The Winchester Mystery House is usually told as a ghost story. Sarah Winchester, heiress to the rifle fortune, building endlessly to appease the spirits of the people her family's product killed — staircases that climb into ceilings, doors that open onto walls, carpenters working in shifts for 38 years.
Drew Breunig's account takes the haunting out. She built it because she liked architecture. No license, no formal training, and no path in her era for a woman to practice it, so she practiced on her own house with effectively unlimited money. At its largest the house had roughly 500 rooms. Projects were planned, completed, abandoned, torn down, rebuilt. "It was anything but aimless."
That reframe should be more alarming than the ghost version, not less.
That reframe should be more alarming than the ghost version, not less.
Because the same description fits a large share of what teams ship right now, and unlike Sarah Winchester, we have users.
Start with the arithmetic. Public GitHub commits attributed to Claude Code now average around 1,000 net lines added per commit on a seven-day average, per Drew Breunig's read of Jordan Alberts' data — roughly two orders of magnitude above a human. Fred Brooks' Mythical Man Month puts a good engineer near 10 lines a day. Antirez, doing his own math on Redis, landed at about 29.
Code got cheap. Nothing else did. Feedback didn't get cheaper, and the "eyeballs" that Linus' Law depends on haven't caught up to the volume of code arriving.
There is exactly one source of feedback that moves at the speed of generation: yourself. You prompt, you review, you use what you built. Which is why so much of this output is idiosyncratic tools for one person — Steve Yegge's Gastown, Jeffrey Emanuel's FrankenSuite with its Rust rewrites of SQLite, Node, Redis, Pandas, NumPy, JAX and Torch. Philip Zeyliger's summary: everyone is building a software factory.
A tool built for exactly one user, by that user, guided by their own taste, is not a failure. It's a workshop. It's Sarah's house.
The problem starts when the same method gets pointed at a product.
Michal Malewicz describes the shape of it better than any roadmap: one idea spawning another, then a side-idea on the side, until you're running 17 terminal windows and 12 subscriptions at $200 a month. "You feel like you're BUILDING. Like BUILDING IS YOUR LOVE LANGUAGE." Technically, you are building. Every stupid idea.
He's blunt about the cost. It took him a year to work out what to kill rather than build, and he says the FOMO the industry pushes is unhealthy — that we're spending cognitive currency at an alarming rate on a technology that moves faster than we think.
Then he says the quieter thing. AI is making him dumber. Not because the output is bad, but because when you can do a lot, you do a lot. More hours prompting, fewer hours thinking. Skills atrophy against the hope that you'll never need them again. He asks designers to remember the day Figma went down and half the design world stopped working. What happens when the model goes down for a day?
Now hold that against the field-level statistic: nearly nine in ten organizations use AI, yet most still haven't rewired their workflows around it. The features arrived faster than any structure to hold them.
Which brings back the 2008 letter, "If Architects Had To Work Like Web Designers," which circulated for years with no name attached. A client writes to an architect asking for a house with somewhere between two and forty-five bedrooms, drawn so bedrooms can be added or deleted easily, with a complete blueprint set before any real design work begins, and the house under roof within 48 hours of approval.
It was funny because it was impossible. It isn't impossible now. And what disappeared along with the cost was the friction that used to force somebody to draw a plan first.
So here is the retrofit, applied to apps you have already shipped. It is not a rebuild.
Before any feature, define the object model. Not the database schema — the shared one: the real nouns your product works in, and the verbs scoped to each. A project can be archived. A document can be shared. An invoice can be approved and paid but never archived, because your business doesn't work that way. Those constraints are the useful part, and they exist only if someone writes them down.
This matters most at the moment of conjuring. When a feature offers to "create a follow-up," that request has to land on something real, with a shape and a persisted home. If "follow-up" isn't an object anyone defined, you have just let that feature build a room that exists on no floor.
Write the noun list before the feature list. Check every create, update and delete against a real object. Publish one page — objects, verbs, relationships — signed off by engineering and design together, so both sides build against the same world.
Above the output sits a smaller library: the interaction patterns. Suggest, draft, summarize, classify, plan, confirm. Each is a named behavior with defined states running from entry through in progress, result, recovery and dismiss.
Most teams have 40 features and zero patterns. That means forty people independently deciding what happens when something fails, and producing forty different answers. Failure isn't an edge case in AI features. It's the default weather. Design it once at the pattern level and every feature inherits it.
The library should govern what a feature shows as strictly as it governs how it behaves, and this isn't an imported discipline. Engineers build controlled languages constantly: an enumeration that permits four values and rejects the fifth, an interface contract that throws on any payload off its schema. They narrow what their code is allowed to say because the alternative is every caller guessing.
A design system is the same move aimed at output, the way Simplified Technical English keeps aircraft manuals unambiguous. A feature doesn't get to render a free-floating prose blob. It returns a structured payload that resolves to components you already own.
One controlled language, spoken everywhere, no bespoke surfaces invented per feature.
The Figma team's framing is the honest one: getting to a prototype is easier than ever, and a polished prototype doesn't answer whether this direction is the best or just the first, whether it feels like you, or whether it's worth building at all. Prototypes that don't use your design system feel launchable and quietly generate the work of unwinding them.
There's a version of this argument that ends in gatekeeping, and it's worth avoiding. The designer Max Kolomatsky has been picking amateur flyers off New York lampposts since 2023 — a dog walker's ad, a hand-drawn sign for a cleaning service — redesigning them for free and putting the new version back up. One was a handyman's, made of business cards taped together because cards were the material he had.
Kolomatsky condensed it to what mattered, added color and one clear illustration, and replaced the fragments nobody on the street was going to piece together. The handyman never asked. He likely never found out.
No AI in that story. The gap between an amateur flyer and a good one wasn't money. It was someone applying design knowledge on the amateur's behalf, for free, by hand, once.
That is the retrofit, in miniature and in full. Not fewer features, and not slower building. Somebody drawing the floor plan before the next room goes on — because the house is going to have strangers walking through it, and Sarah Winchester's only ever had her.