How to make Cursor, Codex and Claude Code designs not look like AI

Someone is building something real with a coding agent, the code is fine, and the interface looks as if it came from the same place as everyone else's. It happens in Cursor, in Codex, in Claude Code, in Copilot and in Windsurf, because the cause isn't any one product. Without direction a model writes the most likely page.

One person on Reddit had tried screenshots, detailed prompts and skills, then uploaded a template they'd paid for and asked the agent to use it. They got a weaker version of it. The threads we read were mostly from Claude users, and the guides for Cursor and Codex give the same advice, so here it is for all of them, in the order we'd try it.

What to do

1. Write the design down where your agent will read it. Keep hex values, font names, the type scale and spacing in a markdown file, and make sure every session loads it. Replies on this problem recommend this more than anything else. The table below shows where it goes in each tool.
2. Add design instructions, kept short. Name the defaults to avoid and how you want type, colour and motion handled. Anthropic publishes a frontend-design skill of about 400 tokens for Claude that does this, and people mention Impeccable for the same job. Codex and others have community equivalents. Read any skill before you install it.
3. Give it real values from a reference. An image leaves the model estimating. Some people have the agent describe a reference in detail and work from that description. Better still is giving it the page's real values, because then nothing has to be estimated.
4. Ask for cheap options first. Ask for several plain HTML mockups in different directions, pick one, then build. It costs little, and it's far easier to react to three options than to describe the one you want.
5. Say how far a change should go. An agent takes latitude from how you phrase things. One poster asked to restyle a single section and watched it spread across the page. Say which part changes and which parts mustn't.
6. Work on one component at a time. Finish and lock the design before building more on it. Each round of "make the whole thing better" tends to pull it back to the average.

Where the design rules go

AGENTS.md is the closest thing to a shared standard, so if your team uses more than one tool, keep the design in one file such as DESIGN.md and point each tool's file at it. These locations change, so check your tool's docs if something isn't being picked up.

Why the template didn't work

Asking an agent to use a template and pointing it at the code are different requests. If it only sees the result, it rewrites from memory and the rewrite comes out average. Copy the template's real files into the project and tell it to build with those components rather than recreating them. Then ask it to change as little as possible.

Design first, or code first?

Some people do the visual work elsewhere, in v0, UX Pilot, Figma or Claude Design, and bring it to their agent to implement. This works if you already like working in those tools, and it does separate two jobs the model tends to blur. The handover still loses detail, so a written set of tokens is worth keeping either way.

The honest limits

Not everyone thinks this is solvable with prompts. Some say to hire a designer, and if your brand is the product, they're right. These steps stop it looking generic. They won't give it an idea of its own.

Where plagio fits

plagio is an MCP server, and we make it. Any agent that supports MCP can call it to read a public page in a real browser and get back the markup, tokens, hover states and motion. It answers the reference step directly: you say "build this like that page" and the agent works from measurements. You'd still want the design file alongside it. If you're weighing tools, here's how the main design references compare.