When I tell people I build apps by talking to an AI, the first question is usually: "But how do you tell it what you want?" It's a fair question. And it turns out the answer is more interesting than "just type stuff."
Over months of building — SpekPlatz, KeepAwake Pro, Music Mouse, and more — I've developed a way of thinking about prompting that makes a real difference in what comes back. It's not magic. It's just clarity.
Start with the feeling, not the feature list
The worst prompts I've written started with a list of technical requirements. "Build a web app with a sidebar, a map, user authentication, and a search bar." That's a spec, not a vision. What comes back tends to be technically correct but soulless.
The best prompts start with: "I want something that feels like..." or "The experience should be..." When I described SpekPlatz, I didn't start with the tech stack. I started with: "I want something between a note app and Google Maps — personal, curated, the feeling of a handwritten travel journal but on your phone." That orientation shapes everything that follows.
"Good prompting is more like being a film director than a software engineer. You're describing the scene, not writing the script."
Be specific about what you don't want
One of the most useful things I've learned: tell Claude what to avoid. "No onboarding screens," "don't use modal dialogs for this," "I want it to feel minimal — not cluttered with options." Constraints are generative. They focus the output and save rounds of "can you remove that thing."
I also find it helpful to reference things that already exist. "Think Notion but without the complexity." "Like iOS, where everything just works." "Like Craigslist, but designed by someone who cares." References carry a lot of information efficiently.
Break it into conversations, not one big prompt
The biggest mistake I see people make is trying to describe the entire app upfront. In one go. Everything they've ever wanted.
That rarely works well. Not because Claude can't handle complexity, but because you don't actually know exactly what you want until you start seeing options. A better approach: describe the core, see what you get, then refine from there.
SpekPlatz started as "a map app where I can save places." Three conversations later it had collections, community features, and a specific visual language. I couldn't have described all of that upfront — I discovered it through the process.
Feedback is a superpower
When something isn't quite right, don't scrap it and start over. Describe precisely what's wrong and what you'd prefer instead. "The card is too dense — I want more breathing room between the image and the text." "The color feels too cold — warmer, please." "The button is correct but it feels generic — make it more distinctive."
Specific, directional feedback — like giving notes to a designer — is far more effective than "I don't like this, try again."
Trust the expertise you don't have
Sometimes Claude will suggest something I hadn't considered. A different interaction pattern. A more elegant solution. An approach I wouldn't have known to ask for. The best thing I've learned to do: say yes, see how it feels, then decide.
Working with AI well means knowing when to steer and when to let it drive. It knows things about web development, about design patterns, about accessibility, that I simply don't. The collaboration works better when I'm open to being surprised.
The apps I'm most proud of have that quality — they're mine in vision, but better in execution than I could have specified in advance.
💬 Discussion
No account needed — just enter your name and comment