AI Kept Inventing a Brand, So One Team Rebuilt Its Design System as Tokens

A design team whose AI assistant repeatedly invented brand elements that do not exist has published a post on Medium explaining how it responded: it rebuilt its entire design system as tokens, replacing prose guidelines with machine-readable values that a model cannot misread.

The title of the post states the problem plainly: "Claude Kept Hallucinating Our Brand, So I Rebuilt Our Design System as Tokens." The author's diagnosis is the concrete part. When brand rules live in written documentation, a language model reads them as suggestions. It samples plausible colors, spacing and type scales from its training data, not from the guideline, and produces output that looks on-brand to a machine and wrong to a designer.

That failure mode is structural, not a prompt-engineering gap.

Why do prose guidelines fail with generative tools?

A traditional design system carries its rules in two layers: components in code, and usage rules in paragraphs. Humans read the second layer. Generative assistants largely skip it or reinterpret it, because natural language is ambiguous to a model in exactly the ways it is convenient to a person.

Tokens change that. A token is a named, versioned value — a color, a radius, a spacing step — that tooling resolves mechanically. When a palette is expressed as tokens rather than described in sentences, the model does not interpret; it resolves. There is nothing to hallucinate, because the value is bound to a name the toolchain enforces.

The Medium author's move follows this logic end to end: rather than asking the AI to behave, they removed the interpretive step the AI was failing at.

What does this change for design teams?

The practical consequence is a shift in where a design system's authority lives.

• Rules move from documents to data: the token file becomes the single source of truth, and documentation describes it rather than defines it.
• Review shifts downstream: instead of checking AI output against a PDF of guidelines, teams check whether the token pipeline was respected.
• The system gains a machine interface without losing its human one — designers keep names and semantics, tools get values.

This is the same architecture principle that carried design systems from static style guides to themed component libraries: encode the constraint where it cannot be bypassed.

Is this a general pattern or one team's workaround?

The post is a first-person account, not a study, and it concerns one assistant, Anthropic's Claude, interacting with one organization's brand. But the underlying mechanism is general. Any generative tool that produces interface code will fill gaps with statistical guesses unless the gaps are closed by explicit, resolvable values.

For practitioners, the takeaway is narrower and more useful than "adopt tokens." It is this: every rule you leave in prose is a rule a generative tool will eventually break, and every rule you encode as a token is a rule it cannot. Teams that already run token pipelines — for theming, for multi-platform delivery — now have a second, defensive reason to keep them strict.

As AI assistants take on more of the routine production of interface code, the systems that survive contact with them will be the ones whose constraints are executable rather than descriptive.