You let the agent decide. The edit is still the job

A solo builder shipped a fifth product in three days, told Claude "you decide" for the entire interface, and got back a terminal look — near-black, amber accent, zero radius — that users thought he had hired a designer to make.

I have run the same test on smaller surfaces and felt the same pull to say "wow, it has taste now." I wrote that why curation is the job now — you kept 116 of 400 AI concepts and the cut was the work. "You decide" does not change that math. It just moves the cut later.

"You decide" works when your system carries taste and you keep the edit. The agent proposes, you approve.

Why does "you decide" feel like taste got automated?

It feels automated because the agent finally makes connected choices instead of one-off parts.

Earlier models executed taste only if you spelled it out — a reference, a spacing scale, a color code. Left alone, they averaged into the blue gradient on dark that the default AI aesthetic warned about: safe and forgettable.

Now the model makes tradeoffs that hold across pages — where to be loud, where to stay empty, what to repeat. That reads as taste. It is pattern consistency with more confidence. Keeping a product coherent for a week still needs a human holding the line.

What actually made that Claude design work?

Consistency plus a short edit.

He did not ship raw output. He fixed an alignment bug twice, moved one call-to-action, and left the rest. The typewriter hero, scanlines, amber on near-black held because it was applied the same way on every page. Users read that as intent.

That is the lesson from consistency as the trust signal — users trust what stays the same more than what pops. You also gave it a narrow job — "make the interface" — not permissions or memory. That boundary is why it felt designed.

How do you make "you decide" work without losing your system?

You make your system readable to the agent before you say "you decide."

An agent cannot respect a system it cannot see. It will invent buttons, spacing, and imports unless you give it the real ones.

That is what Figma's guide to the Figma MCP server is for — it lets agents read your components, variables, and layout directly from the file instead of guessing from pixels. Figma's measured results with Code Connect show why that matters: teams that mapped Figma components to real code imports cut median task time by 19.6%, improved code quality by a full point, and reduced token usage by 29.5% because the agent stopped fabricating.

My setup is simple:

Lock the system first. One button set, one radius, one type scale, one spacing scale. Variables named, components mapped. If the agent cannot read it, it does not exist.

Say "you decide" inside the system. Try "explore a direction using our components and tokens, produce frames on the canvas" — not "make a landing page." The first tells it where to decide. The second tells it to invent.

Edit for two things. Is the primary action obvious without reading? Does the same thing look the same on the second page? The checks from Nielsen Norman Group's guide to minimizing cognitive load still apply: extraneous load is the tax you cut first.

If the agent invents a new component, that is a flag, not a gift. Map it or delete it.

When should you not let the agent decide?

Do not let it decide anything that affects trust.

Letting the agent pick a palette is low risk. Letting it decide what memory to keep, what empty state to show, or whether an error needs an undo is not. Those are product decisions, not visual ones.

I use a simple split:

• Let it decide: visual variations within the system, copy tone within your voice, layout options that use fixed slots so content does not push controls around.
• You decide: what the product remembers, what it forgets, what the default does when data is missing, and where approval lives. Keep those in plain language, one tap to undo, and scoped so a Topic from one project does not leak into another.

The builders winning in 2026 are not the ones who prompt more. They run a tighter loop: agent proposes inside the system, human edits for coherence, agent adjusts. "You decide" proves the loop got faster. It does not prove you can leave it.

Try it on one surface — an empty state or settings page. Lock your components, say "you decide," then time the edit. Under ten minutes and it feels like your product? Your system is portable. Over an hour? Fix the system, not the prompt.