Skip to content
AI-Native WorkflowsDesign Craft

Defaults are the new template

When everyone generates UI from the same models and the same stacks, sameness stops being a design choice and becomes the baseline. Here is where distinctiveness lives now.

Take the last screen an agent helped you build and cover the logo. If it could belong to any of ten other products in your category, you have found the default. There was a stretch of web design when you could spot a site built from a popular template in about two seconds: the same hero, the same three feature cards, the same rounded buttons. Designers spent years learning to escape that look, and AI generation has brought it back at larger scale. An agent with no specification reaches for the most probable answer, and the most probable answer belongs to everyone. Making a product feel like itself is the part of design systems work I care about most, and defaults flatten it first.

Where the sameness comes from

Figma's post on design systems and MCP warns that without a strong system teams risk shipping interfaces assembled from the same pool of AI-generated parts, and that output may default to React and Tailwind unless the agent is configured for the team's own stack. Those are two layers of default, the look and the code, and both leak into the product, while the average designer in Designer Fund's 2026 survey now uses seven off-the-shelf AI tools, so seven sets of defaults feed the same habits. The sameness also settles in places nobody reviews: the default empty state, the default error copy, the default icon size, the default gap between sections. Each is small, and a product made entirely of small defaults reads as nothing in particular even when every screen passes review.

Language has defaults too

The style guide I keep for my own writing is a catalog of the same problem in prose, with nine patterns that make text read as machine-written, each paired with what to write in its place (the guide's anatomy has the detail). Because the guide loads into every session in the repo, an agent drafting for me meets the specific alternative before it reaches for the average. The fix for interface defaults uses the same move: name the default, then put the specific alternative where the agent reads it before it builds.

Three places a point of view becomes readable

Tokens. On 2026-09-27 this site moved to a generated color system with scoped surfaces and on-colors. Each hue scope restates the foreground ramp and hands its own fill to the spot illustrations, so an agent working inside a card on a colored background gets that hue's colors and contrast where it would otherwise get the page default.

Icons. The icon layer gives agents named sizes from 2xs at 12 pixels to 4xl at 64, four weights, color tokens and duotone controls, so asking for a small muted icon picks real values from a short list, and any prop left unset falls back to the library's default.

Type and measure. The design doc names its type scale step by step: heading-xl runs 56 pixels at weight 600 with a line height of 1.08 and tracking of -0.02em, and body-lg runs 19 pixels at a line height of 1.6. Paragraphs cap at 60ch, with the reason written beside the rule: past that width the eye loses its place on the return sweep. An instruction that asks for a clean, modern look describes the default and changes nothing, while a number with its reason attached changes the output.

Build outward from pages

A component pulled from a page with a point of view carries that point of view, and one invented in a vacuum carries none, which is why a practical guide on uxdesign.cc recommends designing full pages first and distilling components from them. The first output an agent returns works as a map of what normal looks like for that request, so read it, name what is generic about it, and give the agent the specific replacement for each generic part. Reading fast depends on taste, and what makes an interface feel polished lists most of what a default will not hand you.

Write the decisions down

With the logo still covered, find the three places on that screen where a decision about type, space, tone or content would make it unmistakably yours, and write those decisions into the place the agent reads first. Then do the same for component names, empty states and error messages, where defaults hide best. The writing takes an afternoon, and every later generation starts further from average.

Greg Sargent
Greg SargentDirector of Design Systems, Spring Health

I write about design systems, accessibility, and the way AI is changing how we build software.

Published October 26, 2026

Be the first to rate this article.