Common Lovable mistakes, and what each one costs
Lovable bills Build mode on the work it does, so a prompting mistake is a purchase: the round that corrects a wrong guess costs credits like any other. Below are the 8 mistakes that cost the most, each with the line that prevents it, what my own runs showed, and the order Lovable's docs give for fixing one that already shipped. One of them has a CVE rated 9.3, critical.
By Marco Kohns · Docs and the CVE record re-read 23 October 2026
8mistakes
Each with the prompt line that prevents it.
9.3CVSS
CVE-2025-48757, missing row-level security in generated apps, critical per NVD.
10free fixes
Per account; each comes back 24 hours after you use it. Then fixes cost credits.
On one screen
Which Lovable mistakes cost the most?
What each costs under usage-based billing, and the line that prevents it. Lines follow this site's prompt standard.
| Mistake | What it costs | Write instead |
|---|---|---|
| Prompting vaguely, then paying to correct the guess | Two Build rounds instead of one | Dashboard, Data-driven and analytical style: Inter with tabular figures, background #f8fafc, accent #2563eb; shadcn Card per metric with the number right-aligned, a 7-day range Select, a Skeleton while data loads |
| Asking for the whole app at once | Six weak features instead of one strong one | Step 1 of a project tool: a task list with statuses todo, doing and done, create and edit forms with validation, empty, loading and error states. Tasks in the built-in backend, each user reads only their own. Nothing else yet. |
| Skipping empty, loading and error states | One retrofit round per screen | Every list and form has an empty state that offers the first action, a loading skeleton shaped like the content, and an error that keeps the input and offers a retry |
| Forgetting the phone | A layout rebuild after launch | Mobile first at 390 px; tables scroll inside their own container; the primary action stays reachable without scrolling sideways |
| Letting Lovable invent your social proof | A claim on a live page you cannot back | Add three testimonial cards labelled PLACEHOLDER; never generate names, companies, star ratings, quotes or user counts |
| Trusting generated access control | The one with a CVE rated 9.3 | Roles in their own user_roles table; every table has row-level security, deny by default; Done when: a non-admin account selecting another user's rows gets zero rows |
| Restating the feature when 80% was right | A large regeneration plus regression risk | The booking flow is right except the time picker, which allows past dates. Limit it to future slots in the venue's time zone. Change nothing else. |
| Building with no way to be found | An app with no acquisition surface | This app grows through search: every public page is server-rendered with one H1, its own title and description, an FAQ with structured data, and a sitemap |
Billing rule: Lovable docs, Credits and usage, read 23 October 2026: “Small, focused edits usually cost less than larger generations, multi-step changes, or requests that require more codebase exploration.”
Before and after
What does each mistake look like in a prompt?
01 · Two Build rounds instead of one
Prompting vaguely, then paying to correct the guess
Make the dashboard look better and more professional
Dashboard, Data-driven and analytical style: Inter with tabular figures, background #f8fafc, accent #2563eb; shadcn Card per metric with the number right-aligned, a 7-day range Select, a Skeleton while data loads
A look left open triggers Lovable's design questions or its defaults, and you pay the round that corrects them. A named style direction with fonts and colours builds directly.
02 · Six weak features instead of one strong one
Asking for the whole app at once
Build me a project management app with tasks, teams, chat, calendar, reports and billing
Step 1 of a project tool: a task list with statuses todo, doing and done, create and edit forms with validation, empty, loading and error states. Tasks in the built-in backend, each user reads only their own. Nothing else yet.
One prompt cannot prioritise six features, so effort spreads evenly. Lovable's best practices ask for one meaningful change at a time with a check in between.
03 · One retrofit round per screen
Skipping empty, loading and error states
(the mistake is an absence: the states are never mentioned)
Every list and form has an empty state that offers the first action, a loading skeleton shaped like the content, and an error that keeps the input and offers a retry
My states run built exactly the states the prompt named and gave first-use and cleared empties separate cards because the prompt said they differ. Unnamed states get no design.
04 · A layout rebuild after launch
Forgetting the phone
(an absence again: no breakpoints, so desktop layouts ship)
Mobile first at 390 px; tables scroll inside their own container; the primary action stays reachable without scrolling sideways
Responsive behaviour added later touches every component. Asked for up front, it shapes which components get chosen.
05 · A claim on a live page you cannot back
Letting Lovable invent your social proof
Add a testimonials section to build trust
Add three testimonial cards labelled PLACEHOLDER; never generate names, companies, star ratings, quotes or user counts
Asked for proof, Lovable writes plausible proof. My landing page run (1.8 credits) came back with an invented brand name, exactly where the prompt had scoped proof to placeholders.
06 · The one with a CVE rated 9.3
Trusting generated access control
Add roles so only admins can see the admin page
Roles in their own user_roles table; every table has row-level security, deny by default; Done when: a non-admin account selecting another user's rows gets zero rows
A hidden menu item is a UI check. Only a policy in the database, tested as the wrong user, protects the data.
07 · A large regeneration plus regression risk
Restating the feature when 80% was right
That's not what I wanted. Build the booking flow like this: [the whole original prompt again]
The booking flow is right except the time picker, which allows past dates. Limit it to future slots in the venue's time zone. Change nothing else.
Lovable's docs: small, focused edits usually cost less than larger generations. Correcting the miss is a small edit; the whole spec again is a regeneration.
08 · An app with no acquisition surface
Building with no way to be found
(an absence: features specified, distribution never)
This app grows through search: every public page is server-rendered with one H1, its own title and description, an FAQ with structured data, and a sitemap
Distribution is structure. An app built for search needs indexable pages; one built for sharing needs a result worth posting. Neither is added the week after launch.
Measured
Which mistakes did my own runs hit or avoid?
Three library prompts pasted unedited into fresh projects in my own workspace, credits read before and after.
| Run | Credits | Avoided | Showed up |
|---|---|---|---|
| Landing page + email opt-in | 1.8 | States and consent present in one round | An invented brand name in the copy (mistake 5) |
| Empty states and skeletons | 9.5 | 2 questions before building; every named state built as its own card | Nothing from this list |
| SaaS MVP with user auth | 10.4 | 4 questions first; roles checked in the database on every read and write | Nothing from this list |

The security mistake
How serious is trusting Lovable's generated access control?
Serious enough to have a CVE. The fix is a policy in the database and a query run as the wrong user.
The record
CVE-2025-48757
Today
What the scans do
The proof
Query as the wrong user
Record: NVD, CVE-2025-48757, read 23 October 2026. More on the scans and the CVE in the Lovable security guide.
After it shipped
What do you do once a mistake has shipped?
Lovable's debugging guide gives the order. Skipping it is how mistake 7 happens.
- 0110 free
Try to fix, once or twice
Each free fix comes back 24 hours after use; after that, fixes are billed as Build work. - 02Build
Describe the bug
What is broken, where, what you expected, what happened. Attach the error or a screenshot. - 031 credit + research
Switch to Plan mode
The docs warn that repeated blind fixes pile up code that hides the cause. Plan finds the cause before more code changes. - 04No refund
Revert if fixes tangled it
Restore a version, or edit the message and Revert and resend. Code only; database rows stay.
Order and quote: Lovable docs, Debug and improve your app, read 23 October 2026. Every action's price is in the app builder guide.
Prompts with the fixes already written in
Disclosure: if you create a Lovable account through links on this page, we may earn a commission. It costs you nothing and does not change what we recommend. Create a free Lovable account
Frequently asked
What is the most expensive Lovable mistake?
Being vague against a large project. Lovable bills Build mode on the work done, and its docs say requests that need more codebase exploration cost more than small, focused edits. A vague instruction makes Lovable explore and guess, and the correction is another billed round. Name the screen, the component and the change.
How many free fixes does Lovable give?
Ten per account, shared between Try to fix and fixes from the Security view; each comes back 24 hours after you use it (Lovable docs, read 23 October 2026). After that, fixes are billed as Build work. After one or two failed attempts the debugging guide says to switch to Plan mode.
Why does my Lovable app break when a list is empty?
Because no prompt named that state. Generated apps cover the path that was described; empty, loading and error states appear when the prompt asks for them. One line in the first prompt covers every list and form; retrofitting costs a round per screen.
Is it a mistake to build the whole app in one prompt?
Yes. One prompt cannot rank six features, so each gets a share of the effort. Build the foundation first, then one feature per prompt in priority order, checking between them, as Lovable's best practices recommend.
Does Lovable's security scan prove my data is safe?
No scan can. The Quick and Deep scans are free and catch common gaps such as tables without row-level security. Whether a member can read another member's rows depends on what your app intends, and only a query run as that member answers it.
What is CVE-2025-48757?
A vulnerability record published 30 May 2025: insufficient row-level security in Lovable-generated apps through 15 April 2025 let unauthenticated attackers read or write arbitrary tables. NVD rates it 9.3, critical. Lovable disputes it on the grounds that each customer is responsible for protecting their app's data.
What we did not publish
- A credit figure per mistake. The costs are rounds, from Lovable's billing rule; I have not run each mistake on purpose to price it.
- The share of Lovable apps that ship without row-level security. No source I could verify publishes it.
- Runs in Bolt or v0.
Sources
- Lovable docs: Credits and usage, read 23 October 2026
- Lovable docs: Debug and improve your app, read 23 October 2026
- Lovable docs: Design guidance, read 23 October 2026
- Lovable docs: Security view, read 23 October 2026
- NVD: CVE-2025-48757, read 23 October 2026
- My runs: 2026-07-31, 2026-09-26, 2026-09-26
Written by

Marco Kohns
Founder of ProtoBites - Venture Growth Studio
Growth PM at a Silicon Valley scale-up (a16z and General Catalyst backed), ex-Techstars where he consulted 13 early-stage startups, Reforge-trained. Every prompt on this site comes out of shipping ProtoBites' own portfolio products.
Related reading
- every Lovable guide, by build stage · Choose a tool, plan, prompt, build safely, launch.
- how to write prompts for Lovable · The structure that prevents most of these.
- Lovable's debugging order, in detail · Try to fix, describe, Plan mode, revert.
- backend prompts that enforce access in the database · The fix for mistake six.
- the states prompt · The fix for mistake three, run and shown.
- build kits with tested row-level security · Policies executed, not described.
- Lovable pricing and credits · Plans read today and what one prompt costs in credits.
- building with a way to be found · Mistake eight, solved before the prompt.
- the Replit incident · What an agent with too much permission did.
- Lovable SEO · The sitemap and internal-link mistakes that cost us months.
- from idea to app, measured · The path that avoids the rebuild.
- vibe coding prompts · First prompts that carry the safeguard.
- Lovable and Supabase · The three RLS failures and the lines that prevent them.
- what is vibe coding · The public record on AI-written code, with sample sizes.
- the opt-in prompt that forbids invented proof · Placeholders where testimonials would be.
- Lovable review · Where Lovable is strong and where it is weak.