Awesome Claude Skills: a curated registry of Claude workflow extensions

Awesome Claude Skills is a curated registry of Claude Skills — self-contained Markdown bundles that extend Claude's behavior with specialized workflows, tool integrations, and bundled scripts. The collection spans roughly 30 hand-picked skills plus 800+ auto-generated SaaS automation stubs, all installable via Claude Code's plugin system.

Why I starred it

Claude Skills launched as a first-class feature in Claude Code — a way to package domain knowledge and tool instructions into reusable units that load on demand. The problem is discoverability: finding a good skill for, say, LangSmith tracing or iOS simulator interaction means knowing it exists in the first place.

This repo solves the discovery layer and adds something more interesting on top: a live plugin (connect-apps-plugin) that bridges Claude to Composio's MCP server, letting Claude fire real API calls — send email, create GitHub issues, post Slack messages — without any bespoke integration code.

How it works

The base skill format is intentionally minimal. Every skill is a directory with a SKILL.md file:

skill-name/
├── SKILL.md          # YAML frontmatter + markdown instructions
├── scripts/          # Executable helpers (Python/Bash)
├── references/       # Reference docs loaded into context on demand
└── assets/           # Templates, fonts, static files used in output

The frontmatter carries just name and description:

---
name: changelog-generator
description: Automatically creates user-facing changelogs from git commits...
---

Claude Code uses description to decide whether to activate the skill for a given task — so quality here matters more than anywhere else in the file. The skill-creator skill's SKILL.md makes this explicit: the metadata is always in context (~100 words), the body loads when triggered, and bundled resources load only when Claude needs them. That's the three-tier loading model the whole system is built around.

The template-skill directory reduces this to its skeleton — a two-line file with placeholder frontmatter and empty body — which is genuinely all you need to get started.

The more architecturally interesting piece is connect-apps-plugin. It's not a skill — it's a Claude Code plugin (.claude-plugin directory) with a commands/setup.md slash-command. When you run /connect-apps:setup, Claude executes:

from composio import Composio
composio = Composio(api_key='YOUR_KEY')
session = composio.create(user_id='claude_user')
print(session.mcp.url)

It then writes the resulting MCP URL directly to ~/.mcp.json:

{
  "connect-apps": {
    "type": "http",
    "url": "<session_mcp_url>",
    "headers": {
      "x-api-key": "<your_key>"
    }
  }
}

After a Claude restart, every app Composio supports becomes a callable MCP tool. OAuth happens the first time you trigger an app action. The setup command also handles merging into an existing ~/.mcp.json — it doesn't clobber existing server configs.

The bulk Composio automation skills (78 original, then 832 added in one PR, totaling over 900 now) each contain real tool slugs from Composio's RUBE_SEARCH_TOOLS API, parameter schemas, known API pitfalls, and quick reference tables. A recent commit shows these were generated by Claude Opus itself:

Upgrade 71 popular app skills with real tool data from RUBE_SEARCH_TOOLS
...
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

That's Claude generating Claude skills — which is either elegant or slightly alarming depending on your perspective.

Using it

Install a skill into Claude Code:

mkdir -p ~/.config/claude-code/skills/
cp -r changelog-generator ~/.config/claude-code/skills/

Then just use Claude Code normally — it activates the skill when context matches the description.

For the Composio plugin:

claude --plugin-dir ./connect-apps-plugin
# then in session:
/connect-apps:setup
# paste API key, restart Claude
claude
# now:
# "Create a GitHub issue titled 'fix login bug' in my-org/my-repo"

The first time you trigger Gmail or Slack actions, Composio handles the OAuth redirect. After that it's stored and reused.

Rough edges

The 800+ auto-generated Composio skills are thin. Each one follows a template and the content quality varies — some have detailed parameter docs, others are generic. There's no testing infrastructure; you won't find a test directory anywhere in the repo. Given most skills are Markdown files, that's somewhat defensible, but the setup command's Python snippet will silently break if composio isn't pip-installed and python3.11 isn't at /opt/homebrew/bin/python3.11 (hardcoded path, macOS-only assumption).

The repo's commit history shows an abrupt expansion: a few dozen curated skills, then a PR adding 874 Composio automation stubs, then a reorg moving those into a composio-skills/ subfolder after apparent pushback. That structural churn is a sign the collection's scope isn't fully settled yet.

There's also no versioning or lock mechanism — skills loaded from this repo are just files you copy. If upstream changes a skill's behavior, you don't know unless you check.

Bottom line

Useful as a reference for what the Skills ecosystem looks like and what people are building. The connect-apps-plugin is the sharpest piece — the three-command setup from zero to live MCP bridge is genuinely low friction for how much it unlocks. Worth cloning if you run Claude Code and want to automate across SaaS apps without writing glue code.