AI is a lever, not a role

"Claude killed design." "Codex killed engineering." Every few weeks a new headline says a tool ended a profession. Scroll the replies and you'll find two camps: one popping champagne, one out for blood. Neither camp is right, and neither is really talking about the tool. They're talking about a role, as if "designer" or "developer" is a fixed job that either survives or doesn't.

That's the wrong frame. AI isn't a replacement for a role. It's a tool, the same category as Figma or a compiler. Tools don't kill roles. They change what the role is capable of, and they widen the gap between someone who's mastered the tool and someone who hasn't.

I've been designing for the web for four years. I'm getting good at it. And when I sit down with Claude Code, what comes out looks nothing like the "AI slop" flooding my feed.

That gap isn't about the model. It's about me.

What slop actually is

Think about someone who just learned Figma. They saw one example in class and they're trying to recreate it. They reach for the tools that are obvious from the surface: type, fonts, a background color. That's the full toolkit they have, so that's the full toolkit they use.

Now think about someone four years in. They know six different ways to build a gradient and when each one is wrong. They know how to layer image fills and color fills to fake depth that isn't really there. They know when auto-layout is fighting them and when to break out an absolute-positioned element instead. They've seen enough layouts that "unique" is actually possible, not just a word.

Same tool. Wildly different range of what gets pulled out of it.

AI works the same way. A generic prompt triggers the model to emulate a designer who knows maybe 25% of what the tool can do. It reaches for the same obvious moves the beginner reaches for, because that's what "generic" looks like statistically. A prompt loaded with specificity, taste, curated references, and real knowledge of how web frameworks actually behave produces something else entirely. Same model. Different designer being emulated.

The gap is the whole story

Here's the concept underneath all of this: there's an idea in your head, and you're trying to materialize it. The distance between that idea and the final result is set by how well you know the tool.

AI didn't remove that distance. It just handed everyone a much longer lever to close it with. But a lever only closes the gap you already know how to see. If you can't tell good from generic, the lever doesn't help you, it just gets you to generic faster.

This is why "AI will replace designers" and "AI is useless for design" are both missing it. The people producing slop aren't failing because AI is bad. They're failing because they don't yet know what they don't know, and the model has no way to fill in taste it was never given.

Sometimes the tool itself is the limit

There's a second version of the gap worth naming. Sometimes the distance between idea and execution isn't about skill at all, it's about the tool's ceiling.

Figma doesn't do real 3D transforms. It never has. For years, if your idea needed actual 3D, you either faked it with clever 2D tricks or you left Figma entirely.

But if you know WebGL and three.js, and you know how to direct Claude Code, that ceiling doesn't apply anymore. The idea that used to die at the edge of the tool now has a path through.

That's the real shift. It's not that AI makes good taste unnecessary. It's that AI makes the tool itself less of a bottleneck, so the only thing standing between your idea and the result is how well you can direct the thing translating it.

What using the lever well actually looks like

This shows up in my own workflow, not just the output.

For years the pipeline was linear: design in Figma, hand it off, watch a developer translate it into code, and lose a little fidelity at every step of that translation. The handoff was the bottleneck. Not the design, not the code, the gap between them.

Claude Code plus the Figma MCP collapses that gap. I can pull design context straight from Figma into code, or skip Figma for a first pass entirely and go straight to code and GitHub, iterating on the actual product instead of a picture of the product. That's not a minor speed-up. It's a different shape of work. The thing I'm iterating on is the real thing, not a representation of it that someone else still has to build.

None of this works if you don't already know what good looks like. The workflow only moves fast because I can tell, in seconds, whether what came back is right. Someone without that instinct has the same tools and a much slower, much worse version of this loop, because every iteration needs an outside opinion to know if it worked.

Where this leaves us

Wielding AI well isn't a separate skill from being good at your craft. It's an extension of it. The designer who knows every corner of Figma, the accountant who's mapped every Excel function, they're not going to be replaced by someone who knows neither and just has better prompts. They're going to be the ones getting the most out of the model, because they already know what they're pointing it at.

The tool got more powerful. What it amplifies didn't change.