The New AI Design Split: Tools That Inherit Your System vs Guess It
By The WAALEE Team · · 5 min read
A quiet split has opened up in the AI design tooling market this year. On one side you have tools that reconstruct your design system by guessing from a prompt, a screenshot, or whatever training data happens to look close. On the other, a newer crop of tools reads your design system's actual source, the real components and tokens you already ship, and builds from that instead of approximating it. The pitch for the second camp is obvious: stop drifting. The catch nobody puts in the headline is that reading your real source only works if a real source exists to read.
Two very different claims about 'knowing' your design system
Most AI coding and prototyping tools, v0, Lovable, Bolt, and friends, work by pattern-matching: you describe what you want, and the model fills in the gaps with whatever UI pattern it's seen most often. Builder.io's own writeup on design-system automation makes the point plainly, most vibe coding tools generate from scratch using generic patterns with no real knowledge of your component library or tokens. That's why two different companies prompting for "a clean SaaS pricing page" tend to get something close to the same page back.
The newer camp, tools like Alloy, argues the fix isn't a better prompt, it's a better source. Alloy's positioning is that it reads your design system's source code directly rather than reconstructing it from screenshots or descriptions, pulling your real components, tokens, and icons so every output is built from the system you already ship, not a plausible imitation of it.
Inheritance works, but only if the wiring is already there
This is the part that matters if you're not already a mature product team. "Read the source" tools assume the source is readable in the first place. Builder.io's own design tokens docs show what that actually requires in practice: tokens registered in code through an SDK call, ideally backed by CSS variables, so the tool can limit edits to only the values your system actually allows. That's real engineering work, done ahead of time, by a team that already has a documented token layer.
- If your design system lives in a token registry or a component library with real source code, inheritance-style tools can plug straight in and stay on-brand.
- If your "design system" is really just a Figma file, a style guide PDF, or whatever your last redesign happened to produce, there's no machine-readable source for these tools to read yet.
- Most small and mid-size teams are in the second bucket, which is exactly why reconstruction-based tools still dominate the market despite the drift problem being well known.
The drift problem is real, and well documented
Design-system commentary this year keeps landing on the same complaint: left unsupervised, AI-generated UI drifts from whatever system it's supposed to match. Sparkbox's piece on finding and fixing design system drift walks through the usual detection methods, visual regression testing to catch pixel-level differences, static analysis to catch hardcoded values that should be referencing tokens instead. Those are good tools for catching drift after the fact. They don't stop a team with no token source from drifting in the first place.
Where WAALEE fits: building the source these tools assume you have
This is the actual gap. Inheritance tools need a token source to inherit from. If your brand's real colors, type scale, and spacing only exist as the rendered output of your live site, nobody's written them down as a registry a coding agent can read. That's the step WAALEE does: it extracts your site's real design system straight off the live, rendered page and turns it into proper W3C design tokens, the same kind of structured source that inheritance-style tools are built to consume, and that plain prompt-based tools can be fed directly to stop guessing.
Three different answers to "does this tool know your design system, or is it guessing":
- v0 / Lovable / Bolt (prompt-based): Fast and flexible, but defaults to generic patterns unless you feed it your actual tokens and components in the prompt or project context.
- Alloy (source-reading prototyping): Reads your design system's existing source code directly for drift-free prototypes. Requires that source to already exist as registered components and tokens.
- WAALEE: Extracts real W3C design tokens straight off your live site, no pre-wired token registry required, then turns them into ready prompts for Claude, Lovable, v0, Cursor, and Bolt.
The split between "inherit" and "guess" is a genuine improvement in how the industry thinks about AI and design systems. But it quietly raises the bar: now the question isn't just which tool you use, it's whether you have a real token source for any of these tools to work from at all. If you don't, that's the gap to close first, before you pick a tool to build on top of it.