Skills Directory / Guides
What are agent skills?
Agent skills are folders of instructions that teach an AI coding agent a
workflow, domain, or tool it doesn't handle well out of the box. One
folder, one SKILL.md file, loaded only when relevant.
The mechanism
Every skill is a directory containing a SKILL.md: YAML
frontmatter (a name and a description) followed
by markdown instructions. At startup the agent reads only the descriptions
— a line or two each. When your task matches one, the agent loads that
skill's full instructions and follows them. This lazy loading is the trick:
you can install dozens of skills and pay context cost only for the one in
use.
Why they exist
Models are capable but inconsistent. They know what TDD is, yet won't reliably practice it; they can write wrangler commands, but from stale training data. Skills close the gap between can and does — encoding the workflow, current facts, and known pitfalls so behavior stops depending on how you phrased the prompt. It's the same shift companies make with human processes: from tribal knowledge to written procedure.
The standard
Anthropic introduced the format for Claude Code in 2025 and published it as an open specification at agentskills.io. It's now supported (natively or with minimal adaptation) by Codex, Cursor, Gemini CLI, and a growing list of agents — see skills for other agents. Skills are complementary to MCP servers and plugins; the taxonomy is untangled in skills vs plugins vs MCP.
The ecosystem
Official vendor teams — Anthropic, Cloudflare, Vercel, Hugging Face, Expo — now ship skills for their own platforms, which matters because platform teams keep them current. Around them sits a community of workflow libraries (like Superpowers) and curated lists. Our directory tracks the collections worth knowing and indexes 195+ individual skills, each with a copy-paste install prompt.
Where to start
- Browse the directory and pick one skill that matches a task you do weekly.
- Install it (or paste its install prompt into your agent).
- When you notice your agent repeating a mistake no skill covers — write your own. That's where the real compounding starts.