The Work Is Deleting, Not Generating

Generating things is no longer the hard part. Describe a screen and you get a screen. Describe a feature and you get a feature. The constraint that shaped design work for twenty years — that making things was slow and expensive — is largely gone.

What replaced it's worse to sit with, because it doesn't look like work. The bottleneck now is judgement: deciding which of the things you can have, you should actually keep.

Twenty-Two Seconds of Nothing

I built a boxing game for the mobile web. The first version opened with a splash screen, then a mode select, then a fighter select, then a tutorial. Every one of those was reasonable. Every one existed in games I had played.

It took twenty-two seconds from tapping the link to throwing a punch. On a phone, from a link someone sent you, twenty-two seconds is an eternity. People were leaving before the game started.

I deleted all four screens. Time to first punch went to six seconds. Nothing was added. The game got better by having less of itself in the way.

The Model Will Happily Build All Four

Here's the part that actually matters now. If I had asked for a splash screen, a mode select, a fighter select and a tutorial, I would have got all four, quickly, and they would have been fine. The model has no opinion about whether they should exist.

It can't tell you that a fighter select is dead weight when there are three fighters and the first one is the obvious choice. It can't tell you that a tutorial is an admission the controls aren't obvious. Those are judgements about a specific product for specific people, and they only come from watching someone use the thing.

Cheap Generation Makes Deleting Harder

There is a trap in this. When something took two days to build, cutting it was easy — you could see the cost of keeping it. When it took ninety seconds, it feels free to leave in. It's not free. The user pays for it in attention every single time.

So the discipline I've ended up with is the opposite of what the tools encourage. Generate freely, because it costs nothing. Then be ruthless, because the thing that costs is the shipped surface area, and that hasn't got cheaper at all.

The full build — sprite pipeline, opponent behaviour, the parts AI genuinely couldn't do — is in the Ring-Rival case study.