Agent skills are folders that package instructions, reference material and sometimes scripts for an AI coding agent. The entry point is a SKILL.md file: a Markdown document with a short header describing when the skill applies, followed by the procedure the agent should follow. Coding agents such as Claude Code and Codex load these skills when a task matches their description, so you can teach an agent a workflow once and reuse it across projects.

This tutorial covers three things: where to find skills, how to install one, and how to write your own.

Where to find agent skills

Stackora lists open-source skills under AI Skills and the broader Agent skills catalog. Several widely used libraries are available under MIT or Apache-2.0 licenses, which means you can read, copy and modify them. Examples in the catalog include:

  • skills by mattpocock, a set of skills for AI coding tools that focuses on alignment, shared language and red-green-refactor testing. It is MIT licensed.
  • agent-skills by Addy Osmani, a library of production-grade engineering skills covering spec, plan, build, test, review and ship workflows. It is MIT licensed.
  • ui-ux-pro-max-skill, a skill that supplies design intelligence for building UI and UX across platforms and frameworks. It is MIT licensed.
  • archify, a skill that turns an idea, plan or codebase into an interactive diagram, for Claude Code, Codex and similar agents. It is MIT licensed.
  • marketingskills, covering CRO, copywriting, SEO and analytics. It is MIT licensed.
  • scientific-agent-skills, a collection of 177 skills for scientific work. It is MIT licensed.
  • DeepTutor, a tutoring system with skill support, licensed under Apache-2.0.

Before installing, check the license, the publisher and the star count, then open the SKILL.md and read it. A skill is executable guidance: it tells the agent what to do, so review it as you would any code you run.

Step 1: Pick a skill and read its SKILL.md

  1. Choose a skill whose description matches a task you repeat, for example a review workflow or a diagramming step.
  2. Open its SKILL.md and check three things: the trigger description in the header, the steps it asks the agent to take, and any scripts or commands it runs.
  3. Note any tools it expects, such as shell access or file writes, and confirm they are acceptable for your environment.

Step 2: Install the skill

Most skills are distributed as a Git repository. The general procedure is the same for Claude Code and other agents that read the same folder layout:

  1. Clone the repository or download the release.
  2. Copy the skill folder (the directory that contains SKILL.md) into the skills directory your agent reads. For Claude Code this is typically a skills folder in your personal configuration directory for global use, or a .claude/skills folder inside a project for team use.
  3. Restart the agent or start a new session so it indexes the new skill.
  4. Confirm the skill is visible, then give the agent a task that matches the skill's trigger description.

Keep each skill in its own folder. If a repository contains several skills, copy only the ones you need to avoid cluttering the agent's context.

Step 3: Test the skill on a small task

Run the skill on a low-risk task in a branch or scratch directory. Watch which steps the agent follows and whether it asks for approval before running commands. If the output is off, edit the local copy of SKILL.md rather than changing the upstream repository, and re-test.

Step 4: Write your own SKILL.md

A useful skill is narrow and procedural. Use this structure:

The header

Start with a name and a description. The description is what the agent reads to decide whether to load the skill, so state the trigger conditions plainly, for example: "Use when writing or reviewing database migrations in this repository."

The procedure

Write numbered steps the agent can execute. Name the files, commands and conventions involved. Specify what done looks like, such as tests passing or a summary to the user.

Reference material

Put long background documents, style guides or schemas in separate files in the same folder and tell the agent when to read them. This keeps the main SKILL.md short enough to load quickly.

A minimal example

A skill for reviewing changelog entries might contain a header saying it applies when editing CHANGELOG.md, followed by three steps: check the entry format, confirm each change links to a pull request, and flag any entry missing a category. That is enough to be useful and easy to verify.

Step 5: Iterate and share

Revise the skill after each run where the agent misread a step. Once it works reliably, commit the folder to your repository so teammates get the same behavior. If you publish it, choose a license explicitly; the open-source skills above use MIT, which permits reuse with attribution.

Recommendation

Start with one established skill from the catalog, such as skills or agent-skills, to learn the format by reading it. Then write a short skill for a single repeated task in your own project. Keep the procedure narrow, test it on real work, and version it alongside your code. Browse the full Agent skills catalog and the publishers index to find more maintainers whose skills match your stack.