Skills: Custom Slash Commands

Type /deploy staging and this entire workflow runs. $ARGUMENTS becomes "staging".

What Are Skills?

Skills extend Claude's capabilities with instructions, templates, and step-by-step workflows, all written in markdown. They come in two flavors:

How Skills Actually Work (Under the Hood)

When you type /deploy staging, here's what actually happens inside Claude Code:

1. File lookup: Claude Code searches for SKILL.md in .claude/skills/deploy/ (project), then ~/.claude/skills/deploy/ (personal), then installed plugins.
2. Frontmatter parsed: YAML frontmatter is read for name, description, allowed-tools, and context settings.
3. String substitution: $ARGUMENTS is replaced with everything after the slash command. ${CLAUDE_SKILL_DIR} becomes the absolute path to the skill's directory.
4. Context injection: The processed markdown is injected into Claude's conversation as instructions. If context: fork is set, it runs in an isolated subagent.
5. Execution: Claude follows the instructions, restricted to the tools listed in allowed-tools (if specified).

Anatomy of a SKILL.md

---
name: my-skill              # Slash command name (/my-skill)
description: What it does    # Shown in skill list
disable-model-invocation: true  # Only runs when YOU type /my-skill
context: fork                # Run in isolated subagent (optional)
allowed-tools: Bash(*), Read # Restrict which tools the skill can use
---

# Instructions for Claude

Use $ARGUMENTS for user input.
Use ${CLAUDE_SESSION_ID} for the current session.
Use ${CLAUDE_SKILL_DIR} for the skill's directory path.

## Steps

1. Do this first
2. Then do this
3. Finally, report the result

Frontmatter Options

Built-in Skills

Claude Code ships with several powerful skills out of the box:

Notable Community Skills

The community has built some incredible skills. These are the standouts:

Creating Your Own Skill

Step 1: Create the file

Step 2: Write the SKILL.md

---
name: review
description: Run a thorough code review on staged changes
context: fork
---

# Code Review

Review the staged git changes (`git diff --cached`) for:

## Checklist
1. **Security** — SQL injection, XSS, command injection, secrets in code
2. **Performance** — N+1 queries, unnecessary loops, missing indexes
3. **Correctness** — Edge cases, null checks, off-by-one errors
4. **Style** — Consistent naming, no dead code, clear variable names
5. **Tests** — Are changes covered? Any missing edge case tests?

## Output Format
- List each issue with file, line number, and severity (critical/warning/info)
- End with a summary: "Ready to merge" or "Needs changes"

---
name: research
description: Deep research on a topic with web search
allowed-tools: WebSearch, WebFetch, Read, Write
---

# Research: $ARGUMENTS

1. Search the web for recent, authoritative sources on "$ARGUMENTS"
2. Find at least 5 distinct sources (prefer: official docs, peer-reviewed, reputable outlets)
3. For each source, extract:
   - Key findings or claims
   - Supporting data or statistics
   - Source URL and publication date
4. Write a research summary to `research-$ARGUMENTS.md` with:
   - Executive summary (3-5 sentences)
   - Detailed findings organized by theme
   - All source URLs as inline citations
   - Gaps in current knowledge / areas needing more research

Step 3: Use it

Skill Scopes

Advanced Features

Forked Context

Adding context: fork runs the skill in an isolated subagent. This keeps heavy operations (like /batch) from cluttering your main conversation. The result is returned when done. Note: forked skills consume additional tokens for the subagent conversation.

Tool Restrictions

allowed-tools limits what the skill can do. Use glob patterns:

• Bash(npm *) — only npm commands
• Read, Glob, Grep — read-only access
• Bash(git *), Edit, Write — full dev access

String Substitution

What Does It Cost?