Authoring Agent Skills

The Goal: Efficient Context Usage

Every skill you create will be loaded into Claude's limited context window. The fundamental challenge is:

Every unnecessary token wastes limited context space, slows processing, costs money, and reduces room for actual work.

Everything else—structure, naming, content—follows from this goal. Understanding why these patterns matter allows you to make good decisions in novel situations, rather than mechanically following rules.

Three Core Principles

1. Progressive Disclosure: Load Only What's Needed

The Intent: Don't force Claude to read 5000 lines when it needs 100.

Why It Matters: When a skill is triggered, SKILL.md loads entirely into context. If you put everything in one file, every query wastes tokens on irrelevant content. A finance query shouldn't load sales data. A basic usage shouldn't load advanced edge cases.

How to Apply: