Google Stitch’s Design Is Weak. But Those 10 Seconds Were Real.

”Ship Before You Hype” Series #2

X Was Buzzing Again

“Google Stitch × design definition prompt = INSANE UI generation” “What Google Stitch’s DESIGN.md is really asking” “New Google Stitch Update is INSANE!”

Google Labs shipped a “vibe design platform.” Type text, get high-fidelity UI mocks. One-click Figma sharing. Free, 350 times a month.

I tried it. Looked at the output. Closed the tab.

The design is weak.

Same Prompt, Three Tools

Stitch, Pencil, Figma Make. Same prompt across all three.

The results were obvious. Stitch output looks “about right,” but every screen has the same face. Color choices are shallow. Whitespace carries no intent. Components are assembled from templates with no sense of “this product demands this layout.”

Pencil was better. Figma Make was better. Same prompt.

Stitch’s UI has no judgment. It doesn’t choose colors. It doesn’t decide placement. It pattern-matches “app-like things” and lays them out. Design is choosing. Stitch doesn’t choose.

Android by Default

There’s another fatal problem.

Stitch outputs Android-first UI. It’s a Google product, so this shouldn’t surprise anyone. But if you’re trying to use the output as a starting point for iOS or web, nothing fits.

Navigation structure, button placement, typographic feel — everything comes out in Material Design grammar. If you’re building for iOS Human Interface Guidelines, it’s not even a rough draft.

A mock that can’t serve as a starting point isn’t a mock.

Is DESIGN.md Really That Good?

The most hyped feature on X is DESIGN.md.

“Describe your design system in plain text and feed it to an AI agent.” Good idea. Define design rules in markdown and enforce consistency in generation. Makes sense.

But this is philosophy, not implementation quality.

Writing a DESIGN.md doesn’t fix a weak generation engine. No matter how precise the input, if the converter’s resolution is low, the output’s resolution stays low.

My own pipeline puts judgment at every layer — from design system definition to asset selection to layout decisions to output verification. If a single DESIGN.md file could guarantee consistency, designers wouldn’t exist. They do for a reason.

10 Seconds of Truth

After all that criticism, one thing deserves recognition.

Figma sharing is absurdly fast.

Copy generated screens from Stitch into Figma. Eighteen screens. Ten seconds. Editable layers. That’s genuinely impressive.

No other tool matches this speed. Getting Pencil output into Figma takes work. Figma Make completes the loop internally, so the comparison is different.

Stitch’s value isn’t in generation quality. It’s in the speed of the Figma bridge. But if the cargo crossing that bridge is defective, speed alone means nothing.

Sharpen Your Own Tools

If Stitch won’t cut it, what should you use?

I’m not giving that answer.

This series doesn’t hand out blueprints. It shows you where the walls are. How to climb them is your job.

One hint: if you want better AI design output, don’t rely on a single DESIGN.md. Add more layers of judgment. What’s the visual reference? What context informs asset selection? What should the whitespace communicate? Decide all of it yourself and bake it into your pipeline.

If you throw “make it look good” at an AI and expect good output, the problem isn’t the tool. It’s the operator.

Everyone has their own We ship. We own. Find your way. Get there yourself.