Vibecoding without losing the customer

With AI tools you can build faster than you can think things through. How I stay anchored to the plan and the customer when ideas branch every day.

Dotun Ogunkanmi · 2026-11-04 · 5 min read

Vibecoding – building software by describing what you want to an AI that writes the code – has gone from curiosity to a real option in a very short time. Since June 2025 I have been building Nielhouse, a service for home buyers, that way, with tools such as Lovable, Cursor and now Claude.

In the previous part of this series I wrote about when it's worth building your own. This time it's about what happens once you do – and about the biggest risk of vibecoding, which isn't technical.

What goes faster than you think

When I started, the tools were much weaker than today, and it was often frustrating. Anyone starting now will get further, faster, than I did then. But even from the start, some things were surprisingly quick.

Coming up with ideas is fast, because all the conversations happen between you and the tool instead of in meetings. Building a first version of a feature takes hours. And above all, it's fast to change direction: you can try something, see that it doesn't work, and switch course from one day to the next.

Every answer spawns three new ideas

What surprised me most was something else. Vibecoding tools constantly suggest new things. You can start with a clear plan – "I want to build these five features" – but as soon as the first one is done, you have two or three suggestions for what you could add. The suggestions are often good. That is exactly why they are dangerous.

If you follow them, you drift, step by step, further and further from what you originally set out to solve. A few weeks later you have a product with many clever features but no clear answer to the question: who is this for, and what problem does it solve?

When every conversation happens in your own head

In a normal development team there are other people who ask questions, challenge you and do quality checks. They keep you grounded. When you vibecode alone, all of that happens in your own head. You decide everything yourself, which is a strength – but it easily becomes a self-fulfilling prophecy.

It also makes it harder to turn back. I have gone down rabbit holes for days, sometimes weeks, before realising that what I was building wasn't needed. The more time you have put in, the harder it is to let go – and there is no colleague to tell you so.

My own journey with Nielhouse

My original idea was that home buyers could describe their life – what they like, which neighbourhoods suit them, what matters in their day-to-day – and be matched with homes that fit. I quickly hit dead ends: it was hard to get homes onto the platform, and hard to translate what buyers really want into something a service can work with.

So I changed course and turned the service into a notebook for the home search that you can share with a partner or family. It solved a real problem – much of what you collect is scattered across text messages, notes and your own head – but only a small part of what makes buying a home difficult. Each turn took days, not months. But each one also taught me how easy it is to confuse what I find exciting with what users need.

How I stay anchored to the customer

This is what works for me today:

1. Write the plan down before you start building. A short document is enough: which problem, for whom, the three to five most important features – and what is explicitly out of scope.
2. Park the suggestions. When the tool suggests something new, I add it to an idea list instead of building it straight away. Only what supports the plan gets in.
3. Decide what "done" means for each feature. Otherwise nothing is ever finished. That applies just as much to one person with AI tools as to a development team – I wrote about it in Why do IT projects stall?
4. Let real users test early. Start with people who know you and understand your vision. But don't take their answers as the truth. You get the truth from potential users who don't know you and have no loyalty to you.
5. Don't take feedback personally. You have thought through how everything should work. But for the person who will pay, or whose everyday life should get better, their needs matter more than your vision. That is harder to accept than it sounds.
6. Give every side track a limit in time and money. It's easy to burn credits and days. A simple limit makes it easier to turn back in time.

What it means for companies

More companies now let employees and smaller teams build their own tools with AI. That can bring big gains in speed and creativity. But the same risks apply – and they grow when more people build at once. It doesn't take heavy governance, but it does take a few ground rules: a clear owner and plan for each solution, real users testing early, and an answer to who will maintain and secure the solution once it's in use.

Speed isn't the problem

Vibecoding makes it possible to build faster than ever. That's fantastic. But the speed also makes it easier to build the wrong thing, very efficiently. The most important discipline isn't technical. It's returning, again and again, to the question: who am I building this for, and what are they telling me?