Knowledge

Tokens tell an agent what things look like. They say nothing about whether a button belongs on the screen, how many primaries a surface may carry, or what a table does when it runs out of width. That judgment is the part teams keep in people’s heads, and it is the part an agent cannot infer.

The knowledge repo is where Border UX wrote it down. It is plain markdown, MIT licensed, and it is the knowledge base the agents actually read.

Skills are the whole knowledge base

Every skill is a directory under skills/ holding a single SKILL.md, following the Anthropic Agent Skills specification. Nothing else in the repository is agent knowledge. The repo says so explicitly, and it matters: there are also human-facing component pages under docs/, parts of which are out of date, and where the two disagree the skill wins.

That convention is the entire index. A directory with a SKILL.md is a skill, so nothing has to be registered anywhere, and the list below is read from the repo rather than maintained by hand.

Load the router first

For any screen-level work an agent loads recursica-skill-design-router before anything else. It sets the order the decisions get made in, routes each one to the skill that owns it, and states what happens when two rules disagree.

Both layers, for every component placed. The component skill says what the component is and how to make it accessible. The design-rules skill says whether it belongs on the screen at all. Working from a component skill alone is the most common way to build something that is individually correct and collectively wrong.

When two skills conflict, the router states the precedence. In short: the design system owns styling, a prohibition beats a permission, and a design-rules skill beats a component skill on composition. The component skill wins only on which variants and states actually exist.

When nothing applies, the agent asks. The rule is to ask rather than guess, and never to import a convention from another design system to fill the gap.

What a skill looks like

Frontmatter first: a name, a license, and a description that is doing real work. The description is how an agent decides whether to load the skill at all, so it lists the triggers in the words a person would actually use, and just as importantly the anti-triggers that route elsewhere. The button skill’s description ends by sending navigation to the link skill, single-choice to the segmented control, and cross-screen hierarchy to the buttons-and-links rules.

Then the body: when to use the thing, when not to, and a table of what to use instead. The rules are specific enough to be checkable. A button that changes the URL is a link, no matter how it should look.

Editing it

It is your repo once you clone it. The rules are Border UX’s opinions, and a rule you disagree with is a file you edit. Adding your own research is adding a directory with a SKILL.md in it, and the contributing guides cover the conventions. Agents pick up the change on the next run, with no build step between writing a rule down and having it enforced.

Agents that consume all of this live in agents/. See the team for what each one does.

The four categories