Independent product retrospective · 2026

When AI built the wrong product

I asked AI for a prompt, gave it to Lovable and watched a plausible-looking app appear. Now I know better: a prompt is not a product definition.

Paused product and retrospective

alliwant.is was built, then paused when the product model and the overgrown AI context became too unreliable to keep extending.

The built interface. It looked coherent before the complete journeys were coherent.

The challenge

I asked AI for a Lovable prompt before I had properly defined the people, language, relationships and complete journeys. AI filled the gaps, and the database quietly became the product.

My role

My idea, my direction, my mistake. AI did exactly what I allowed it to do.

What changed

I stopped instead of piling more work onto bad foundations. The useful result is the stricter process I now use before asking AI to build.

01The app I thought I was building

alliwant.is was meant to make wishlists and occasion planning calm, useful and private. The real concepts were people, wishes, occasions, reservations and selective sharing.

I had an idea. I did not yet have a defined product. Instead of stopping there, I asked AI to write a prompt and gave it to Lovable.

02It looked like a product

Lovable produced navigation, forms, cards, responsive layouts and consistent styling astonishingly quickly. Then AI turned those screens into polished promotional images. The whole thing looked increasingly convincing.

AI-generated promotional board showing alliwant.is on laptop and phone mockups with polished feature panels — AI-generated promotional concept using the built interface. It made the product look more convincing. It is not evidence of shipped behaviour.

03Then I tried to use it

The interface worked. The product did not. Ordinary tasks exposed all the decisions I had never made.

Circle cards led nowhere useful.
Some things looked clickable but were not.
The Wish form was dense and ownership was unclear.
Development tools and placeholders appeared inside the real app.
Basic actions around people, wishes and occasions were weak or missing.
Labels followed database concepts instead of ordinary language.

The built alliwant.is People screen with connections, removal actions, circles and a create-circle form — Built interface. Connections, circles and account actions exposed the unresolved product model.

The built alliwant.is mobile home screen in its light theme — The responsive implementation worked. That did not make the product structure right.

What I learned

04Stopping was the correction

The product never reached a neat before-and-after redesign. Once the Lovable context had become bloated, it started losing decisions and inventing new ones. Fixing one area could quietly break another.

I paused the project. The result is not a prettier replacement screen. It is a much stricter process: define the product, people, language and complete journeys before asking visual or coding AI to build.

Stop extending unreliable foundations

More prompting would have added more work without restoring a coherent product model.

Keep the failure visible

Good-looking output is not proof of a good product.