From the Barcelona Marathon to an MVP

One of us is going to run the Zurich Marató Barcelona. Family and friends want to cheer you on, but they face a pretty simple problem nobody has solved well: what time will you pass a specific point?

Garmin and Strava track you live, sure, but you need to be set up as a contact to receive an email or SMS, and even then they don’t tell your supporter what time you’ll reach the corner where they want to cheer you on.

That gap is where RaceRoar comes from. An app that shares your position in real time with whoever you choose, estimates when you’ll arrive at each point along the course, and requires no sign-up, no account, no personal data. You share a link, you run, and when you cross the finish line you leave the race and the link stops working. Use and discard.

The idea was clear. Two weeks to get it on the stores. A small team can’t pull that off alone, no matter how good they are. But at Lostium we’ve been working as an AI First company for months: an AI agent joins every project from minute zero, not as a support tool, but as another team member. It takes part in ideation, proposes solutions, executes and corrects.

In this case the chosen agent was Claude Code. As of today, March 2026, it’s the tool that gives us the best results in our workflow. It’s not the only one we’ve tried, and we don’t rule out that something better will come along tomorrow. But right now it’s the one that fits best with how we think and build.

A new teammate in the brainstorming

Before writing a single line of code, we brought the agent in to talk about what product we wanted. Not about technology. About needs.

The session started with the usual questions: who’s the user?, what frustrates them about what already exists?, what’s the bare minimum that has to work on race day? But the agent came back with questions we hadn’t asked ourselves. Privacy as a competitive advantage: if your app doesn’t require sign-up, doesn’t store history, and the link expires, you’re not competing with Strava on features — you’re competing on trust. Anyone can see where you are without you having to open the door to your sports profile.

Would we have reached the same value proposition on our own? Probably yes, but not in one afternoon. The agent shortens the path between the initial idea and the good idea because it doesn’t carry the assumptions you’ve been turning over since the project first came to mind.

From concept to the first visual prototype was also fast. With Stitch, Google’s design tool connected to the agent via MCP, we had real screens to discuss before touching code.

Trusting someone who knows what you don’t

After scoping the product, the agent proposed the tech stack. The backend and the web frontend fit within what we already handle at Lostium, though with more innovative approaches: free maps without depending on Google, a geographic database to calculate distances and arrival times, real-time communication so viewers don’t have to refresh anything, and the entire server set up automatically with a single command.

The native app is a different story. A single codebase for iOS and Android with technologies we hadn’t mastered.

Here comes the real business decision: do we take the risk of building with tools we don’t control?

The app is very simple: two screens. The scope is contained. If something goes wrong, the cost of pivoting is low. So we said yes.

It’s not blind faith. It’s weighing the project’s scope against the risk of the unknown. For a 200-screen ERP we would have said no. For a two-screen app with an expiration date, the calculation is different. And that ability to evaluate is precisely what sets using AI with judgement apart from using it recklessly.

10 languages and two stores without losing your mind

Races attract a diverse, international audience. So we internationalised from day one, not as an afterthought. RaceRoar is available in the four main languages in Spain: Spanish, Catalan, Basque and Galician, plus English, German, French, Italian, Portuguese and Dutch. 10 languages from the very first release.

The agent generated descriptions, metadata and screenshot texts in all of them. And we’re not talking about literal translations: each language has its own tone, its own store conventions, its own maximum title length. Work that without AI would have taken days, done in hours.

Was it a walk in the park? No. The stores have their own rules and apply them with shifting criteria. Several submissions rejected, metadata that didn’t comply, screenshots that didn’t fit that day’s guidelines. Back and forth until everything clicked. But even with the rejections, the correction cycle with the agent was fast: read the reason, adjust, resubmit.

Your teammate is not infallible

Just because the agent is good doesn’t mean it foresees everything. The app stopped broadcasting GPS as soon as you locked the screen, a native-app classic that we discovered running — literally — down the street. Building native apps is still a nightmare of certificates and SDK versions, though Expo EAS helped a lot with cloud-based builds. Some things only come to light when you get your hands dirty.

What’s next

As we write this, the store review process has us on edge. The Barcelona Marathon is approaching and it’s not up to us whether the app makes it in time. If it doesn’t, that’s fine: we’ll wait for the next race. The product is already built and developing it this way has been a complete success.

And precisely because it’s ready, we’re already thinking about what comes next. Online cheerleaders who can send live messages to the runner. A Spotify-style wrapped with the messages, photos and videos your people left you during the race. And why not, opening the door to organisers and sponsors who want to offer official tracking at their events and explore potential revenue streams.

These are just ideas, but that’s what happens when you build a working product in two weeks: you reach the roadmap before the momentum goes cold.

The method, not the magic

Without an AI-First approach, this project doesn’t ship in two weeks. It’s not that the agent works magic — it’s that it radically changes what’s possible for a small team in a short time. The ideation was deeper, the technical decisions better argued, the execution faster. And when something breaks, the correction cycle is measured in minutes, not days.

But none of this works without human judgement. The agent doesn’t know your app needs to broadcast GPS in the background until someone goes for a run and checks. AI multiplies, but it doesn’t solve everything.

RaceRoar is a concrete example of how we work at Lostium. If you’d like us to help you bring this way of working to your team, request the AI First Kit: our free preliminary guide with a maturity assessment and first steps for your organisation.