OpenCode released Skills documentation this week: AI-compliant rules are written in Markdown and loaded on demand, so unused rules don't consume context space. This isn't a minor optimization—it separates "telling the AI how to do things" from "giving the AI tools" and "assigning the AI assistants."
What this is
A Skill is essentially a manual for the AI to read (a SKILL.md file), specifying coding standards, framework usage, business terminology, etc., which the AI only calls and reads when needed. OpenCode also clarified three mechanisms: Skill is the "manual," telling the AI what to do; MCP (the protocol for AI to call external tools) is the "toolbox," enabling the AI to do something; Subagent (an independent sub-AI process) is the "assistant," executing sub-tasks for the AI. Three distinct things, three distinct solutions—they shouldn't be conflated.
Industry view
We note this direction is becoming an industry consensus: no matter how large the context window, it cannot hold all knowledge; on-demand loading is the inevitable path. Cursor and Windsurf are doing similar things. But every layer of abstraction adds maintenance costs—teams must maintain not only code standards but also Skill files that AI can read, and the two won't always stay in sync. Some developers also question: when the number of Skills balloons, how does the AI judge which one to load? The cost of loading the wrong Skill might be greater than not loading one at all.
Impact on regular people
For enterprise IT: There is a new path for internal knowledge management—writing team standards into Skills, which AI coding assistants automatically follow, instead of relying on word of mouth. For individual careers: The essence of writing Skills is "training AI," which may become a new skill for technical management roles. For the consumer market: No direct impact in the short term, but the concept of "on-demand knowledge loading" will gradually spread from coding scenarios to broader AI products.