Slash Commands

Principle

A slash command is a Markdown file that defines a workflow. Claude executes it as if it had received that text as a prompt, with the ability to pass dynamic arguments.

CC 2.1.3: Custom slash commands now live in .claude/skills/ with disable-model-invocation: true. The old .claude/commands/ directory is no longer used for new projects.

File locations

.claude/skills/              # Project skills (team) — since CC 2.1.3
├── commit/
│   └── SKILL.md             → /commit
├── review-pr/
│   └── SKILL.md             → /review-pr
└── tech/
    └── deploy/
        └── SKILL.md         → /tech:deploy

~/.claude/skills/            # Global skills (personal)
├── release/
│   └── SKILL.md             → /release
└── sync/
    └── SKILL.md             → /sync

Global skills are available in all sessions, regardless of the project.

SKILL.md frontmatter for user-invocable skills

---
name: commit
description: Generate a conventional commit from staged changes
allowed-tools: [Read, Bash]
disable-model-invocation: true
---

The disable-model-invocation: true flag is what makes a skill user-only — it prevents the model from auto-loading it based on context matching.

Invocation and arguments

/commit                    # No argument
/tech:deploy production    # Positional argument
/release minor             # Version bump

In the Markdown body, arguments are accessible via variables:

Deploy to environment: $ARGUMENTS[0]
With strategy: $ARGUMENTS[1]

If no argument provided: use "staging" by default.

The $0, $1 syntax is equivalent to $ARGUMENTS[0], $ARGUMENTS[1]. The old dot notation ($ARGUMENTS.0) is deprecated since v2.1.19.

Recommended structure

# Command Name

## Objective
What this command does in one sentence.

## Process
1. **Step 1** — detailed instructions
2. **Step 2** — detailed instructions
3. **Step 3** — detailed instructions

## Arguments
- $0: target environment (default: staging)

## Output format
Description of the expected result.

Native commands (built-in since recent versions)

Useful custom skills (user-invocable) to create

Skill (user) vs Skill (auto) vs Agent

A user-invocable skill cannot embed additional reference files without a skill folder — for that, create a full skill directory with SKILL.md plus resource files. A skill has no persistent memory — for that, use an agent.

Practical tips

One skill, one responsibility. As soon as a skill does two distinct things, split it. Name with an action verb first (commit, release, sync) so that the intent is immediately readable in the list of available commands.