The last five percent (ft. building with AI)
Introduction
Agentic coding models and harnesses have become so effective that a large share of new software is now written by agents. For a good reason, it’s faster, code is at times higher quality than if human written, and there are seemingly no downsides. We thought.
But the majority of new Software feels average. Very often you can feel that if was created to check a box instead of being crafted with care to bring delight to the user.
In the following I focus on frontend development, because it is easier to explain based on that. The same applies to all other forms of development and research.
Back then
None of this has to do with AI per se, at all. It’s about the organisatorical structures and processes that often fall when development goes too quickly. AI produces impeccable results, if done right. Let me explain.
Years ago we used to start at an idea, sketch it out by hand or on a whiteboard, then refine the user experience and exact optics in Figma, then start development and iterate until it really looked like the design wants it to look. The designers’ job is it to design a beautiful button, the developers’ job is it to create that button to look exactly like the designer intended. Things took time and consequently it (usually) means that more thought has been put into.
Now
Now many teams go from the feature idea directly into code, because coding is easy now, right? You describe the idea and observe the outcome, then give feedback. This quickly gets you an MVP, but not something that excites users. There’s this well known saying that getting the last 20% right takes 80% of the work, and here I would go even further and say that getting the last 20% right takes 95% of the work. And the reason for AI slop is getting 80% there and calling it a day.
The idea of jumping from idea to code is not new, and many experienced (for example) web developers have done that. But it’s harder to get right, because more than ever it requires having a very strong mental picture of what you want to create and how it should look like. A very strong vision that you keep in mind at all times. If you have a vague idea, you’ll see the result and think “okay that works”. It “works”, it always does. But it’s missing the 20%. It takes people who have a vision so strong that they are willing to go through extra iterations until they are really exactly there, at their vision, and not stop before that. We live in an age where we expect everything done quickly. If you’ve watched The Playlist, you know that scene where Daniel Ek repeatedly asks Spotify’s engineers to re-iterate on their playback delay until it’s near instant. This is exactly the same thing.
AI slop
Defining “AI slop” in product development is hard if you try to nail it down on single things. Sure, there are giveaways, certain patterns AI likes to use (the left border on a rounded card, gradients, etc.) but AI slop is more than certain patterns. Something can simply feel like it has been created without care. You can feel when something didn’t have craft and thought put in.
For product designers this might be perfectly clear. For anything else, it’s much easier to notice the concept of AI slop in written text. Because everyone has read a text with virtually value, a word salad without soul. And the same applies to UI, although many people don’t see good/bad UI directly, but in the long term, everyone feels it subconsciously. So when I say “slop” in product development, what I mean is creating something that ticks the boxes, but doesn’t tick them with care. And AI make it really easy to do exactly that, tick the boxes. And it makes it just as hard than ever before, to tick them with care. And idea, craft and effort will always shine out.
Here’s a video recording of an animation in Apple’s music app. I know for a fact that this has been implemented with AI. It’s hard to describe, but again, you feel it. It’s just too … linear?