Anti AI slop design
Your agent writes code that works and pages that look like everyone else's. Why AI output turns generic, why longer prompts stop helping, and what fixes it.
It works. It looks generic.
The spacing is whatever the model saw most in training. The type is the default. The colours are the framework's. The button on one page is not the button on the next. Nothing is broken, and nothing looks like you.
The prompt treadmill.
So you write longer prompts. You paste screenshots, hex codes and Tailwind classes. It works for a page or two. Then the next chat drifts and you paste it all again.
Prompts are instructions. They do not carry over to the next session, nothing enforces them, and they fade as the conversation grows.
Why AI output looks like everyone else's.
No design judgment
A model has seen a great many pages. Asked for one, it writes the average of what it has seen: competent, safe and like everyone else's.
No memory of your brand
Each session starts from nothing. Your colours, type and voice are gone unless you paste them in again.
No check
Nothing holds the output against your brand before it ships. Off-brand work gets through until someone notices.
What fixes it.
- Your brand, described once.
- Read by the agent while it works, the same in every session.
- The same brand decisions behind every page, so the tenth page is built from the same brand as the first.
- A check on what comes out.
Where Designless comes in.
Describe your brand once. Every page your agent builds in Claude Code, Codex or Cursor reads that brand while it works, and Arbiter checks the work against your rules, on request or before every delivery when your brand requires it. Any other MCP agent connects through LESS MCP.
See agents work in your brand on one canvas