Anyone can make an app now. Describe it and an AI agent writes the code, the screens and the tests. We have done it many times over: the games on our games page, our novels and Garden, our HR product. Getting something that runs has stopped being the hard part.

Getting something good has not. Everyone can buy the same tools, and only some people make furniture you want to keep. With AI, anyone can make an app, but only the artisan will make a good product.

Where the work moved

When an agent builds one of our games, it usually works the first time we open it. The work has moved to everything after that.

Two of our games installed fine on a phone and then refused to open. Nothing in the code looked wrong. The cause was one line in the install file that made every game on our site claim to be the same app. Nobody found it until Mas tried to install them on a real Android phone.

Another game looked perfect on its first screen, but its battle pictures were missing. The code that loaded them assumed the game sat at the top of the site, and the first screen was the only one the checks had opened.

Then there is the work no test can find. Miso’s Bookshop and Paper Reels had icons that looked lovely on a home screen, and the other games did not. Nothing was broken. Someone had to look at a phone and say so.

Taste becomes rules

Each time we caught something like this, we wrote it down for the agents: how an install file names its game, that an update waits for the player to tap it instead of reloading in the middle of a round, that sound is on from the first tap, how an icon sits inside the shape a phone cuts it to. Most of those rules now come with a script that checks them, so the next game cannot repeat the miss.

We do the same with writing. Every time Mas marks a sentence in a chapter as the kind to avoid, the example goes into a taste book. The editing agent reads it before it touches the next chapter.

That is how an artisan’s knowledge gets into the work: one specific example at a time.

What the rules cannot do

A rule has a limit, and we learnt it the hard way. Show the writing agent the checklist and it writes to pass the checklist. The faults disappear and the chapter is still flat. As Mas puts it, a measure you use in production stops measuring. So the taste book goes to the agent that edits and never to the one that writes.

We also tried handing judgement to AI readers: agents that read a whole outline cold and reported what was wrong. Two rounds of them cost a lot and caught less than one sitting where Mas went through the outline chapter by chapter. We stopped using them.

Judgement also decides what not to build. We put AI into Garden’s hiring tools, then took it out when the EU AI Act classed AI that ranks job applicants as high-risk. An agent would have kept improving the feature. Deciding it should go was a person’s call.

The gap is closing

Look at those rules again and you can see where this goes. Every rule we write hands a piece of an expert’s judgement to the agent, and the models get better at the same things on their own every few months. The distance between an ordinary build and an expert one is getting shorter, and AI is what is shortening it.

That does not make the artisan unnecessary. It moves the work. Once the agent knows how an install file should look and where an icon should sit, what is left for us is the part nobody has written down yet: what the game is for, what would surprise a player, which idea is worth a month. The people who do well will keep moving to that edge. They will need to be more creative, and quicker to learn what they did not know last year.

This post was drafted by Claude Opus 5.5 from Mas’s notes and our project records.