Progressive Disclosure

Progressive disclosure is a museum placard: six words on the wall, the essay in the drawer underneath.

You show the agent the titles of twenty skills. It opens the one it needs. The rest stay shut, so the context window is not full of recipes for jobs it is not doing. Thirty full files would eat the room before the work started. Thirty honest labels cost almost nothing.

The label has to be true

A placard that undersells the drawer means the drawer never opens. A placard that bluffs — and buries the real step at the bottom — is worse. This plugin's scaffold-project description is a label on purpose: start a Next/Hono/stack, prefer MCP, don't hand-roll. The long workflow lives in SKILL.md, one step away.

Same trick in a UI. The builder shows the bag, not every CLI flag on the first screen. CLAUDE.md and AGENTS.md should point at detail, not recite the manual. The system prompt stays short; the cards stay on the shelf.

You have used the human version for years. Advanced settings behind a click. Onboarding that asks three questions, not thirty. Dump the textbook on the first screen and both humans and models drown.

What this unlocks

You can install a lot of know-how without filling the room. The agent finds the right drawer. The desk stays readable.

Show the short version. Keep the depth one honest step away.