A "skill" is a bundle of task-specific tools plus instructions (e.g., how to set an issue's priority) that an agent loads on demand through a meta-tool, rather than having every possible action and the full underlying schema available at all times. Linear's team found that exposing the model to everything at once — every action, the raw GraphQL schema — caused hallucination and overload; narrowing what's in scope at any given moment, and loading skills dynamically as the task requires, fixed it. The lesson generalizes: agent quality comes from curating what the model can see and do at each step, not from maximizing raw capability or access. Skills also double as a place to encode product opinions (e.g., a team's own priority-setting convention) rather than leaving that judgment to the model.
Related but distinct: Skills & Plugins Marketplace is about surfacing and distributing skills/plugins to end users; this is about the runtime mechanism an agent itself uses to keep its own context small. See also Context Engineering and Model Context Protocol (MCP).
Notion adopted the skills + progressive disclosure pattern directly from coding agents to fix "tool and instruction fatigue" in its own AI product: rather than dumping every capability and instruction into one bloated system prompt, capabilities are only loaded/disclosed to the model as needed. When an AI product's system prompt grows bloated with tools and instructions, restructuring it into discrete skills that are progressively disclosed is the fix.