← All posts

Why the same model designs better on a canvas than in your repo

We build our websites with coding agents. A pattern kept showing up: a page drafted in a visual design tool came back composed and confident, while the same page built in a coding session came back with cramped sections, a flat hierarchy and a layout that looked like every other landing page. In both cases the work was done by the same model. So we asked why, and searched for what others have found.

None of what follows is a controlled benchmark. Anthropic doesn't publish how Claude Design works inside. The sources are product announcements and hands-on tests by individual developers. They agree with each other, and with what we see in our own sessions.

The canvas lets the model see its work

Claude Design renders every change on a canvas, and the model iterates against that render. In a coding session the model writes Svelte or JSX plus Tailwind classes and judges the result by reading the code. Spacing, alignment and visual weight are hard to judge from code and easy to judge from pixels.

One tester gave the same landing-page prompt to both tools and got similar output. The difference was the loop: the coding agent made them "re-prompt, re-render, and inspect in your browser every single time", while the design tool got there about ten times faster. The person becomes the model's eyes, and every look costs a round trip.

Without a direction, the model falls back on the average

If nobody names a look, the model reaches for the one most common in its training data: tasteful but anonymous. Anthropic's frontend-design plugin for Claude Code exists to break that default. Before any CSS, it makes the model commit to a small design system: four to six colours, typeface roles, ASCII wireframes and one signature element. Then the model critiques that plan against "the generic default you would produce for any similar page" and rewrites whatever fails. The plugin also bans three looks the model kept producing unless the brief asks for them.

Several guides call this the biggest single improvement to design quality in Claude Code. A comparison of design skills found the results vary a lot from skill to skill. The more opinionated ones produced stronger visuals, and they used much more compute to do it.

A canvas starts from the layout; a repo starts from a diff

Claude Design works top-down. During onboarding it builds a design system from your codebase and design files. Guides recommend getting the structure right first, then the content, then the polish, with comments on the canvas for small edits like padding or spacing.

A coding session starts from files that already exist, and good engineering practice pushes the agent toward the smallest change that works: reuse the component, don't touch what the request didn't name. That's the right instinct for code and the wrong one for a page whose composition is the problem. The result is the old layout with tweaks. Rules about imports, module layout and error codes also take up the model's attention before taste gets any of it.

Where the coding session wins

The reports agree that inside an existing repo, the coding agent wins time-to-production: edits across many files, the real component library, git, and build and test loops all happen in one place (MindStudio). A canvas prototype still has to be ported into those components, and the port is work a coding agent does well.

The same model, three differences

The model is the same in both tools. What changes the result is whether it can see its output, whether it was given a look to aim for, and whether it started from the layout or from a diff.

Sources