Zed just shipped v1.4.2 with a feature that quietly changes how you collaborate with your AI agent: Skills. Think of them as reusable instruction packages, each one a folder with a SKILL.md file, that teach the agent a specialized workflow. Test-driven development, PDF processing, your team's internal code standards, a deploy procedure you never want triggered by accident. You define it once, and the agent knows when to reach for it.

Out with Rules, In with Skills

The old Rules Library has been removed and replaced with skills support. Upgrading requires swapping the old @rule syntax to @skill in agent chats, but all other configuration files carry over without forcing a manual rewrite. Existing rules are migrated automatically, so there's no cliff edge here.

A global AGENTS.md file now sits alongside settings.json and feeds instructions into every project automatically, preventing context drift when jumping between unrelated repositories. That's a separate but complementary addition: Skills handle task-specific knowledge, while AGENTS.md handles the always-on baseline.

The Architecture Behind It

The key design decision here is progressive disclosure, a loading strategy borrowed from the Agent Skills open specification. Rather than dumping all your instructions into the system prompt at once, the agent only loads what it needs, when it needs it.

The system breaks knowledge loading into three levels: L1 metadata (~100 tokens per skill), which is just the skill name and description loaded at startup so the agent knows what's available; L2 instructions (under 5,000 tokens), the full skill body loaded only when the agent explicitly activates a skill; and L3 resources, external reference files like style guides or API specs, loaded only when the skill's instructions require them.