How do skills stay cheap in context?

The reason skills scale is progressive disclosure. The agent does not load every skill in full. It works in three stages: at startup, only the short name and description of each skill load, just enough for the agent to know it exists and roughly when it applies. When a task actually matches that description, the full SKILL.md instructions load into context. Only then, if the instructions call for it, do the heavier bundled resources — a reference doc, a template, a script — load too.^[1]

The published numbers make the cost concrete: metadata runs about 100 tokens per skill and is the only part paid on every single turn, the full SKILL.md body costs under 5,000 tokens and only once a task actually matches, and bundled resources cost nothing at all until something inside the instructions decides to read them.^[1] A hundred-skill library therefore pays roughly 10,000 tokens at startup for descriptions alone, not the hundreds of thousands it would take to keep every skill’s full instructions resident all the time. That first stage is capped by the spec itself, not just by convention: a skill’s name tops out at 64 characters and its description at 1024, a hard ceiling on how much any single skill can cost before it has even been triggered once.^[1]

The result is that a large library of skills costs almost nothing in context until one becomes relevant. This is what makes skills different from stuffing every procedure into a single giant system prompt: a system prompt pays its full token cost on every turn whether or not any of it is used, while a library of a hundred skills pays only for the handful of short descriptions plus whichever one skill’s full instructions actually fired. See skill, tool, or prompt for that comparison directly, and inside a skill for what sits behind each stage.