Open the original Tallybook dashboard and give yourself five seconds to answer one question: which invoices need a follow-up? You will probably notice the cards before you notice the invoices. Open the revised version and try again. The rows, amounts and dates are the same. What changed is the relationship between the summary and the work below it. You can compare the two screens without knowing anything about the tool that wrote them.
That is the kind of judgment people mean when they say an app looks AI-generated. A reviewer cannot identify a coding tool from a screenshot. A human can make all nine choices below; an agent can make none of them. Lovable is in the title because an owner working from a generated first draft often has a specific practical problem: the app runs, and a person seeing it for the first time still cannot tell what makes it theirs. Read the following signals as questions you can test, not as a list of styles you have to ban.
1. A gradient where the product should be
A wash of pink and violet behind a headline can look pleasant. Look at what it occupies. On the original Pourbook landing page, the interesting fact is that a person can choose a dose and see the water divided into three pours. A large decorative area has to earn its place against that working example. If the first viewport promises a better coffee routine but never shows the recipe, the background is taking up evidence space.
Try covering the decorative area and putting an actual recipe there instead. Write “18 g coffee, 288 g water” beside the control. Ask someone what they would change to make a larger cup. If they can answer from the page, the new composition explains something the old one did not. A gradient that carries information, such as the scale on a heat map, passes a completely different test.
2. One accent with six unrelated jobs
Find everything purple on the screen. You may see a primary button, a selected tab, a positive badge, an avatar background, a promotion and a chart line. They have different jobs, but the color tells you to give them equal attention. Changing purple to green preserves the problem. Try assigning meanings before changing hues: one color for an action, one treatment for selection, a written label for status and a quiet surface for decoration.
Test the result without color. An overdue invoice should still say “Overdue”; a selected filter should still have an identifiable state. Then restore color and see whether it shortens the path to the next decision. If every button needs a stronger accent than its neighbors, look at the number of competing actions instead of searching for a brighter shade.
3. Emoji standing in for a feature
A rocket above “Launch faster” describes neither a feature nor a speed. Three such bullets can make a row feel complete while leaving a buyer unable to name what the product does. Replace one bullet with a sentence you could demonstrate: “Save the dose and reopen the recipe tomorrow.” Keep the emoji if it belongs to the product's vocabulary. A feed about birds may have a good reason to show a bird; an invoice export probably does not need a rocket.
Check the narrow version too. Icons often occupy an entire extra row when their labels wrap. Removing one decorative symbol might bring a useful control into view. If the label still sounds vague without the emoji, edit the sentence before paying someone to draw a better icon.
4. Default gray components that never meet
A familiar input or dialog is useful. Trouble starts when the project list, proof, owner and due date arrive in separate gray boxes because the component library made boxes convenient. In Folio those details describe one decision: which proof needs my attention? Give the title, review state and date a common alignment. Let a boundary separate projects, not related facts within a project.
Temporarily remove the backgrounds from the cards. If everything becomes a heap of labels, there was no grouping underneath. Put related information in reading order and use spacing to hold the group together. Add back a border only where the reader changes objects. You can keep the library's input behavior, focus outline and dialog semantics while replacing its incidental arrangement.
5. A headline that could sell the next product
Copy the main headline onto a blank page. Could it advertise a scheduling tool, a recipe app and a billing dashboard without changing a word? “Your all-in-one solution” passes that unfortunate test. Useful copy makes a claim small enough to inspect. “See which October invoices still need payment” sets an expectation that the ledger can meet or disappoint.
Look one level below the headline. A specific heading followed by three identical “Learn more” links leaves the decision unresolved. Name the destination or the action: “Open the October ledger” or “Change coffee dose.” If no destination exists yet, do not imply that it does. A modest description of a working feature is more useful than a large claim attached to an empty button.
6. Every line centered
Centering can suit a short invitation. It is a poor default for a list people scan repeatedly. Each centered line starts at a different horizontal position. Now put amounts, dates and changing customer names into that arrangement. A reader looking for the third overdue item has to keep finding the start of the row.
In a table, align names to a stable left edge and comparable amounts to a stable right edge. At phone width, an invoice can become a compact row with its status below the name. Preserve the same order throughout the list. The goal is not to make a mobile page look like a spreadsheet; it is to let the eye repeat a successful movement. Watch a reader find the second item, not just the first one.
7. Equal emphasis for unequal facts
Four identical metric tiles are often easy to implement. On a working billing screen, “total invoiced,” “already paid” and “still to collect” do not necessarily deserve equal weight. Decide what someone should do next. A person chasing late payments needs the overdue amount, the records it includes and a way to act. They do not need every historical number to shout at the same volume.
Take a screenshot, shrink it until the labels are difficult to read, and note which shapes dominate. Then return to full size and check whether those shapes contain the right information. A large numeral can be honest and still distract from the action. You can leave all the accounting facts on the page while changing their order and prominence.
8. A screen designed for its sample data
Search the original portal for a project that does not exist. A design comfortable with three perfect records may become a blank rectangle when there are zero. Now try a long project name and open the page at phone width. Does the status disappear? Can a client recover without guessing which filter is active? Empty and narrow conditions reveal decisions the attractive first screenshot hides.
Write the missing state in the language of the task. “No projects match Atlas. Clear search” distinguishes a filter from a new account with no projects. A failed request needs a different answer, since clearing the filter cannot restore the connection. Test each state on its own. Pasting the same “Get started” card into all three creates a consistent appearance and three incorrect instructions.
9. Spacing that follows the template instead of the task
Measure two gaps: title to supporting date, then one project to the next. If they are equal, readers have to infer which date belongs where. Choose a small set of spacing steps and use the shorter gap inside a group, the larger one between groups. Preserve comfortable controls at narrow widths. You do not have to turn a one-pixel border into an eight-pixel stripe just because the main spacing scale uses eights.
Finally, return to the five-second question. Ask a person who did not make the changes to identify the overdue invoices, the proof awaiting approval or the control that changes the dose. Record where they first looked and what they tried. Someone who finds the right row but cannot open it has uncovered a different defect. Fix that next, even if the revised screenshot already looks better.
A borrowed palette can become a second template. We put the five finished demo screens beside each other and found that coffee, chat and billing all used the same serif, cream and green. The billing screen now uses a compact Utility ledger; the chat uses Dark to keep the thread in focus. Look at the five working comparisons and ask what would be harder to do if you swapped their design decisions.
In the real-repo trial for studio billing dashboard, the plain agent won in both orders. The two votes were 3 and 3 on a seven-point scale oriented toward the kit, with four neutral. This measures a billing task. A vote on one ledger does not establish that a palette works across every product.
Try the free audit on your Lovable preview. Keep a pair of captures, a short task and the evidence that the task still works. If you hire help later, those three things let you describe the repair in terms that survive a change of designer, framework or model.