I am retiring FrunkFriends. It was a social network for Tesla owners, and it failed: in the months after the launch, very few people signed up.

It was also my first attempt at building a whole platform with the new AI tools, Claude Code first, in early 2026. I made every mistake a first-time founder can make. I regret none of it.

FrunkFriends banner: the private community for Tesla owners, with a Cybertruck on a desert road at sunset

What FrunkFriends was

The idea was simple: a place where every member is a real Tesla owner. The only way in was to sign in with your Tesla account, so there were no bots and no fake profiles. Inside, you found a feed, a group for each model, local events, private messages, and Frunky, an AI that answered Tesla questions in the comments. Your car could even post: battery level, odometer, tire pressure, the song playing in the cabin.

On paper, the timeline looked like a startup’s. First line of code in February 2026, a closed alpha on Telegram in April, a public beta on iPhone, Android and the web in June, and a launch on Product Hunt and Uneed on 24 June, where it finished third of the day on Uneed.

Then very little happened. Signups trickled in, a few at a time, for months.

Everything I got wrong

I built what nobody asked for

With Claude Code, a new feature took an evening, sometimes less. So I kept adding them: 49 badges in six families, live cards from the car, encrypted messages, an AI in the comments, a web app next to the two mobile apps, translations into each reader’s language. Every one of them worked. None of them brought in new members, because nobody was waiting for them.

The cheaper code gets, the easier it is to hide in it. Shipping a feature feels like progress. It is not, if nobody uses it.

I never focused on marketing

I did what most new founders do: I built, and I hoped people would come. A launch day is not marketing. After the launch, I went back to the code, the part I knew how to do, instead of reaching Tesla owners where they already were, week after week, and giving them a reason to switch.

A product nobody hears about is a product nobody uses, however good the code is.

Founders never fail, they learn

I like that line because it is true when you let it be. FrunkFriends failed as a product. As an education, it worked, and I learnt a lot.

The best proof is what came next.

How FrunkFriends led to EVPick

In July, I started a side piece for FrunkFriends: a knowledge base about Tesla cars that Frunky could draw on, so its answers came from real specs instead of from memory. A RAG, in AI jargon, served from a FrunkFriends subdomain and meant for internal use.

It turned out to be the first stone of an epic founder journey. “Which Tesla fits me?” quickly became “which electric car fits me?”, and in September the side piece got its own name and domain: EVPick.

Today EVPick compares more than 1,900 versions of electric cars from 90 brands, with prices in 15 countries, in ten languages. You ask in plain words, it answers from the catalogue and underlines every figure it took from it. I wrote about how I built it and about how it now looks after itself. It launched on Product Hunt on 13 November, with this video:

Car photos in the video are credited on each car's page at evpick.ai.

What I do differently now

One lesson became a rule. The others went into EVPick.

  • No more social networks. I will never build one again. Getting users is far too expensive, and an empty room is a deal breaker: when a new member opens the app and finds nobody there, they don’t come back.
  • Start from a question people already ask. Every day, people search for which electric car to buy. EVPick answers that question, with thousands of pages that search engines can read, before it adds anything else.
  • Marketing is part of the work. EVPick has a video project that renders its social clips (45 so far, including the one above), a brief for whoever posts them, and a log of every launch site and directory it was sent to.
  • Agents carry the routine. Data, bugs and search traffic are watched and fixed by Claude Code routines, so my own time goes to the product and to the people using it.

No regrets

I started FrunkFriends with @kleak_dev, and I am proud of what we built together. We both learnt how to work with AI, and not only how to prompt it: the harness around the agents, the instructions, skills, checks and routines that make their work reliable. That is what makes us much better founders than we were in February.

FrunkFriends did not find its users. It found us our next step.

To every Frunker who signed up: thank you. And if you are choosing an electric car, give EVPick a try.