The short version

A skill is an automated workflow. You define a process once - what steps to follow, what tools to use, what the output should look like. The agent runs it on command. Every time. The same way. Repeatable work stops being manual work.

What a skill is

In Claude Code, a skill is a Markdown file inside your project's .claude/skills/ folder. Each skill lives in its own subfolder - .claude/skills/skill-name/SKILL.md - and defines one automated process: the steps, the tools, and the expected output. When you invoke it, the agent executes the workflow exactly as written - not its own interpretation of what you might want.

At AOL 4 you are no longer writing prompts for individual tasks. You are building automated workflows that live in your project permanently. Other AI coding tools use the same concept under different names and file structures.

Example automations
check-design Audit all components against the design system. Flag any border color or spacing that does not match the tokens.
generate-component Create a new Svelte component following the project conventions - props interface, scoped styles, barrel export.
verify-build Run the build and type check. Report all errors and warnings before anything is committed.

What automation gives you

Processes run, not explained
You define the workflow once. It runs the same way every time. No prompt variation, no inconsistency, no re-explaining how to do the same task.
Quality becomes structural
The skill enforces the standard. Output quality no longer depends on how you phrased the request that day.
Complex workflows in one command
A skill can chain multiple steps - search, verify, fix, report - and run them as a single automated sequence.

Skills build up into a library

The first skill you write is useful. Ten skills is a project automation library. Each new skill removes another recurring manual task from your workflow. Over time the agent handles more of the routine work and you handle more of the high-level decisions. Skills compound.

When to create a skill

The signal is regularity, not complexity. A simple task you run every day is a better candidate than a complex task you run once a year.

Write a skill
  • Any process you run daily or weekly
  • Any check you do before every commit or deploy
  • Any workflow that takes more than one step and repeats regularly
  • Any task where inconsistency has cost you time
Keep it a prompt
  • Tasks you do a few times a year
  • One-off requests with no clear pattern
  • Exploratory work where the steps change every time

Skill scope

Skills exist at three levels. Most live at the project level - specific to your stack and workflows. The scope controls where the skill is available.

Global
~/.claude/skills/
Available in all your projects. Good for universal workflows that apply everywhere.
Project
.claude/skills/
Specific to this project. Your stack, your conventions, your deployment process. Most skills live here.
Local
.claude/skills/*.local/
Personal workflows not shared with the team. Add to .gitignore.

Skills and /docs work together

A skill defines the process. A doc holds the standard. The skill tells the agent to check border colors. The design system doc tells it what the correct border colors are. They are separate for good reason: standards change independently from processes.

/docs - The Standard What the output is supposed to look like. Maintained separately. Updated when standards change.
Skills - The Process How to verify or produce the output. Points to the docs for the standard. Updated when the process changes.

Where skills live

Skills belong in the execution layer, separate from your code and your knowledge. The three layers stay clean by design.

/docs
Source of truth. Standards, decisions, architecture.
/src
Application code.
/.claude/skills/
Execution and automation. If it acts, it belongs here.

Review before you run

A skill is not just instructions - it is permission for the agent to act. Running a skill can mean shell access, file system writes, environment variable reads, network calls, or deploy triggers. That is real authority inside your project.

Copying a public skill without reviewing it imports that authority blind. Before running any skill you did not write yourself, read every step. Check for shell commands, network calls, and anything that touches outside your project boundary. Study public skills for patterns - then rewrite them for your system.

Agentic power without review is automated risk.

The skill is the workflow. The hook is the trigger.

At AOL 4 you still invoke skills manually. You type /check-design and the workflow runs. That is already a real shift - you are executing an automation, not writing a prompt.

AOL 5 removes the manual step. Hooks fire skills automatically at the right moment - after a file changes, before a commit, when a build fails. You do not ask the agent to check its work. It checks automatically, every time the trigger fires.

What AOL 4 gives you

You enter the Integration Zone at its peak. The agent has context and a library of automated workflows - it can handle complex, multi-step tasks on its own. You spend less time directing and more time reviewing. The human in the loop shifts from giving instructions to approving results.

How to set it up

One folder. One file. That is the entire installation. The skill system in Claude Code does not require configuration - it requires content.

1
Create the skills folder
.claude/skills/
Inside your project root. If it already exists, skip this step.
2
Add a skill file
.claude/skills/your-skill-name/SKILL.md
One subfolder per workflow. The folder name becomes the command name. Keep it short and specific.
3
Invoke it
/your-skill-name
Type the slash command in Claude Code. The agent reads the file and runs the workflow.

The starter file below includes a working example skill - a code review workflow - with annotations explaining each section. Download it, read it, then adapt the workflow sections for your own process.

Starter file
SKILL.md Starter A working code review skill with every section annotated - description, methodology, output format, edge cases, and examples. Download, adapt, and invoke immediately.
Next step: AOL 5 - Integrated
Hooks trigger your skills automatically. You built the automations. Now they fire on their own.
See AOL 5