LLMs Didn't Change What's Hard About Building Software. They Changed What's Easy.

There is a version of the AI story that goes: everything changed. A better version goes: the bottleneck moved.

The hard parts of building software are exactly as hard as they were five years ago. Figuring out what to build. Deciding what not to build. Understanding who you are actually building for versus who you think you are building for. Having the conversation with a customer that reveals your entire roadmap was answering the wrong question. Being wrong about the market, updating your model, and deciding whether to pivot or hold. None of that got faster. None of that got easier.

What changed is everything that used to feel like work but was actually just execution.

Writing code is execution. Drafting a spec is execution. Generating copy for a landing page, sketching out an API contract, stubbing test cases, translating a design into a working prototype -- execution, execution, execution. LLMs are very good at execution. Not perfect, not autonomous, but fast and capable enough that the time cost of most execution tasks dropped by an order of magnitude.

The builders who are winning right now are not the ones who learned to prompt better. They are the ones who recognized the bottleneck moved and responded accordingly. When execution gets cheap, judgment becomes the constraint. When you can go from idea to working prototype in a weekend, the question is no longer whether you can build it. The question is whether you should. That question requires judgment, and judgment comes from talking to people, sitting with uncomfortable data, and thinking clearly under uncertainty. None of those have an AI shortcut.

Most founders and product teams have not updated where they spend their time. The calendar still looks the same: sprint planning, stand-ups, code reviews, design reviews. The rituals of an execution-bottlenecked world running inside a judgment-bottlenecked one. That mismatch is expensive. You are optimizing the thing that is no longer the constraint.

Here is what I think the right response looks like. More time in front of customers. More time asking whether the thing you are building actually connects to a problem someone is losing sleep over. More willingness to kill a feature that is technically impressive but strategically irrelevant. Less celebration of shipping velocity as an end in itself. Shipping fast is a capability, not a goal. The goal is building something that matters, and LLMs did not change what that requires.

The gap between judgment work and execution work is widening. Execution is becoming a commodity. Judgment is not. This is not a warning about AI replacing developers. It is a more uncomfortable point: the skills that were never really technical are now the ones that determine outcomes, and a lot of people who got good at the technical stuff did not spend as much time getting good at the other stuff.

You now have leverage over execution that would have seemed unreasonable a few years ago. The question is what you are doing with it.