Skills on disk.
Descriptions on demand.

Pi and Claude Code both put a catalog of installed skills into the model’s context. Each entry is a name plus a description, and the catalog is rebuilt on every request. Nothing about it changes between turns; it is simply there, every time, ahead of the work.

Deferred Skill Catalog keeps the skills on disk where they already are. What changes is where the descriptions live: out of the prompt, behind two tools the model calls when it needs them.

search_skills

Answers “which skill fills a PDF form?” with at most 8 matches, each naming the skill, its trimmed description, and how to load it.

list_skills

Pages through the full inventory, 20 per page by default and 30 at most.

Human listing

/skills [query] [page=N] in Pi, and /deferred-skill-catalog:skills [query] [page=N] in Claude Code.

It works with any skill directory: skills written for one harness and used from another, plain repositories that aren't packages, and any directory listed in roots.

The two harnesses defer differently. Pi replaces its <available_skills> block with a short pointer to the tools, so its deferral is automatic. Claude Code has no hook that can rewrite its system prompt, so descriptions are dropped by its own skillListingBudgetFraction setting and served back by an MCP server bundled with the plugin. In Claude Code the skill names stay in the prompt; only descriptions are removed.

Version 0.2.1, MIT licensed, installable as a Pi package and as a Claude Code plugin from its Git repository. The two harnesses are not symmetric, and this site keeps that distinction: Pi's deferral is part of the package, while Claude Code's depends on a setting you turn on in its own configuration.

This site publishes no benchmark, percentage, or savings figure. The source README records one measurement taken on one machine; that measurement has not been reproduced here, and a single machine is not a promise about yours.