Author Skills for progressive disclosure: SKILL.md under 500 lines, push detail to reference files

Claude Docs·medium signal

Only Skill metadata (name + description) is preloaded at startup; Claude reads SKILL.md when the skill becomes relevant and reads linked reference files only when a task actually needs them, so unused detail costs zero context tokens. Keep SKILL.md under 500 lines of high-level logic and split anything longer into separate files referenced inline ('For complex form-field mapping, see forms.md'). The highest-leverage authoring move is a distinct, intent-rich description — that string is what triggers the skill, and a vague one means it never loads.