Skills: expertise the agent loads when it needs it

A skill is a packaged procedure: a markdown file (SKILL.md) with YAML frontmatter (name, description, and optionally which tools and model it should use) followed by a body of instructions. Think of it as a runbook the agent can pull off the shelf. The "Seven Pillars" breakdown calls this "skills as workflow encoding," pillar 3.

The crucial mechanism is progressive disclosure. At session start, the agent sees only the skill's name and description (a line or two, costing maybe 100 tokens). The full body loads into context only when the agent actually decides to use that skill. So you can have fifty skills available and pay almost nothing for the forty-nine you do not use this session. The "Seven Pillars" breakdown gives the token math: skill names and descriptions might cost about 500 tokens total at startup, while a full skill body (often 5,000-plus tokens) is attached only on invocation and dropped when the agent moves on.

A skill body can do more than give instructions. It can orchestrate a multi-step deterministic workflow, bundle scripts and templates alongside it, and degrade gracefully (if a script fails, the agent can fall back to the prose instructions). This is how a single "investigate" skill can encode "run these three analysis scripts, compare results, write a report," turning a fuzzy request into a reliable routine.