Intent is distributed. That's the problem.

A design system stores knowledge across at least three places. Documentation explains when to use a component, when not to, and what it's trying to do. The codebase holds the behavior: props, states, constraints as implemented. Figma holds the structure: variants, naming, and visual relationships.

No single source is complete. Docs don't know what the code does. Code doesn't know what the docs mean. Figma knows what something looks like, but not what it's for. Point AI at any one of them, and it gets a partial picture. It fills in the rest with assumptions.

That's not an AI problem. It's an architecture problem.

The framework defined how to extract intent across all three.

The methodology worked in layers. First, identify all sources of truth for a given component — documentation, codebase, Figma — and treat each as holding a different kind of signal. Second, normalize across them: standardize component names, variant properties, states, and relationships so the inputs could be read together rather than separately. Third, extract meaning rather than structure — not what a component is, but what it's for, when to reach for it, when to reach for something else, what constraints govern it, what accessibility requirements apply.

The output was a component contract: a YAML file that packaged extracted intent in a form AI could actually use. Not structural data alone, but structured meaning.

Defining the heuristic was the work.

Engineering built the Figma plugin that generated the YAML output. The technical implementation belonged to them. What required a different kind of expertise was the question the plugin had to answer: what does intent need to contain to be decision-useful to AI?

That heuristic didn't come from the codebase or from Figma. It came from years of structuring documentation precisely enough that designers and engineers could work from it without having to fill in the gaps themselves. The context extraction framework was where that accumulated understanding became a methodology.

Component contract output was evaluated against documentation standards to ensure intent was captured completely. The schema was defined, the heuristic was established, and the evaluation criteria were clear enough to know when the output was right. The plugin was still in progress when the layoff came. The loop never closed in production.