[PROXY-VALIDATION] The Coherence Gap: Why AI Needs Text Files More Than Specifications

We're approaching design systems backwards.

Google Labs just released DESIGN.md—a format specification for describing visual identities to coding agents. The impulse is understandable: give AI a structured, machine-readable blueprint and it will faithfully execute. But there's a quieter insight hiding in the recent wave of "text files as UI" discussions: sometimes the most coherent way to communicate with machines is to not optimize for machines at all.

The paradox is real. Specifications promise clarity through formalization. DESIGN.md does this for design tokens, constraints, and hierarchies. But formal specifications have a hidden cost: they force us to choose what matters. The moment you commit "button blue = #0066FF," you've eliminated every other dimension of judgment that led to that choice. Why that blue? Because it contrasts well on light backgrounds, because it matches the brand's trust positioning, because users in our market associate it with action, because the previous designer chose it and changing it would break habituation. The specification captures zero of this.

This is the coherence gap: the distance between what we specify and what we actually understand.

AI agents amplify this gap. They optimize for the specification, not the intent behind it. Give them a rule, they'll find the edge case that technically satisfies it while violating every principle that motivated the rule. This isn't new—AI safety researchers call it specification gaming. But it's newly urgent because we're about to hand design systems to agents that operate at scale and speed we can't monitor in real-time.

The alternative isn't to abandon structure. It's to recognize that text files—prose, narrative, context-rich documentation—actually preserve coherence better than formal specs do. A designer reading a well-written design system document absorbs not just rules but reasoning. An AI reading the same document can be trained to extract intent-level signals, not just token-level rules.

YAGNI—"You Aren't Gonna Need It"—was never about avoiding preparation. It was about avoiding premature specificity. We're committing to machine-readable specifications before we've actually understood what needs to be preserved in translation from human judgment to machine execution.

The path forward: design systems that layer narrative over specification. Keep the structured tokens. Add the story. Let agents learn coherence through context, not just compliance through rules. The cost of doing this isn't higher—it's distributed differently. We spend less time debugging edge cases where specifications broke and more time upfront articulating why decisions matter.

Text files won. They won because they're flexible, human-readable, and preserve context. Now we need to learn how to make them machine-coherent without losing that edge.