Taste survives when you say 'you decide' — if your system lets it

I told an agent "you decide" and got a clean dashboard I could have drawn myself. A builder I follow said the same and got something he wouldn't have dared to spec — users asked if he'd hired a designer. The difference wasn't the model.

I wrote that design taste is the new moat for AI startups. The proof is simpler now: taste is what ships when you stop describing it.

The "you decide" test shows if taste lives in your system or only in your head — and on September 3 a solo founder proved it ships when you get out of the way.

What does "you decide" actually test?

It tests whether your agent can make trade-offs that hold with real users.

Jacob Yang runs five AI products alone — no team, no designer — and builds with Claude every day. For his fifth, Tracetify, he gave the interface one instruction: "you decide." No references. No palette. What came back was a point of view: terminal type, zero radius, amber on near-black, scanline sweeps, a typewriter in the hero. He fixed two alignments and moved one button. Done.

You can read it in Jacob Yang's solo run of five AI products in a year on HackerNoon — three days from idea to live, with dated receipts for every claim. People asked if he'd hired a designer. He hadn't. He just stopped re-briefing taste line by line.

That's the test. Earlier models executed taste you handed them. This one had to pick where to be loud, where to stay quiet, and what to leave empty. It picked, and the picks held. That doesn't mean models "have taste." It means a short, opinionated brief beats a long hedging one when trade-offs stay visible.

Why does product feel still beat a polished screenshot?

Because users don't live in screenshots.

Lorenzo Alvaro put it well in Forbes piece on why product feel is becoming harder to copy: layouts get cloned, colors get lifted, screens get rebuilt — and copies still don't feel the same. People notice hesitation, risk, feedback, behavior under pressure. That's feel.

AI widened that gap. Templates made polish cheap. A screen can look finished and still read slow. When every app is clean, clean stops meaning much. The product that feels considered stands out more. Yang's amber terminal worked because it was consistent — every page kept the same constraints. Feel is consistency over time.

How do you run the test without losing your system?

You don't hand the agent a blank canvas. You hand it a system that says what truth is.

Mine is short. One page in DESIGN.md: palette, type, spacing, radius, two things we never do. One Figma library where names are real, variables hold tokens, components carry states.

What changed is the system can travel. Figma's TL;DR on how the MCP server brings design context into code says it directly: with MCP, an agent doesn't guess tokens from a screenshot — it reads components, variables, and Code Connect mappings. The file becomes the brief.

Run "you decide" this week without drifting:

• Write the one-page file. "Amber only on primary CTA, 8px grid, zero radius, muted grays for hierarchy."
• Make Figma readable. Real names, auto layout, variables not raw values, Code Connect mapped.
• Expose it. Connect the Figma MCP server so Claude Code or Cursor reads before it writes. Fetch context first, use tokens, no hard-coded colors.
• Give one direction. "Terminal-inspired, quiet where it can be, loud only at the decision."
• Review against the last two shipped screens. Fix one drift per ship and add the rule to the file.

With that, "you decide" isn't a gamble. It shows which trade-offs your system makes clear and which you still need to write down.

When shouldn't you say "you decide"?

When you haven't decided what good means yet.

If your taste file is empty, the agent returns the average of what it saw most. If your Figma file has "Frame 1337" and detached instances, MCP just shows the mess more clearly. The open brief amplifies the system you have — it doesn't replace it.

I also wouldn't use it where you need hard checks: checkout, permissions, empty states with real data gaps. Those need explicit constraints, not vibes. I wrote about this as let the agent decide but keep the edit — the prompt can be short, the edit is still yours.

Use it for direction, not certainty — a landing concept, a report card, a tip you can test tomorrow. Keep what feels opinionated, cut the rest.

Try it once this week on one small surface. One line of direction, one system to hold it, one review against your last ship. You'll learn where your taste is written down and where it's still in your head.