Every Project Is the Same Project: A Failed Architect Trial

I’ve run enough software projects to think I knew the shape of a project. Then we started renovating a cottage above the Balaton, and it went wrong in exactly the ways software goes wrong. I’d told myself building was a more honest kind of work than the abstract stuff I do all day, more grounded, harder to fake. That was wishful. A renovation fails the same way a feature ships wrong, for the same reason, and my whole career should have warned me.

We hired an architect. Sat down for a long afternoon, walked through every room, handed over the vision and the mood boards, talked through the panorama window and the bunk room and the wine cellar and the wellness deck. A few weeks later a set of drawings came back. Most of it missed.

So this is the honest post. Round one didn’t work, we’re parting ways with the studio, and I want to write it down properly, both because a handful of things in the deliverable are worth keeping, and because the way it missed is the most familiar failure I know.

What we asked for

The two earlier posts have the full version. The short one: turn a thatched-roof présház above the lake into a year-round place for families to gather. Inside, we wanted a kitchen that opens onto the terrace, a continuous bench wrapping the green tile stove, the master bedroom’s Balaton-facing wall cut open into a floor-to-ceiling window, a six-bunk kids’ room, and a wine cellar showcased under the brick vault. Outside, a sauna, a hot tub, a fire pit, and an outdoor kitchen, all in natural materials. Wabi-sabi, warm, and built to be touched.

We didn’t hand over a vague feeling. We handed over a written concept and a hundred reference images, room by room.

What came back

The deliverable was a munkaközi terv, a work-in-progress set: a measured survey of the existing house, a demolition plan, a proposed layout, a site plan, and a few 3D massing views. It was competent and properly drawn. It was also almost entirely about the building envelope and almost nothing about the experience we’d described.

The interiors read like a standard cottage renovation. The tile stove stayed, but the bench we wanted wrapping around it wasn’t there. The kitchen was still a room you stand in facing a wall. The master bedroom’s lake wall was still a small window. The six-bunk kids’ room didn’t exist in the plan at all. A new internal staircase landed right in the middle of the ground floor, taking the open space we’d asked to keep open.

None of it was wrong. It just wasn’t ours. We’d handed over a vision and gotten back a permit drawing.

The frustrating part is that the gap wasn’t skill. The drawings are accurate, the dimensions are real, the structure is understood. The gap was intent. The studio heard “renovate this cottage” and we meant “design an experience.” Those are different commissions, and we’d quietly assumed the mood boards would bridge them. They didn’t.

Why it felt familiar

The most common way a software project goes wrong isn’t that the team can’t build. It’s that they build the wrong thing competently. You write the spec, hand it over, and get back something that compiles, runs, passes review, and isn’t what you meant. The brief said one thing and the reader filled in the average. I’ve watched this happen a hundred times, and I’ve been on both sides of it.

That’s precisely what came back from the studio: an accurate survey, real dimensions, sound structure, and a generic cottage renovation sitting on top of it. Not wrong, just the mean of every renovation, which is another way of saying it belonged to nobody. I keep telling myself building is different because the constraints are physical and the feedback is concrete. The actual dynamics are the same as a sprint that ships the wrong screen.

The texture of the result is worth naming, because it’s the one I now see everywhere. Give a model a prompt and it converges on the most probable output, sanding off the specific thing you asked for. That’s slop. These drawings were the most-probable cottage, the design equivalent of a competent, on-distribution paragraph that answers a question nobody asked. And architecture and construction aren’t immune to the literal tool either. Rendering, layout, spec generation, and planning are all sprouting AI features that pull hard toward the same generic center. The failure mode I know from software is showing up on the building site with help.

The parts worth keeping

And yet I’d do the trial again, because four things came out of it that we’d have struggled to produce ourselves.

The first is the legal reality of the plot. It sits in an agricultural zone (Mk, kertes mezőgazdasági terület), and the rules are tight: on the 6,186 m² we can build to at most 150 m² of footprint and 3% coverage, at least 80% has to stay under cultivation, only one building is permitted, and the heights are capped. That single sheet reframed everything. The wellness cabins and outdoor kitchen all have to fit inside those numbers, and now we know the numbers.

The second is the garden, zoned. The site plan placed everything we’d asked for and put it in sensible relationships: the outdoor shower, the fire pit, an outdoor grill, the playground, a rest spot built around an existing tree, the pergola-covered terrace with a swing. It also drew something we hadn’t thought about at all. The hilltop slopes, and water moves across it. So the plan added drainage: a geotextile drain line along the slope, a gravel soakaway, a channel to carry runoff away from the house. We’d never have specified that, and we’d have regretted skipping it.

The third is the cellar. Under the brick vault, the plan set a round tasting table on a polished concrete floor, with the existing stair kept and finished. It’s close to what we pictured: walk down, sit under the vault, open a bottle. Of all the interior moves, this is the one that landed.

The fourth is the bones, measured. We now have an accurate survey and a clear demolition plan, so we know what’s load-bearing, what comes out, and what the buildable shell actually is. Whoever designs round two starts from real drawings instead of a tape measure and a guess.

The pages above are excerpts. If you want the whole thing, survey, demolition, layouts, cellar, and site plan, here’s the full work-in-progress set: download the PDF (5 MB).

What we learned

A survey-and-permit architect and a design partner are two different jobs. Both are real, both are worth paying for, and we hired for one while needing the other. That’s on us as much as on the studio.

The trial was cheap insurance. A few weeks and a modest fee told us the alignment wasn’t there, before we’d committed a whole renovation to it. The alternative, finding out halfway through construction, doesn’t bear thinking about.

Keep the good, restart the rest. The zoning constraints, the site plan, the drainage, the cellar layout, and the measured survey all carry straight into round two. The interior design begins again from zero, and that’s fine.

And next time we’ll test the gap early. This is the one move my day job should have handed me for free, and I skipped it. In software you don’t hand over the whole spec and hope; you build one thin slice end to end and look at it before you commit. So next round we’ll ask for a single room, fully designed, and see whether what comes back has any of the feeling in it. If it doesn’t, we’ll know in a week instead of at the foundations.

That’s the part I keep turning over. I thought the cottage would teach me something new, and instead it taught me that the project I’ve been running for years and the one I’m running now are the same project wearing different clothes. The materials change. The way it succeeds or fails doesn’t.