How to Write Prompts for AI App Builders That Work

A good prompt for an AI app builder asks for one feature and says who uses it, what they see, what data is saved, who may see that data, and what "done" looks like. Prompts like that work in Lovable, Bolt, Replit, v0, Cursor and Claude Code alike. This guide gives you the formula, examples, and the mistakes that cost the most credits.

The formula

Every good build prompt answers five questions:

1. Who uses this feature?
2. What can they do, and what do they see on the screen?
3. What data is saved?
4. Who may see or change that data?
5. What does "done" look like, so you can check it?

Optional, but helpful: what the screen shows when there's no data yet, and what happens when something goes wrong.

Before and after

Before:

Add bookings to my app.

The builder has to guess who books, how, where bookings show up, and what's allowed. It will build something, but probably not what you had in mind.

After:

Owners can book a walk. On a walker's profile page, the owner picks a date and one of the walker's free time slots, then clicks "Book". Save the booking with the owner, the walker, the date, the time and the status "unpaid". Owners see only their own bookings on a "My bookings" page; walkers see bookings made with them. If the walker has no free slots, show "No free times this week". Done when an owner can book a slot, sees it under "My bookings", and the slot is no longer shown as free.

It's longer, but the builder doesn't have to guess anything, so you'll fix less afterwards.

Seven rules that save the most time

1. One feature per prompt. Two features in one prompt means twice the guessing, and mixed-up mistakes.
2. Build in order. Ask for what other features depend on first, like accounts before bookings.
3. Say who may see what. It's the detail beginners forget most, and the one that matters most for privacy.
4. Describe the empty state. What does a new user see before there's any data?
5. Define "done". One sentence you can check by clicking through the app.
6. Fix with specifics. "The booking button does nothing when I click it on my phone" beats "it's broken".
7. Don't restart, refine. Starting over throws away what already works.

Differences between builders

The formula is the same everywhere, but each tool likes a slightly different style:

• Lovable and Bolt (browser builders): describe screens and actions in plain words. Skip file names and terminal commands.
• Replit: same as above, plus ask the agent to test the feature in the preview when it's done.
• v0: focus on layout, states (empty, loading, error) and what each button does.
• Cursor: mention which parts of the project the change touches, and ask for small changes you can review.
• Claude Code: give the whole feature with clear acceptance criteria; it can read your project and run checks itself.

Let Flowmaps write the prompts

Writing a prompt like the "after" example for every feature takes time. Flowmaps writes them for you: you plan your app on a visual map, pick your builder, and each feature gets a ready-to-paste build step in the style that builder understands best, in the right order. See the example maps, or read how build prompts work in Flowmaps.

Frequently asked questions

How long should a prompt for an AI app builder be?

Usually one paragraph per feature: long enough to answer who, what, which data, who may see it and what "done" means.

Should I write prompts in English?

Most builders understand many languages, but English often works most reliably. The words on your app's screens can be in any language; just say so in the prompt.

Why does the AI builder ignore part of my prompt?

Usually because the prompt asks for several things at once. Split it into one feature per prompt and try again.

Can I reuse prompts between builders?

Mostly, yes. The five questions stay the same; adjust the style a little for each tool, as described above.