Crabhaus

What it is

A skill is a folder of procedural know-how an agent loads on demand. Its SKILL.md opens with YAML frontmatter — a name and a one-line description — followed by markdown instructions, with scripts, templates, and reference docs alongside. Instead of cramming every procedure into a system prompt, expertise ships as files the agent reads only when a task calls for them.

How it works

Progressive disclosure, in three stages. At startup the agent preloads only each skill's name and description — tens of tokens apiece. When a task matches a description, it reads the full SKILL.md into context. Bundled references are read, and scripts run, only if that run actually needs them, so a skill can hold megabytes while idling at near-zero context cost.

A concrete example

Anthropic's official skills cover procedures like building .xlsx spreadsheets or filling PDF forms: the frontmatter costs a sentence, and the full checklist plus helper scripts load only when a matching task appears. DeepSeek's Harness, previewed this week, treats skills as one of its swappable plugin types alongside models and tools.

Why it matters

Context is scarce, and procedural knowledge is most of what teams want agents to retain. Skills turn it into portable, versionable files instead of prompt sprawl. The new vendor-neutral Agent Plugins 1.0.0 spec standardizes the packaging — one directory with plugin.json, skills/, and mcp.json — with maintainers from Amazon, Cursor, Microsoft, OpenAI, and Vercel.