--- title: "When you can build anything, what's worth building?" description: "Codex and Claude Code can ship a first releasable version of practically anything. The constraint isn't capability anymore. Picking what's worth committing to is the new hard problem." date: 2026-06-02 image: "/uploads/2026/06/whats-worth-building.jpg" categories: - "AI" - "Product" - "Method" --- Most of what used to make my work hard has quietly dissolved over the last year. Codex and Claude Code can take more or less anything I can describe and turn it into a first releasable version. Not a sketch. Not a wireframe. A real app, with a real backend, real data, real UI, running on a real device. The technical constraint that defined most of my career is basically gone. Which has left me sitting with a question I didn't expect to have to answer. If I can build almost anything, what's actually worth building? ## Twenty apps, two launches I've built more than twenty apps in the last few years. Mobile apps, web apps, browser tools, internal tools, experiments. Some came together over a weekend. Some took longer because I cared more. I've released two. I quietly pulled some of those back down. The ones still up haven't moved much. I haven't done much to push them since launch week. I've [written about how fast the building part has become](/thoughts/two-days-to-build-what-used-to-take-months), and that's still true. The thing I keep noticing about my own behaviour is that building wasn't the hard part. Committing to one of them was. ## A v1 is fun. A real thing is not. Starting a project is mostly play. The first weekend is exhilarating. New repo, new icon, new logo, new mockup, working demo by Sunday night. You can feel the thing wanting to exist. Then comes everything else. Supporting users. Fixing the edge cases you didn't want to think about. Finding a name that isn't already taken and a domain that isn't already parked. Marketing. Distribution. Talking about the thing every day, week after week, for months, when it stopped being fun in month two. The cost of starting just collapsed. The cost of finishing didn't. ## The shelf is full The App Store is saturated. Every category has hundreds of competitors and most of them are abandoned, which doesn't really help, because the user still has to scroll past all of them. Every good domain is taken. Every good brand name is a defunct startup or someone's holding page. Building has never been cheaper. Getting noticed has never been more expensive. I [wrote about that gap a few months back](/thoughts/building-is-easy-standing-out-is-hard) and it has only got more pronounced since. ## What happens when people just ask AI? This is the shadow question behind all of it. If I'm honest about how I actually use my devices now, I don't open apps as much as I used to. I ask. I describe what I want, and a model does most of it. Booking, summarising, drafting, planning, looking things up, deciding what to cook, working out a route. The interface has moved from a grid of icons to a conversation. So what happens to apps? Some categories get absorbed straight into the assistant. Calculators. Currency converters. Simple utilities. Generic content tools. If a model can do it well enough in line, the app version stops being worth installing. Other categories don't, at least not anytime soon. Apps with real workflows. Apps with proprietary data. Apps with hardware integration. Apps with network effects. Apps where the brand and the feel and the specific design are the product. Those stick around, but the bar for them being worth installing is much higher than it used to be. So before I commit to building something, the question I keep coming back to is which side of that line it falls on. If a user could get most of the value by just asking, I'm building a worse version of a feature that's already inside the assistant they actually use. ## Where the moats might be I don't think there's only one answer. But three places feel like real ground to me right now. **Physical AI and hardware.** Atoms move slower than bits. Manufacturing, supply chains, certification, regulation, distribution, support. All of it is hard, and all of it gives a product time to breathe before anyone can catch up. The software part might be a commodity in two years. The robot, the sensor, the device on someone's kitchen counter, isn't. I [wrote about this recently](/thoughts/physical-ai-the-robot-revolution-starts-now), and the more I work on it the more it feels like the most defensible layer left. **Systems that combine many things.** A single feature is easy to clone. A system that wires together hardware, services, data, models, and people is not. The moat is in the integration and the orchestration. Anyone can rebuild one node. Almost nobody can rebuild the whole network. **Great, simple consumer interfaces.** AI doesn't dump out taste by default. Most AI-built interfaces look the same and feel the same. They're forgettable. A really considered interface for an everyday problem, with a strong point of view, a clear voice, and a brand that actually means something, is still rare and still valuable. That's an opening for designers, not a threat to them. ## What I'm doing about it I don't have a clean answer to any of this yet, and I'd be lying if I tried to write one. What I'm actually doing is staying close to the work. I'm building for clients. I'm putting time into Minutemailer, my side hobby startup, because I like working on it and I use the product myself. I'm testing a handful of small things to see which ones earn more of my time. I'm prototyping physical AI and robotics in software, using open-source tools like [Genesis](https://genesis-embodied-ai.github.io/) and parts of the NVIDIA stack. Half of it is research, half is just play. My Mac mini can't run anything that needs CUDA, which is genuinely annoying right now, but it's also forcing me to be more deliberate about what I try. I'm having fun. I'm learning a lot. I'm not committing to a single thing yet, and I think that's fine for now. Sitting in the question is part of the work. ## The new hard problem Building used to be the gatekeeper. If you could ship the thing, you were already further than most people. That's gone. With AI, plus any technical instinct and any background in design or branding, building is genuinely easy. What's hard now, maybe harder than it's ever been, is choosing what to build at all. Picking the one thing worth your next year. Saying no to the other twenty ideas, including the ones you could ship by Sunday. Trusting that the commitment is the work, not the keyboard. I've [thought about this before in a different register](/thoughts/what-problems-are-worth-solving). It feels more urgent to me now than it did then. The supply of possible things is effectively infinite. The supply of attention and energy and time inside a single person hasn't changed. So before the next repo, the next domain, the next icon, the next launch tweet, I'm trying to sit with the question a little longer. When you can build anything, what's actually worth building? I don't think you can outsource that one to an agent. Not yet, anyway.