I gave two AI Builders the same brief. None of them won!

Last month, I spent some time building a small product on the side.

A tool for the people who run hospitality at a Formula 1 race weekend.
The ones tracking VIP guests, paddock access, and a 3 day itinerary for each person across practice, qualifying, and the race.

This is not a toy market.

F1's Paddock Club served 65,000 guests last season at around $7,000 a ticket. That is over $450 million in revenue, growing 10 percent a year.

Behind those numbers, someone is doing real coordination work. That is the user I was designing for.

Before building it myself, I ran an experiment.

The experiment

I wrote one detailed brief. A full handoff doc. Data model, routes, brand, the F1 aesthetic, all of it.

Then I handed the exact same brief to two AI app builders -

1. @boltnew
2. @lovable

Same words, same day, no extra steering.

I did not just glance at the previews. I logged into both with a real headless browser and clicked through every screen. Dashboard, guests, itinerary builder, share flow, the public guest page, even invalid links.

What Lovable did

Lovable was the obedient one.

• It read the calendar in my doc and got it right. June onward, correct weekend dates, even the new Madrid round.
• It stubbed the email sending exactly as asked.
• Routes, data model, access-tier badges, CSV import, colour-coded itinerary. All there.
• The brand mark was genuinely decent. Strong engineering. It de-risked the boring 70 percent in minutes.

Then I tried to do the one job the whole product exists for.

Give a guest a schedule.

I could not find it.

The flow was buried behind a 2 pixel grey label that said "0 assigned." No guidance, no empty state, nearly invisible. I wrote the spec, and I could not find my own core feature. That is not a nitpick. That is the product failing at the one thing it is for.

What Bolt did

Bolt had better instincts and worse discipline.

• Assigning a guest was obvious. The guest showed up as a chip right on the itinerary item, handled inside the form. Exactly the workflow Lovable hid.
• Typography was calmer. Title Case, not all caps everywhere.
• It showed the full guest link in plaintext instead of hiding it behind a button.

Then I looked at the calendar.

Bolt ignored my doc and invented its own. It put Bahrain as Round 1.

It brought back races that are not on the schedule. For an F1 audience, that is an instant credibility killer. They will clock it in two seconds.

So I had a real trade on my hands.

• Lovable: faithful to the data, weak UX.
• Bolt: sharp UX, careless with the facts.

Neither was sendable. And here is the part that actually taught me something.

Why were they so similar?

Strip away the calendar bug and the assign flow, and the two apps looked like siblings.

Both defaulted to dark mode. Both used the same font, the same neon red on near-black, the same all-caps labels, the same flat grid of identical cards. Neither had a single image. Two
different companies, two different models, near-identical screens.
That is not a coincidence. That is the brief.

I had over-specified the visuals. Exact hex codes. Exact font. Sharp corners. n every token, you collapse the design space to roughly one answer. The briefdid the "design." The tools just painted it in. The sameness was baked in before either of them started.

One more tell. The headless browser I used had no dark-mode preference set, and both apps still rendered dark anyway. Neither built a light option. They went dark because my brief said near-black background, full stop.

The lesson

Tokens are not design.
A hex code and a font name produce a screen. They do not produce hierarchy, re deliberate choice about what to leave out. My brief specified all the paintand none of the architecture, so both tools defaulted to the same safe admin panel wearing F1 colours.

Real design was the part I never wrote down. Layout. Emphasis. What the eye lands on first. What gets cut.

So I rebuilt it, and the first thing I threw out was my own aesthetic.

• Light by default, dark as a choice, not a cage. - Editorial and image-led. Real circuit photography carrying the emotion, not
• Space Grotesk for character, Inter to do the quiet reading work. - Red used sparingly as an accent, not a coat of paint.
• And I kept the one thing Bolt got right. The discoverable, chip-based way of assigning a guest.

Same problem. Same category of tools. A completely different result. The difference was judgment, not technology.

If you brief an AI tool or a teammate, where is the line between giving enough direction and accidentally handing over the one decision only you should make?