The first AI draft is always slop. Here’s the taste check that fixes it

The first AI draft will always fool you. It looks clean, it renders, it feels finished for about 30 seconds. Then you ship it and realize it looks like everything else.

I’ve written that AI has no taste — and I mean it as a practical problem. The model gives you the average of what it has seen. Your job as a solo founder is to move it from average to intentional before anyone pays for it.

Anthropic’s own move proves the point. Claude Cowork puts the agentic engine from Claude Code into a workspace where it edits files and runs tasks on your behalf. Capability keeps climbing. Judgment still sits with you.

Why does the first AI draft always look like slop?

Because the model is trying to be helpful, not specific.

A founder I watched recently put it well: she’d generated a full prototype in Figma Make in a thousand iterations, showed her husband, then saw it — “Everybody’s product looks like this.” Same gradient, same card, same safe hierarchy. It worked. It just didn’t feel like hers.

That’s the pattern. The first draft picks the most common spacing, button, and voice. Nielsen Norman Group’s research on crafting AI explanations makes the trust point: people don’t judge AI on whether it answered, they judge whether they can understand and steer what it did. A generic draft hides its defaults.

What are the four checks that catch it?

They’re boring. That’s why they work.

1. Intent. Does this screen do the one job it promised? A pricing page should make the choice obvious. An onboarding flow should get someone to a first win. If the draft tries to do three jobs, you haven’t edited — you’ve accepted.

2. Consistency. Does it use the system you already set? Same radius, same type scale, same motion. Each new prompt rebuilds the system from scratch if you don’t have a short reference file the agent must follow. Without that file, you have suggestions, not a system.

3. Hierarchy. Can a new user tell what matters in two seconds? Size, weight, and spacing should say it before copy does. Models give everything equal weight and the screen goes flat. Fix the order first, then the words.

4. Edge states. What does it look like when it fails or has nothing to show? Jellyfish’s August 2026 AI Engineering Trends report finds teams generate more code while PR throughput flattens — more output, same review cost. Edge states are where that cost shows. If empty, error, and loading weren’t designed, you didn’t finish the feature.

If a draft fails one of these, it’s not “almost there.” It’s not ready.

How do you run a taste check in under 15 minutes?

You don’t need a design ritual. You need a loop you’ll actually run.

Capture four screenshots: the happy path, one edge state, mobile, and the same screen next to a product you admire. Side by side, your eye catches weight, spacing, or tone drift immediately — even if you can’t name why.

Then ask the four questions in order. Pick the biggest miss and fix that one thing. Not four. One. Ship the fix, screenshot again.

The founder who caught that stray colored stripe on a hover state did it this way. Nine things shipped correctly overnight via agents. One hover had a leftover accent bar — textbook slop. She was the only one who saw it because she was the only one looking the next morning. Keep a tiny reference set the agent can read — five screens you love, with a line on why each works. Without it, every prompt starts from zero and drifts.

When should you ignore the check and ship anyway?

When you’re still learning if anyone wants the thing at all.

If five people are testing a private prototype, taste isn’t the bottleneck — distribution is. I wrote in Building got easy. Distribution didn’t that the scarce resource in 2026 is a user who comes back, and no amount of polish creates that. In that phase, the first draft is fine because you’ll throw it away.

Once you decide to keep a feature, run the check. Slop compounds. One generic screen feels harmless. Twelve become a product no one remembers. You don’t need better taste by a lot — just catch one default choice per feature and replace it with something you meant. Do that consistently and the product stops looking generated. It becomes edited.