Coinbase fed its design system to the agent and cut token costs 22.5%

The tools stopped guessing. For two years, AI design tools made pretty screens from a blank prompt. Now they read your tokens, your components, your code, and the little markdown file where you wrote down your rules. That flips the job. The thing you feed the model matters more than the thing you type into it. Let me catch you up.

The deep cut

• The system is the prompt now. Coinbase and Figma both moved the design decision into tokens and Code Connect, not the chat box.
• Structured context beats a longer prompt. Coinbase's Code Connect cut token cost 22.5% and killed icon hallucinations overnight.
• A blank prompt returns the average of the internet. Anshu Chimala got one-of-a-kind pages only by forcing the model off its defaults.

Why the bg-indigo-500 button keeps showing up

Ask Claude, Cursor, v0, or Lovable to build a screen and you get the same thing: Inter font, centered hero, three rounded cards, and an indigo button. Tailwind's creator publicly apologized for that button, joking it was "leading to every AI-generated UI on earth also being indigo." That is what the model does when nobody tells it otherwise. It predicts the safest token at every step.

This is not a taste problem you can fix with a better prompt. A CHI study of 227,000 website screenshots found layout distance between sites fell 44% over the 2010s, tracking framework adoption. AI speeds that up. The model regresses to the mean because the mean is what it was trained on.

Feed the model your rules, not a paragraph

The fix is plumbing. Two things landed that make it work now. Token adoption jumped from 56% to 84% of teams in a year, and the W3C shipped a stable token format in October 2025, backed by Adobe, Google, Meta, and Figma. So your tokens can travel as real data, not a moodboard. Second, the tools speak design system natively through MCP servers that let an agent read your components and variables straight from the file.

Coinbase proved the payoff. Technical Lead Erich Kuerschner ran Code Connect against the same design three times, and it cut token cost 22.5% and time to implementation 22%. Better, agents stopped inventing the Stepper component out of two progress bars, and the icon hallucinations "disappeared overnight." The context did the work the prompt could not.

Structure kills guessing, but it does not create taste

There is a split worth holding in your head. Kuerschner found that agent skills shape how the agent builds, while Code Connect gives it the right starting context. Both matter, and one does not replace the other. His team writes a COMPONENTS.md and a CLAUDE.md file, the "tribal knowledge" a screenshot cannot carry: hierarchy rules, spacing rhythm, tone, what not to do.

But a clean system reduces inconsistency without reducing sameness. Those are two different problems. Anshu Chimala, who ran R&D design at Apple for 12 years, showed why the model stays bland: asking for "something totally unique" still returns the same purple gradient and the same pottery metaphors. He had to force real variety from outside, seeding the model with a random string or a specific reference like an industrial control panel. The point of view still has to come from a person.

What actually changes on your team

The daily loop shrinks. Instead of wireframe to mid-fi to hi-fi to handoff to wait, you describe a screen and watch it generate inside your visual language. Figma's own team rebuilt their working process around this, shipping generative plugins in two months with no PRDs, starting every design as a working prototype. Rogie King put it plainly: code on the canvas is not so different from design on the canvas.

The catch is who owns the system when it drifts. Tokens, MCP configs, and markdown playbooks go stale the moment nobody maintains them, and a stale system feeds the AI wrong rules with total confidence. Xinran Ma, who builds design sandboxes for teams, notes the teams doing this well have kept iterating on theirs for months, some over a year, often with a paired designer and developer assigned to it. The auditing work does not vanish. It moves to catching the ghost component, the checkout flow that skips the payment step and looks plausible anyway.

Three questions for your team

1. Are our tokens machine-readable data yet, or still a Figma file a designer opens once a sprint? If it is the latter, the model is guessing every time.
2. Who owns the playbook and the MCP setup when the design system changes? Name the person, or plan for it to go stale and mislead.
3. Where in our review loop does a human push against the model's default look? A clean system gives you consistency, not a point of view. That is still your job.