Agent skills are the new plugins

An agent skill is a packaged set of instructions and expertise that gives an agent a capability it did not have. The format is described as simple and open. This set covers it, along with directories for finding skills, bundles of them, and apps that run several agents at once.

The word "plugin" fits loosely, so the sections below follow what each item does: the format, places to find skills, packaged configurations, and tools for running agents.

[Corey: why do you see skills as the plugin layer for agents, and what made you start paying attention to them?]

The format itself

Agent Skills is the home of the format. It describes itself as "a simple, open format for giving agents new capabilities and expertise." The directory tags it for both Claude and ChatGPT, so it is the place to read what a skill actually is before you install anything.

Start there if you want to know what a skill contains before you install one.

Where to find skills

Two directories in this set exist to help you search and install skills.

The Agent Skills Directory at skills.sh is built to "discover and install skills for AI agents," and the link goes to its trending page, so you can see what other people are reaching for. It is tagged for Claude.

Claude Code Skills Hub lists 308 skills as of this writing, with fuzzy search and filters. It also has interactive tutorials, and it says everything in it is open source. If you want to learn how skills are put together while you browse, this one is built for that.

UI Skills is a single collection rather than a directory. It is a set of skills for polishing interfaces that agents have built, so it is a concrete example of what a skill for one job looks like.

Whole configurations in one install

Everything Claude Code Plugin bundles agents, skills, hooks, commands, rules and MCPs into one Claude Code configuration collection. Its description credits the configs to an Anthropic hackathon winner, and the link goes to a GitHub repo (affaan-m/everything-claude-code), so expect to read the README to see what you are taking on.

Ralph Wiggum is a different kind of bundle. It is an agentic coding loop that is open source, spec-driven and community-led, aimed at real-world teams. Its listing includes a line telling an AI agent to read an INSTALLATION.md file, so the install is written for the agent to follow.

In both cases you are installing a lot at once. The underlying thing to check is what each piece will change about how your agent behaves, so go through the repo before you adopt the whole set.

Running agents around your skills

Skills tell one agent how to work. The two tools below deal with the agents themselves.

Conductor creates parallel Codex and Claude Code agents in isolated workspaces. You see at a glance what each one is working on, then review and merge their changes.

Letta is a platform for stateful agents. You build agents with long-term memory, advanced reasoning and custom tools through the Letta API and its Agent Development Environment (ADE). Memory is its own idea, separate from skills, so read it as the neighboring approach to giving an agent context.

FAQ

What is an agent skill?

It is a packaged set of instructions and expertise that gives an AI agent a new capability. Agent Skills describes the format as simple and open.

Where can I find agent skills to install?

The Agent Skills Directory is made for discovering and installing them. Claude Code Skills Hub lets you search 308 skills with filters.

Can I get skills, hooks and commands in one install?

Everything Claude Code Plugin collects agents, skills, hooks, commands, rules and MCPs for Claude Code in a single GitHub repo.