Building Got Cheap. Judgement Didn't.

On Saturday I sent one message: "do all 10 except number 9". It was a reply to a list of ideas for Split Race, the racing game on this site.

By the evening there were three cars with their own physics, a third circuit, night racing, rain, recorded engine sound, a championship, ghosts you can race, and tilt steering on phones. Three pull requests. 557 passing tests. All of it in production.

I have been building software for a long time, and that sentence still looks wrong to me. It is also not the interesting part.

The cost of building fell off a cliff

Over the last two weeks this site gained a cron explainer, a text-to-art generator, a tiny language model trained on this blog, a quiz show about prompting, a three-lane runner and an interactive film. Each one went from idea to live in about a day.

None of that is because I got faster. The building is done by an AI agent working in the repo, and my part is mostly saying what I want and whether I got it.

So the price of making something has collapsed. Not halved. Collapsed. That changes what is scarce, and the scarce thing is not where most people are looking.

What did not get cheaper

Three things took exactly as long as they always have.

Deciding what is worth building. The list of ten ideas took minutes to produce. Choosing which nine to build, and that the tenth (replays) was not worth it yet, was still my call, made the old way.

Knowing whether it is any good. The interactive film, a meeting you play as the middle manager, was built in a day. Filmed shots, six voices, lip-sync. A day later I hid it. Nothing was broken. It was just bad, and no test could have told me that. I had to watch it.

Knowing what "right" means. When the racing game got three cars, the leaderboard needed a rule. Should a lap in the fast car count against a lap in the slow one? A wet lap against a dry one? The code could do any version. Picking the fair one is a judgement about what the board is for, and it is now one line in a pull request description that took longer to think about than to write.

Cheap building makes more of the wrong thing

This is the part I did not expect. When building is expensive, the cost filters out bad ideas before anyone writes a line. When it is cheap, the filter is gone, and the bad ideas get built too. Quickly, and well tested.

Three weeks ago I found that five personal-best boards on this site had never received a single score. Not because nobody played. Nothing had ever sent one. Each board was built properly, with tests, an API and a database table, and the games around them were never wired up to post. Every part worked, and together they did nothing.

And when we went back to wire them up, "one save call per game" turned out to be wrong for two of them. One game never ended, so it had no score that meant anything. The other was a translator, where a score would just reward pasting. The fix was not code. It was asking what a score is.

Ten games here have sound that was chosen by measurement, not by ear: no clipping, no silence at the start, the steadier take of two. That is real rigour, and it still cannot tell you whether an engine sounds like an engine. Someone has to listen.

For the person doing the work

If building is what you sell, the price just dropped. That is not a reason to panic. It is a reason to be honest about which part of your job was the building and which part was the judgement.

The judgement is the part that survives: knowing what good looks like before you see it, noticing the thing that works but is pointless, saying no to the easy feature. Those used to be hidden inside the typing. They are not hidden any more, and they are worth a lot more now that the typing is cheap.

It is the same point as Productivity Is Not Leverage, only sharper. Output was already a weak signal. Now it is close to free.

For the person deciding

If your team can now build five times as much, you will get five times as much. Most of it will work. Plenty of it will not be worth having.

The bottleneck has moved from making to reviewing, and most organisations are still staffed and measured for making. Count what gets shipped and you will get more shipped. Ask who decided it should exist, and who checked it was good, and you will find out where your real capacity is.

The people worth paying more are the ones who can look at a working feature and say "this is bad", or "this should not exist", and be right.

The takeaway

Building got cheap. That part is real, and I am not going back.

But the hard parts were never the building. They were deciding what to build, knowing what good looks like, and having the nerve to throw away something that works. Those cost the same as they always have, and they are now most of the job.

If you want to test your own instincts on the prompting side of this, Prompt Master is fifteen questions about exactly that kind of judgement.