How I work
The hybrid method: why human-led, AI-built is the future
AI writes code fast, but a person understands the business, checks the work and answers you. Why human-led, AI-built is the right way to build.
In May 2025 the CEO of Klarna, the Swedish payments company, said something tech companies almost never say out loud. The company had replaced a large part of its customer service with an AI bot, and now he admitted that the focus on cost had led to lower-quality service. Klarna started hiring people again, so that a customer who wants to talk to a human always can.
Klarna didn't throw out the AI. It put the person back where one was missing. That is exactly the idea my way of working is built on, and I think it's also where the whole field is heading.
What is the hybrid method?
Three steps. You describe an idea in a few lines, in your own words. I build a working first version within days. Then we decide together, in front of something real, whether and how to continue.
Inside each step the division of work is clear. The AI writes most of the code, drafts text, runs tests and does the heavy, repetitive work. I listen, decide what gets built and what doesn't, check what the AI wrote, and answer you. On my home page that fits in four words: human-led, at AI speed.
Why not let the AI do everything?
Because it makes mistakes in ways that are hard to see. In Stack Overflow's 2025 developer survey, more developers said they distrust the accuracy of AI tools (46 percent) than said they trust it (33 percent). The most common complaint, from 66 percent of respondents, was solutions that are "almost right, but not quite". Almost right is the most dangerous kind of mistake, because it looks fine until something breaks.
A July 2025 experiment by the research group METR shows how hard that is to notice. 16 experienced developers worked on 246 real tasks, some with AI tools and some without. With AI they were 19 percent slower, and yet after the experiment they believed the AI had sped them up by 20 percent. To be fair: the experiment tested tools from early 2025, and METR themselves write that the results are no longer current. But the gap between how it feels and what actually happens is exactly why someone has to check.
Then there's security. A Veracode study found that in about 45 percent of cases AI-written code introduced a security flaw, and in the summer of 2025 a Replit AI agent deleted the database of a live system. I went through this in what to check before you hire someone who builds with AI. The bottom line: AI with no person responsible for it is a risk that someone else pays for.
And why not give up the AI?
Because without it the whole method wouldn't be possible. Building a working version before talking about money is a nice idea, but when every version takes a month, nobody can afford it. With AI that changes. When I built a room booking system for the Shiluv clinic, a first booking went all the way through on the day of the first commit, and the system was live twelve days later.
That speed is what lets me build you a first version for free and let you decide only after you've seen it. Anyone working without AI in 2026 is asking the client to pay for the time the machine would have saved.
What does the person do that the AI can't?
They understand the business. The AI can write a booking system, but it doesn't know that at this clinic there are therapists who rent the same room at the same hour every week, and that a standing slot like that has to work differently from a regular booking. You learn that from a conversation with the people who work there, in their words. I wrote about this in Domain Driven Vibe Building: a shared language first, screens after.
They decide what not to build. The AI is happy to add another feature and another screen, because it's cheap for it. But every addition is one more thing to check, explain and maintain. Someone has to say "not now".
They take responsibility. When something doesn't work, you don't write to a bot. On my home page, next to the WhatsApp button, it says "A real person on the other end", and that isn't a slogan. The Shiluv system has a feedback button, and anyone who gets stuck writes straight to me, from the screen where they got stuck.
What does it look like inside the product itself?
The hybrid method doesn't end with the build. It goes into the product too. In the Shiluv system the clinic's Claude can ask the system questions and prepare actions, such as a booking or a cancellation. But every such action goes through two steps: the AI prepares it, and a person explicitly approves it. The AI saves the work, and the person keeps control.
The build itself worked the same way. The team tried each version, wrote to me about what didn't fit, and I fixed it. The AI wrote the code fast, but what decided what was right was a real working day at the clinic.
Why is this the future and not just a phase?
Because the models are getting better at writing code faster than they're getting better at judgment. Veracode's July 2026 report got a headline that says it precisely: the models are getting smarter, but not safer. As writing code gets cheap, the truly expensive part becomes knowing what to build, checking that it's right, and standing behind it.
So I don't think the future has fewer people in the process. It has people in the right place in the process: fewer hours of typing, more hours of listening, deciding and checking. Klarna learned that in front of its customers. I've built on it from day one.