Interfaces built with agents come out too wordy

The longer I design a screen with an agent, the more the screen turns into the conversation. Paragraphs explaining why a section exists, navigation to everywhere, labels that read like notes. An agent reads that fine. I get tired in about ten seconds.

It happened on several projects, so I know it’s not one bad session. I’ve written about why AI products have terrible UX, and this is one cause. The same habit shows up in agent replies, which I handled with a rule in my instructions files.

The screen ends up written for an AI

In August an agent built me a landing page with a lot of text on it. It read like it was written for an AI, not for a human. A visitor needs their pain pushed, their questions answered in the right order and clear buttons, with much less text. When I had the content rewritten it came back long again, and it left odd notes on the exact pages, which the frontend agent then carried into the code.

Genvas looked like slop

On September 19 I sent a screenshot of my media canvas app and said it looked like pure ai slop. I sent the agent to clean it with Impeccable, a design skill, plus a rule: never overexplain, as little text as possible, more nice elements with attention to detail, minimal and clean.

I want Genvas worked on together with an agent. In the brief I wrote, the agent reads and resolves my comments on the canvas, and shows where it’s working and when it’s finished. The same brief says the interface should be very clear, lightweight and optimized. A canvas like that has no room for paragraphs.

Explanation text is build context

Early October, Herdtrude, the agent app I’m building. I went through a batch of screenshots. The sessions list was huge and needed a show more, and the work patterns screen was ugly. The explanations were the worst part:

all of this overexplaining text is context flowing into interface no need

That’s what it is. The agent had the context from building the screen, and the context ended up on the screen. The landing page I wanted for the app got the same brief: stand out, little text, clear about what the app is.

A customer site full of factory notes

A game studio website was worse. In late October I told the agent it was flowing too much context from the factory to the customer-facing site, and it was too deep into one game. It should feel like a good professional game studio, but playful. I wanted a clean agent to sort the content and marketing properly. The hero was still too much text the same night.

Prose where the visuals should be

A game project’s interface has the same problem. The interaction and UI are bad, and the screens are slop with just prose. I want the whole interaction built in Paper first, so it feels like a proper indie game you can interact with. Over the last week of October I kept asking for the same things: less prose, more visual stuff, icons instead of text wherever possible, one coherent UI. Yesterday it was “just bunch of the text and prose wow”.

Impeccable out, unslop in

Impeccable has been my design pass for weeks, across several projects, and the slop keeps coming back. On Sunday I wondered how heavy it was and whether something better existed. I found unslop, which looks really good. That day I also went looking for a term for UX for AI models, and whether anyone works in that field.

I decided to drop Impeccable for now and start from unslop. From a trimmed Impeccable I want the direction of UX and wording, so we start doing good design right away, always on Paper first.

That became my craft skill. The visual system is adapted from unslop by Damir Sokolovsky (CC BY 4.0), and the UX and wording come from a trimmed-down Impeccable (Apache 2.0). The credit lives at the top of the skill file. Unslop is installed so it only runs when I call it, not on every task.

Polish doesn’t cut text

Yesterday I wanted the preview interfaces inside Genvas unified for image, audio, video and 3D. They open a poor preview window, and 3D and audio open nothing. It has to be on Paper first.

Then I ran unslop on the interface, proposals first. The numbered list came back and I went through it: no, no, no, no, then fix on every item from 5 to 14, with the settings checked too. Herdtrude got the same treatment, propose, then do it.

A polish commit had already gone into Genvas yesterday morning: tokens, sentence case, an 11px floor, a 4px grid, numbered labels. That’s real cleanup, but it changes how the text looks and says nothing about how much of it there is. What I saw afterwards was this:

it looks like too much prose and navigation etc its fine for ai agent but its hard to human its overwhelming

So the job went to a new fresh agent, with time to think about UX and UI. I also approved all the Paper boards for the media viewers and told the agent to build from there, with no placeholders or hardcoded values like the estimated price. I’ll see what the fresh agent cuts.