The design system AI agents actually follow.

Great Learning runs a learner app, a suite of internal staff tools, and a public marketing site. In practice, all three used Magna, the learner system, and the other two improvised around it. They don't share a runtime, a component vocabulary, or an audience, so each one drifted on its own.

Then AI coding agents arrived: Claude Code, Cursor, and PMs and designers generating UI from prompts. Without an authoritative source, an agent fills the gaps with values it makes up.

So the brief was never "make it prettier." It was: three surfaces, three first-class systems, one source of truth that humans and agents build from in the same way.

The founding decision: define each system by who it serves, not by a shared component list. No abstraction layer forcing them into a fake common API. Each system's audience decided its rules, down to the breakpoint.

The deepest work started with blunt feedback: the component definitions "weren't working." There were two separate failures, found months apart, and each taught a different lesson. One component, the Switch, shows the whole path.

Before the pivot, the right values were two hops away and nothing pointed to them. After it, they sit inside the one file every agent reads first, written as a rule rather than a suggestion.

The skill is what gets the system to an agent. Every system ships one in two parts, design rules and content rules, but they aren't copies of one template. Each system's purpose decides what comes first, and what is forbidden is spelled out as carefully as what is allowed.

The old docs described responsiveness and accessibility in prose, and nothing a machine could follow. Both became per-component fields. Contrast is computed rather than assumed, and AA failures are flagged inside the definition, which gave designers a concrete list to fix.

The same idea applies to content. A first attempt copied Great Learning's whole writing-review workflow into the docs, including intake questions and output formats. The fix was to keep only the rules, nested under the right section. Where a rule sits mattered as much as what it says.

Each system now has a self-contained design-system doc covering colour and type, responsiveness, content rules, the AI protocol, the embedded component reference, and anti-patterns. Every definition holds exact dev values and plain-language guidance (what it is, when to use it, when not to) in the same file, so an engineer's spec and a PM's explanation can't fall out of sync.