A `SKILL.md` body stays under 500 lines, with depth in `references/`
The whole body loads into context when the skill activates, competing there with the conversation and with every other active skill. The specification recommends fewer than 500 lines and 5000 tokens, with supporting files loaded only when the task calls for them.
Observable check
every skills/*/SKILL.md is under 500 lines, and any skill needing more material ships it under that skill's own references/ directory.