How to build an app with AI

What actually changed

There used to be a long stretch between the idea and the first screen: pick a stack, set up a project, create a database, wire the screens together. That stretch shrank to a paragraph of description. It is the biggest practical change, and it is real.

What did not shrink is everything else: deciding what the app does, finding out what is missing once somebody uses it, and handling anything involving money and permissions.

A chat assistant is not an app builder

This is the confusion that costs the most time. A chat assistant hands back code. A builder hands back a running app.

Between the two sits all of this: where the app runs, where the data lives, how somebody logs in, how it gets to an address, how you change one screen next week without pasting files anywhere. Asking an assistant for code and having nowhere to put it is where most attempts end.

The loop that works

1. describe it in one paragraph: who uses it, what they do, what gets saved;
2. open it and click. The first version comes out navigable, and what is wrong shows up faster by using it than by reading it;
3. fix on the spot, one thing at a time: "this list needs a date filter", "only the account owner can delete";
4. publish early and send it to the people who will use it.

Asking for too much at once is the common mistake. A targeted correction lands better than a long specification, because the AI does better when it can see the screen that is wrong. How to write that request is in how to write an app prompt.

Where AI still gets it wrong

• who sees what. It generates the screen; the rule that one customer cannot see another's order has to be stated, and it has to be checked;
• the sad path. Missing data, a dropped connection, the same thing clicked twice;
• money. Amounts, tax, refunds and what happens while a payment is pending;
• what it does not ask. If the description did not say, it invents something reasonable, and reasonable is not your process.

The check is short and always worth it: sign in as an ordinary user and try to do what that user should not be able to do.

What to verify before investing weeks

• the app publishes to an address of its own instead of living in a preview;
• you can export the code;
• the data is yours and can be taken out;
• fixes happen through targeted conversation, not by regenerating everything each time.

The summary

• AI shortened the beginning, not the product;
• an assistant returns code, a builder returns a running app: the difference is where it lives;
• describe briefly, open, fix on the spot, publish early;
• always check permissions, errors and money, which is where it errs most;
• before investing weeks, confirm publishing and export.