An author writing a skill pastes every reference table, schema and worked example into SKILL.md so that nothing can be missed. A reviewer objects that this will cost context in conversations that barely touch the skill. Which account of how skills load supports the reviewer?

The principle — A skill is a package whose parts are read progressively. Skills load resources on demand, keeping context usage efficient, and that is what stops a thorough skill from taxing every conversation it brushes against.

Why the key is correct — The recommended shape is a focused SKILL.md with detailed material moved into a references/ directory, alongside optional scripts/ and assets/ directories. Those files are read when the work calls for them, so a long lookup table costs nothing until the Agent actually needs it. Pasting the same table into SKILL.md inverts the arrangement: the body is the part read whenever the skill comes into play, so the cost is paid every time rather than when it is earned.

Why the others are wrong — Nothing pulls a whole folder into the prompt at discovery time; discovery presents the Agent with what is available so it can decide relevance from each description, which is why the description field deserves real care. Loading everything on invocation would make the split between SKILL.md and references/ pointless, and the documentation would have no reason to recommend it. And the body of SKILL.md is certainly read once a skill is in play, because it is the procedure the Agent follows, so treating it as free is the mistake that produced the bloated file in the first place.

Remember this — Keep SKILL.md to the procedure and the decisions, and push tables, schemas and long examples into references/. Write the description carefully, since that short line is what decides whether the skill is chosen at all.

Sources — Cursor's Agent Skills documentation.