If you can't afford to load it all, load an index and let the agent pull the rest.

Progressive disclosure

Starting the agent with a small entry point and showing it where to look next: a short index now, details when they are needed. A skill's description is the index entry, and its body is the detail.

You keep a whole shelf of procedures for the price of their labels. But the risk moves. The detail arrives only if the label matches, so a vague description means a procedure that never loads, and nothing tells you. Test the label, not just the procedure.

Say it to your agent

“Move the 400-line release procedure out of AGENTS.md into .claude/skills/release/SKILL.md, with the description "Use when cutting a release or bumping the version." Leave one line in AGENTS.md. Start a new session, ask it to cut a release, and show me that the skill loaded.”

Not to be confused with Just-in-time retrieval

Progressive disclosure is how you arrange knowledge, so there is a path to follow. Just-in-time is the agent following that path, or searching, at the moment it needs something.

Where you see it

OpenAI: "agents start with a small, stable entry point and are taught where to look next, rather than being overwhelmed up front." Anthropic: "Progressive disclosure is the core design principle that makes Agent Skills flexible and scalable." Anthropic reported the same effect for tools in November 2025: about 77K tokens used by tool definitions before any work, down to about 8.7K with tool search, a vendor's own test. Claude Code (November 2026) gives the skill listing 1% of the window and drops descriptions of your least-used skills when it overflows, and each description is cut at 1,536 characters.

Progressive disclosure

Before the agent

The agent

The harness

Everything it can get wrong

The operator

The workshop