How to build a good app with AI

An AI assistant will happily try to build whatever you ask for — so it’s tempting to write one enormous prompt describing every screen, field, and feature, hit enter, and hope. That’s the surest way to get a tangled app that half-works and is hard to fix.

The better way is the way good software has always been built: start small, check it, then grow it. Here’s the flow that reliably gets you a solid app — whether you’re chatting with Claude, ChatGPT, or Cursor over MCP, or using Croft Build.

1. Start with the smallest useful version

Ask for the core of the app and nothing else. Not “build me a CRM with pipelines, email, reporting, and permissions” — just “build me an app where I can add a customer with a name, email, and a note, and see the list.”

One screen, one job. A small first version is quick to build, easy to understand, and — crucially — easy to tell whether it’s right. You’re establishing the foundation before you pile anything on top.

2. Open it and actually use it

Don’t judge the app from the chat transcript — open it and use it like you would for real. Add a couple of records. Click the buttons. Try the thing you actually need it to do.

You’ll spot the gaps immediately: a field that’s missing, an order that’s wrong, a button that should be somewhere else. This is the step people skip, and it’s the most valuable one.

3. Fix it by describing what’s wrong

When something’s off, don’t start over — just say what you want changed, in plain words. “Move the note to its own column.” “Sort the list by most recent.” “The email should be required.” The assistant changes that one thing and redeploys.

Because you kept the first version small, each fix is small too. You’re never debugging a giant black box — you’re nudging a simple thing into shape.

4. Add one thing at a time

Now grow it, one change per step, checking after each:

Add a second view — “let me filter to just this month.”
Add data — connect Stripe or your CRM so the app can pull real numbers, with the key held safely by the platform, not pasted into the code.
Add a schedule — “email me this list every Monday at 9am.” Now it’s an automation, not just an app.
Add people — invite a teammate and choose what they can see.

Each addition is easy to verify precisely because everything before it already works. If a change makes things worse, you’ll know at once — and you can undo it.

5. Ship it when it’s right — and keep going

On Croft the app has been real the whole time: a live URL, a private server, a database, a login. So “shipping” is just inviting the people who should use it. Then keep iterating — apps grow the way your needs grow.

Two things make this safe to do:

Every deploy is versioned and reversible. A change from the AI that isn’t right is a one-click rollback, not a crisis. That’s what lets you experiment freely.
You’re not locked to one AI. Build the first version in ChatGPT today, refine it with Claude next month — the app is provider-agnostic and lives on a server that’s yours.

The short version

Resist the urge to specify everything at once. Build the smallest thing that’s useful, open it and use it, describe what to fix, then add one capability at a time. It’s slower for the first five minutes and much faster for everything after — and you end up with an app you actually understand.

Start with a prompt → · See how it works →