Plan Before You Prompt: I Built the Same App Twice in Lovable, Once With a Miro Plan
The most expensive habit in vibe coding is opening the builder before you know what you want.
Every AI app builder I’ve used charges for each round of changes. When the first version misunderstands you, you pay to explain again. When the fix breaks something else, you pay again.
Lovable has added some protection since I last wrote about its costs. Paid plans now let you set credit limits for a whole workspace or a single project, and long tasks pause to check in once they’ve used 20 credits. That’s welcome. But a limit only tells you when you’ve run out. It doesn’t stop the credits going on fixes you didn’t need.
So I ran a simple experiment. I built the same small app twice in Lovable:
- Build A: a one-line prompt, then fixing things as they came up, the way most people start.
- Build B: a few minutes mapping the app in Miro first, then one structured prompt built from that map.
Same app, same tool, same goal. The only difference was the planning.
What was builtA booking request app for a dog walker: clients pick a service and time, the walker sees requests and approves or declines them
Tools usedMiro (free plan), Lovable
Time takenAbout 11 minutes and 3 prompts without a plan. About 10 minutes and 1 prompt with one, including 7 minutes in Miro.
Approximate cost4.83 Lovable credits without a plan, 2.30 with one. Both on Lovable's free plan. Miro free plan.
Difficulty for a non-developer
No code. The Miro step needs no design skills, only boxes and arrows.
The app
A dog walker needs a simple way for clients to request walks. Clients choose a service (a 30-minute walk, an hour walk or a drop-in visit), pick a day and time, and add their dog’s name. The walker sees all requests in one list and approves or declines each one.
It’s small, but it has the things that usually trip up AI builders: two kinds of user, a status that changes (pending, approved, declined) and data that needs to be saved.
Build A: just start prompting
I opened Lovable and typed:
Build a booking app for a dog walker.
The first surprise came before any building. Lovable asked me three questions: should bookings be saved to a private dashboard or sent as a simple request, what walks I offer and at what prices, and which colour palette I liked. That’s new, and it’s Lovable trying to fill the gaps itself. I clicked Skip all, because the point of Build A was to see what happens when you don’t plan.
What came back three minutes later looked great. A business called Happy Tails, a hero photo of a golden retriever, and a tidy three-step booking flow. Underneath, it had made a lot of guesses. It invented three walks I don’t offer (one of them a 90-minute walk that cost less than the 60-minute one), and it had no walker view at all. Bookings only lived in the browser of whoever made them, which Lovable flagged itself in its summary. As a walker, I would never have seen a single request.
Then came the fixing. I typed what a beginner would: “I’m the dog walker. I need my own page where I can see all the booking requests clients send me.” That one prompt did more than I asked. Lovable added a walker login, a dashboard of requests, confirm and decline buttons, a filter for pending, confirmed and declined, and saved bookings in its own database. It cost 2.30 credits, the most expensive step of either build.
The second fix was the walks: “The walks are wrong. I offer a 30-minute walk, a 60-minute walk and a drop-in visit.” Lovable asked what the drop-in should cost, I skipped again, and it made up a price and a length and told me so. Nothing broke along the way.
Final count for Build A: 3 prompts, 4.83 credits, about 11 minutes.
Build B: a few minutes in Miro first
This time I started on a blank Miro board and mapped three things, with no design skills involved.
1. Who uses it. Two sticky notes: Client and Walker. Under each, the things they need to do.
2. The screens. One box per screen, with arrows for how you get between them: the client’s request form, a confirmation, the walker’s list of requests.
3. The data. What a booking actually contains: client name, dog name, service, date, time and status. And the three statuses a booking can be in.
I didn’t use Miro’s AI for this. The map took seven minutes with plain sticky notes and three boxes, and free accounts only get a small number of AI credits each month, so I saved them.
Then I turned the board into a prompt. It took a few minutes, because the thinking was already done:
Build a booking request app for a dog walker. There are two roles. Clients fill in a request form with their name, their dog’s name, a service (30-minute walk, 60-minute walk or drop-in visit), a date and a time, then see a confirmation. The walker sees a list of all requests, newest first, each showing status as pending, approved or declined, with buttons to approve or decline pending ones. Save all requests so they persist after refresh.
This time Lovable asked no questions. It had nothing to ask. Three minutes later I had a request form with my three real services, a confirmation screen, a walker login and a list of requests, newest first, with approve and decline buttons on the pending ones. Requests were saved. Everything on the board was there on the first try, and nothing needed fixing.
It was plainer than Build A. No hero photo, no marketing copy, a form on a cream background. That’s fair, because I didn’t ask for any of that. If you want the polish, ask for it, and it’s cheaper to ask once than to strip out what you didn’t want.
Final count for Build B: 1 prompt, 2.30 credits, about 10 minutes, including the Miro time.
What planning actually saved
Planning cut the cost roughly in half: 2.30 credits instead of 4.83, and one prompt instead of three. On Lovable’s free plan, where you get 5 build credits a day, that’s the difference between finishing in one sitting and running out halfway.
The gap was smaller than I expected, and that’s worth being honest about. Lovable is better at guessing than it used to be. It asked clarifying questions before Build A, and one well-phrased fix prompt added a whole walker dashboard. If you answer its questions instead of skipping them, you’ll probably land somewhere between my two numbers.
Where the plan really paid off was in surprises. Build A invented a business name, a 90-minute walk and prices. Build B built exactly what was on the board. Both versions made the same risky choice, though: the first person to create an account becomes the walker. Neither plan nor prompt mentioned it, so whoever builds this needs to create that account before sharing the link.
Why this works
AI builders guess when you leave gaps, and every wrong guess costs a round of fixes. A one-line prompt is almost all gaps. A plan fills them before you start paying.
Lovable seems to agree. It now has its own Plan mode, where it talks through an approach before touching any code, for one credit per message. That’s a sensible step, and worth using. What it can’t do is decide what your app is for. That part is still your job, and it’s cheaper to do on a board than in a chat that charges by the message.
You don’t need Miro specifically. Paper works. So does a notes app. What Miro adds is space to see the whole app at once: who does what, which screen leads where and what data moves between them. For apps with more than one kind of user, that bird’s-eye view is where I catch the problems that would otherwise turn into five extra prompts.
Miro’s free plan gives you three editable boards, which is plenty for this. Try Miro, and spend ten minutes on a board before you open Lovable.
My planning checklist
Before I open any AI builder now, I write down:
- Who uses it, and what each type of person needs to do
- Every screen, and how you get from one to the next
- What gets saved, and what each saved thing contains
- What changes, such as statuses, approvals or anything that moves from one state to another
Four questions, ten minutes. In this test they saved me two prompts and just over half the credits, and I got the app I’d described instead of the one Lovable imagined.
Get vibe coding project updates
Real build times, real costs, real failures. From someone who is not a developer.