How to Write Your First Claude Code Skill (SKILL.md Guide)

Agent Skills turn Claude Code from a smart assistant into a programmable automation layer. Instead of typing out the same multi-step workflow every session, you define it once in a SKILL.md and invoke it with a single slash command.

This tutorial builds a /smart-commit skill — it runs lint, formats the diff, writes a commit message following conventional commits, and commits. A workflow most developers do manually, every day.

Prerequisites

Claude Code installed (npm install -g @anthropic-ai/claude-code)
A git repository to test in
Basic familiarity with Markdown

Skill directory structure

A skill lives in a directory inside your project (or globally in ~/.claude/skills/):

.claude/
└── skills/
    └── smart-commit/
        ├── SKILL.md       ← required
        └── README.md      ← optional docs

Create the directory:

mkdir -p .claude/skills/smart-commit

Write the SKILL.md

Create .claude/skills/smart-commit/SKILL.md:

# Skill: Smart Commit

## Trigger
/smart-commit

## Description
Runs lint, reviews staged changes, writes a conventional commit message, and commits.

## Tools
- shell: true
- read_file: true

## Instructions

You are helping the developer commit their staged changes cleanly.

Follow these steps in order:

### Step 1: Check for staged changes
Run `git diff --cached --stat` to see what is staged.
If nothing is staged, stop and tell the user to stage files first with `git add`.

### Step 2: Run lint
Run `git diff --cached --name-only` to get the list of changed files.
For JS/TS files, run `npx eslint --no-eslintrc --rule 'no-console: warn' <files>` (skip if no JS/TS files).
If lint fails with errors (not warnings), stop and show the errors. Do not commit broken code.

### Step 3: Review the diff
Run `git diff --cached` to read the full diff.
Understand what changed: what was added, removed, or modified.

### Step 4: Write a commit message
Based on the diff, write a commit message following Conventional Commits format:
- `feat:` for new features
- `fix:` for bug fixes
- `refactor:` for code changes that do not fix a bug or add a feature
- `docs:` for documentation changes
- `chore:` for build process, tooling, config
- `style:` for formatting changes only

Rules:
- Subject line: max 72 characters, imperative mood, no period at end
- If the change is complex, add a blank line and a short body (2-4 sentences max)
- Do not mention file names in the subject — describe the intent

Show the message to the user BEFORE committing and ask for confirmation.

### Step 5: Commit
After user confirms, run:
`git commit -m "<the message>"`

Report the commit hash and subject line when done.

## On Failure
If any step fails, explain what went wrong and what the user needs to fix before trying again.

Test it

Stage some changes and run:

git add src/some-file.ts
claude /smart-commit

Claude will:

Check what is staged
Lint the changed files
Read the diff
Propose a commit message like feat: add user authentication middleware
Wait for your confirmation
Commit

If you want to change the message, just say so and Claude will revise it.

Add a lifecycle hook

Hooks run shell commands at key points. Add this to your SKILL.md to log every skill invocation:

## Hooks
before_start: echo "[smart-commit] Starting at $(date)" >> ~/.claude/skill-log.txt
after_success: echo "[smart-commit] Committed: $(git log -1 --oneline)" >> ~/.claude/skill-log.txt

Now every commit is logged with a timestamp.

Make it team-shareable

Add a skills.json to your project root so teammates get the skill automatically:

{
  "skills": [
    {
      "name": "smart-commit",
      "path": ".claude/skills/smart-commit"
    }
  ]
}

When a teammate runs claude skills sync, Claude Code installs all project skills automatically.

Publish it as a community skill

To share your skill publicly:

Create a GitHub repo named skill-smart-commit
Put SKILL.md at the root
Add a skill.json manifest:

{
  "name": "smart-commit",
  "version": "1.0.0",
  "description": "Lint, review, and commit with conventional commit messages",
  "trigger": "/smart-commit",
  "author": "your-github-username"
}

Others can install it with:

claude skills install github:your-username/skill-smart-commit

Extending the skill

Once the pattern is clear, extending is easy. Some ideas:

Add test running: After lint, run pnpm test --passWithNoTests and stop if tests fail.

Add a co-author line: Append Co-Authored-By: Claude <claude@anthropic.com> to every commit body.

Support monorepos: Detect which package was changed and scope the commit type accordingly (feat(auth):, fix(api):).

Add a dry-run mode: Pass --dry-run as an argument to preview without committing.

Each addition is just more steps in the ## Instructions block. No code required.

Key things to remember

Instructions are plain English — Claude interprets them, you do not write code
Keep steps numbered and explicit — Claude follows them literally
Use stop and tell the user language to create checkpoints
Test incrementally — run the skill after each addition to verify behavior

Next steps

What Are Claude Agent Skills and How Do They Work? — conceptual overview
Browse awesome-agent-skills for more examples
Claude Code skills documentation

What Are Claude Agent Skills and How Do They Work?

A clear explainer on Claude Agent Skills — what they are, how the SKILL.md format works, lifecycle hooks, and why they matter for developer workflows in 2026.

Vishnu Damwala Mar 11, 2026

Agent Skills 8 min read

Vercel AI SDK Tools: One API for Claude and OpenAI Skills

Vercel AI SDK's unified tool interface works with Claude, OpenAI, and Gemini. Write your skill once and switch AI providers without rewriting the agent loop.

Vishnu Damwala Mar 14, 2026