An AI-native design system as a Claude skill

I prototype constantly with AI coding agents, mostly Claude Code and Cursor, against the Horizon design system. The system itself is well documented for human consumption — a website, component pages, token references, the works. But every prototype I generated came out close to right and not quite right. Wrong tokens. Wrong spacing scale. Component compositions that looked plausible but were not how Horizon is actually used.

I was paying a correction tax on every prototype: a hand-cleanup pass to make AI-generated UI look like ServiceNow. After the fifth or sixth time of doing this, it stopped feeling like a one-off and started feeling like a missing tool.

The reason the design system was not working for AI was that it had been built for an entirely different audience. A human designer learning Horizon flips through pages, scans examples, builds intuition over weeks. An AI agent does not flip through pages. It reads what you put in front of it and follows the rules it can extract from that text.

Most design system documentation, written for humans, is the wrong shape for that. Long. Discursive. Full of context the agent does not need. Light on the specific, declarative rules the agent does need. The system was not failing the agent — the agent was reading the system the only way it knows how, and finding noise where it needed signal.

I built a skill — a compact instruction file — by scraping my production Figma files, extracting AINPX color tokens, component specs, and AI-native interaction patterns, and packaging the result as a layer that sits on top of the standard Horizon DS skill.

The skill is not a stylesheet or a documentation page. It is a set of rules an AI follows when generating any AINPX surface. Use these tokens. Compose these components. Refuse these patterns. Invoke these AI-native primitives. Short, declarative, opinionated.

The strongest thing I learned packaging this work is counterintuitive: the more comprehensive I made the skill, the worse the agent performed. Comprehensiveness becomes noise. Beyond a certain point, every additional rule you add dilutes the signal of the rules that matter most.

The skill format forces a useful discipline: if a rule cannot be stated in two sentences, it is probably not a rule — it is a preference, or a context-dependent judgement that belongs in the designer's head, not in the agent's instruction set. That has changed how I think about design system content generally, even outside AI tooling.

Prototypes I generate now come out in the AINPX visual language without correction passes. The skill is in active use by me and a handful of designers on adjacent teams who hit the same correction tax I did. It is not an org-wide rollout — it is a personal tool that turned out to be useful to people working on similar surfaces, which I think is the honest size of the thing.

More importantly, the skill format has become how I package any reusable design context for AI agents. The portfolio you are reading right now was built using the same approach.