2026 — Skills, plugins, and the agentskills.io standard

Why this matters

Skills are the reusable instruction layer of the three-layer stack — the answer to “I keep telling Claude the same thing across projects; how do I package that once?” A skill is a folder with a SKILL.md (the instructions for the model) and optionally some supporting scripts and resources. The model loads the skill into context when relevant — not always, not never.

What made skills the headline 2026 ecosystem story:

  1. The plugin marketplace shipped in Claude Code itself — /plugin is a built-in command, with the Anthropic-managed marketplace pre-connected.
  2. The agentskills.io open standard formalised the SKILL.md format with cross-vendor portability — Claude Code, OpenAI Codex CLI, Cursor, Gemini CLI, GitHub Copilot all read the same format.
  3. The supply exploded: from a handful of official examples at the start of 2026 to over a million community contributions indexed across marketplaces by mid-year (per the tonsofskills.com aggregator, numbers as of 2026-05: 425 plugins, 2,810 skills, 200 agents).

The combined effect is that a 2026 Claude Code workflow is built much more around composed skills than a 2025 one was around hand-written CLAUDE.md content.

Skill anatomy

A skill is just a directory with a SKILL.md in it:

my-skill/
├── SKILL.md           # required — the instructions Claude reads
├── scripts/           # optional — helper scripts the skill can run
│   └── do-thing.sh
└── reference/         # optional — extra docs the skill can reference
    └── api.md

A minimum-viable SKILL.md:

---
name: my-skill
description: One sentence on when this skill should be activated.
---

# My Skill

When the user asks for X, do Y, then Z. Use the script at
`scripts/do-thing.sh` when running step 2.

If the input matches `<some-condition>`, instead do `<other-thing>`.

Two things matter about the format:

You can disable a skill without uninstalling it by renaming SKILL.md to _SKILL.md (the harness’s discovery walks for SKILL.md and won’t pick up the underscore-prefixed variant — see the Mark Chen walkthrough for the same trick documented in detail).

A subtle but consequential clarification on skill identity: the harness resolves skills by folder name; the name: frontmatter is display metadata. Folder my-skill/ with frontmatter name: my-skill is the conventional shape, but if the two diverge the folder name wins for resolution.

Plugins: how skills get distributed

A plugin is a packaging convention — one repo with one or more skills, plus optional hooks, MCP server configs, slash commands, and subagent definitions. A plugin’s job is to bundle these into a single installable unit.

my-plugin/
├── .claude-plugin/
│   └── plugin.json    # required manifest
├── .mcp.json          # optional MCP server config
├── commands/          # optional slash commands
├── agents/            # optional subagent definitions
├── skills/            # optional — usually the bulk of the plugin
│   └── my-skill/
│       └── SKILL.md
└── hooks/             # optional hook scripts

The required .claude-plugin/plugin.json is small:

{
  "name": "my-plugin",
  "version": "0.1.0",
  "description": "What this plugin does, one line.",
  "skills": "./skills"
}

A marketplace is just a repo containing a .claude-plugin/marketplace.json that lists multiple plugins. The Anthropic-managed marketplace at anthropics/claude-plugins-official follows this shape; community marketplaces (netresearch’s, the tonsofskills.com one, Whittlery’s eventual Genno-curated one) all do the same thing.

Installing — the 2026 flow

From inside any Claude Code session:

/plugin

Opens the plugin browser. Tabs: Discover (browse marketplaces), Browse (browse installed plugins), Manage (uninstall / update).

A 2026 quality-of-life addition: the Discover and Browse screens show what a plugin contributes before you install it — its commands, agents, skills, hooks, and any MCP / LSP server config. And: per-session projected token cost is shown for plugin components. So you can audit “this plugin will cost ~X tokens per session of overhead” before committing.

The CLI equivalent:

claude plugin marketplace add <github-owner>/<repo>
claude plugin install <plugin-name>@<marketplace-name>
claude plugin details <plugin-name>      # the per-session cost view

A subtle 2026 change: the keyboard shortcut for “remove” in the plugin manager changed from r to d — r was collidng with “retry” and causing accidental removals. Worth knowing if your muscle memory says r.

Cloning is SSH by default when the source is github, which trips up users without a global github.com SSH key (common — many people use per-org aliases like github-genno). To force HTTPS, use the environment variable CLAUDE_CODE_PLUGIN_PREFER_HTTPS (shipped in v2.1.141):

export CLAUDE_CODE_PLUGIN_PREFER_HTTPS=1

Or set up a global gitconfig rewrite:

git config --global url."https://github.com/".insteadOf "git@github.com:"

(See the agensi.io marketplace guide for the full plugin install lifecycle.)

The agentskills.io open standard

The reason skills matter as an ecosystem play, not just a Claude Code feature: the agentskills.io open standard formalised SKILL.md as a cross-vendor format. The same skill folder runs in:

The wrapper manifests vary (each tool wants its own plugin.json shape pointing at the skill), but the skill itself is portable. For someone writing a skill that’s worth multiple tools’ attention, this is a real win — you write it once, you bundle it for the tools you care about, you don’t fork the body.

The community has converged on a hierarchy

The 2026 best-practice writeups (firecrawl, buildfastwithai, nimbalyst, ofox, tembo) all land on the same hierarchy:

  1. Start with skills. Easiest to author, immediate value, low risk. A skill is just instructions; if it’s wrong, you fix the SKILL.md.
  2. Add hooks when you need deterministic enforcement. Skills are advisory; hooks are mandatory. Use hooks when “the model complies most of the time” isn’t good enough.
  3. Use subagents when context isolation or parallelism is worth the 7× token cost. Not “feels parallel” — structurally needs parallel.

Almost every team’s first three weeks of CC-platform work fall in the same order: write a CLAUDE.md, write skills, add a hook or two, maybe later set up a subagent.

Try this

  1. Browse the Anthropic-managed marketplace inside CC: /plugin, Discover tab. Note that before you install anything, the browser tells you what the plugin contributes (commands, hooks, skills, MCP servers) and an estimated per-session token cost.
  2. Install one specific skill — frontend-design from the official marketplace is a good test. Restart your CC session. Ask Claude to design a button. Notice the skill activate (you’ll see the skill name in the response or in tool-use traces).
  3. Write your own minimal skill in a project: create <project>/.claude/skills/<skill-name>/SKILL.md with a tight frontmatter description and a one-paragraph body. Restart, ask Claude to do the thing the skill describes, see it activate. The per-project skill folder is the fastest way to feel skill semantics without dealing with plugin packaging.

Sources