not syntax.
Why so much preparation before generating the code?
Because an agent doesn't guess the things a client left unsaid — it makes them up. Every decision you make beforehand — flow, screens, palette, components — is one less thing invented at random that you'll have to redo later.
How much time does it really take?
Reading it takes a bit over half an hour. Redoing the process on your own project — brief, document, choices, generation, and the review pass — takes roughly two and a half hours, about half of which goes into preparing the briefing.
