I burned through my weekly Pro credits in an hour on day one. I was designing the onboarding flow for a side project I'm building in Claude Code. I went in curious. I came out with a lighter wallet and a sharper thesis.
The reviews were already splitting by then. On one side: Claude Design is remarkable. Screenshot to design. Inline edits. Interactive prototypes in minutes. The full creative loop, inside the same tool you already use for everything else.
On the other side: inconsistent buttons. Drifting states. Components that vary across screens in ways that would fail any design review.
Both are true. They have the same cause.
My team runs the same loop most design teams are running right now. You design something in Figma. You hand it off. Tradeoffs get made in development. Good ones, necessary ones. The shipped product drifts from the original file. Nobody goes back to update Figma. The next evolution starts from files that no longer reflect what's in production. More decisions. More drift. Repeat.
Figma solved a real problem when it solved it. One file, real-time collaboration, a single source of truth. That was genuinely valuable.
The issue is a decade of building on top of that model produced systems nobody would design from scratch today. 946 color variables. Components nested inside components. Layer names that read like error logs. The file became canonical. The product became secondary.
Claude Design makes the opposite bet. It's HTML and JavaScript all the way down. When it reads your system during onboarding, it reads your codebase. Not your Figma primitives. That's why the handoff to Claude Code is structurally cleaner than anything Figma Make can offer. There's no translation layer. The medium is already the medium.
The inconsistent buttons make sense once you understand this
Claude Design creates real structure. Components, assets, inspectable code.
The problem is it's inventing its own structure from scratch because it had nothing tight to pull from. Every button a new interpretation. Every state a new decision. Every screen a new guess at what your brand actually is.
That's not just inconsistency. That's compute. Every judgment call the tool makes costs you.
One account I read this week described stripping their Figma file down before connecting it to Claude Design. Kept only tokens, core atomic components, and key accessibility and error patterns. Flattened the variants. Cut the file size by roughly ten times. The hypothesis: a focused file produces more intentional output and stops you from burning your weekly limit on artifacts the model was never meant to consume.
Design systems are built for humans using Figma. Annotations, templates, slideware, work in progress, organizational scaffolding. Useful to a designer navigating a library. Noise to a model. You're paying for every token of it.
That's not a Claude Design tip. That's a design systems discipline problem that just became expensive enough to care about.
The obvious objection: Claude Design will get smarter. Patterns will improve. Componentization will tighten. That's probably true and it doesn't change the argument. Anthropic just ended the flat-fee era for enterprise. Every token now has a price attached to it, billed at standard API rates on top of your seat. The tool improving and the costs decreasing are not the same curve. A smarter model running against a bloated, drifted system still burns money on noise. The economic pressure toward tight systems is structural. It's moving in one direction.
This is judgment work
Making a system tight enough for Claude Design to work isn't AI work. It requires a designer who can audit what exists, reconcile the drift between Figma files and production code, define component states, and establish token structure. Someone who knows the difference between what your library claims your system is and what's actually shipping.
Claude Design has no native concept of patterns yet. Error states, messaging, notifications. Your system is full of them. The tool currently has nowhere to put them. The first release is genuinely impressive and genuinely incomplete.
The tool cannot do this for you.
The calculus isn't "Claude Design replaces the designer." It's that the designer who prepared the system is the reason everyone else can move fast. Product. Development. The AI itself. They're not cleaning up after the tool. They're the reason the tool works. The clean handoff to Claude Code that everyone is excited about is only clean if the system behind it is.
That's the multiplier. And that's a different version of dangerous than we were talking about a few weeks ago. But it's still the same word.
Figma won the last decade by becoming the source of truth. Claude Design is making a credible argument that the source of truth was always the product, and that the decade of infrastructure built to maintain the fiction otherwise is exactly the debt you're paying for now. At AI speed.
The machine doesn't know your brand. That's not its failure. It's our unfinished work.
This is the seventh piece in a series on design leadership, internal tools, and what it actually means to become dangerous in an AI-accelerated profession. Start with Why Internal Tools Are the Most Underrated Design Challenge.