Slash Command Design

Slash Command Design

This skill provides interpretive guidance and best practices for creating Claude Code slash commands. It helps you understand how to create excellent commands.

Fundamentals

Prerequisites: This skill builds on box-factory-architecture. If not already loaded, core concepts like isolation model and component selection may be unclear.

Four core principles for slash command design:

Commands Are User-Triggered Prompt Expansions

Commands execute when users type /command-name. They expand to prompts that Claude processes. Typically invoked directly by user, but Claude Main Agent can also invoke them autonomously via the SlashCommand tool.

Use Delegation When It Provides Clear Benefits

Commands should delegate to sub-agents to achieve: context isolation, parallelization, or leveraging the abilities of specialized agents.

Delegate when:

• The goal of the command can be satisfied clearly and cleanly by a sub-agent (preserves context)
• There is a clear specialty agent designed for the task (eg, 'researcher', 'test-runner')
• The task can be parallelized into independent subtasks (efficiency)

Do not delegate when:

• There is no clear, appropriate agent
• The task needs user interaction or broad context from the main conversation

Arguments Are Simple String Substitutions

Use $1, $2 for positional arguments or $ARGUMENTS for all arguments as a single string.

Tool Restrictions Are Optional

Commands can use any tools needed for their task. Use allowed-tools frontmatter only when you specifically want to constrain scope (security-sensitive operations, read-only review commands, etc.).

Workflow Selection

Quick Start

Create a command file based on the path-resolution rules from box-factory-architecture component-paths.

---
description: Brief description of what this command does
argument-hint: expected-args
---

Command prompt content here.
Use $1, $2 for individual arguments or $ARGUMENTS for all.

With delegation:

---
description: Run full test suite and analyze failures
---

Use the test-runner agent to execute all tests and provide detailed failure analysis.

Official Documentation

Claude Code changes rapidly and is post-training knowledge. Fetch these docs when creating commands to ensure current syntax:

• https://code.claude.com/docs/en/slash-commands - Core specification and examples
• https://code.claude.com/docs/en/settings#tools-available-to-claude - Verify tool names

Decision Framework

Command vs Agent vs Skill

Quality test: If you want this to happen automatically based on context, it's an agent, not a command.

Use Command when:

• User wants explicit control over when it runs
• Simple, deterministic operation
• Wrapping a bash script or tool sequence
• "I want to type /something to make X happen"

Use Agent when:

• Need isolated context window (large outputs shouldn't pollute main conversation)
• Want parallelization (multiple agents working simultaneously)
• Reusable workflow (same logic invoked by multiple commands)
• Want autonomous delegation based on context detection

Use Skill when:

• Multiple contexts need the same knowledge
• Substantial procedural expertise
• Progressive disclosure would save tokens

Deep dive: Load the box-factory-architecture skill for full component selection guidance. Traverse when: ambiguous component choice, edge cases not covered by summary. Skip when: summary above clearly answers the question.

Frontmatter Fields (Official Specification)

All fields are optional:

Best Practices for Frontmatter (Opinionated Guidance)

Always include description even though it's optional - improves discoverability and Claude's ability to use the SlashCommand tool for autonomous command invocation. Use strong language guiding Claude when it is appropriate to use this tool.

Use argument-hint when arguments are expected - helps users understand command syntax in /help output.

Use allowed-tools when:

• Security-sensitive operations (restrict to specific git commands, read-only access)
• Single-purpose commands with clear scope (status checks, simple queries)
• Preventing accidental file modifications (read-only tools only)

Don't use allowed-tools when:

• Command delegates to agent (let agent define its own tool restrictions)
• Need full tool access for complex operations
• Multiple tool combinations required

Use disable-model-invocation only in edge cases where SlashCommand tool auto-invocation causes issues. Most commands don't need this field.

Argument Patterns

Official Specification

All arguments as single string:

$ARGUMENTS

Example: /fix-issue 123 high-priority → $ARGUMENTS = "123 high-priority"

Individual positional arguments:

$1, $2, $3, etc.

Example: /review-pr 456 high alice → $1="456", $2="high", $3="alice"

When to Use Which Pattern (Best Practices)

Use $1, $2 when:

• Need arguments in different parts of the prompt
• Want to provide defaults or conditional logic
• Arguments have distinct semantic meanings

Use $ARGUMENTS when:

• Passing all arguments directly to agent
• Single compound argument (like a description or message)
• Don't need to reference specific positions

Argument Handling

Commands use simple string substitution ($1, $ARGUMENTS). Claude handles the actual interpretation and validation as part of processing the expanded prompt.

---
description: Deploy to specified environment
argument-hint: environment
---

Deploy to $1 environment. Validate the environment name, check deployment prerequisites, and proceed with appropriate rollback strategy.

Complex multi-step workflows with error recovery may benefit from agent delegation for context isolation, but straightforward argument handling works fine in commands.

Advanced Features

Bash Execution with ! Prefix (Official Specification)

Execute bash commands before the prompt runs:

---
allowed-tools: Bash(git:*)
---

!git status

Review the git status above and suggest next steps.

When to use: Quick status checks before analysis (!git status, !npm test), gathering context for Claude's response, read-only commands that inform the prompt.

Avoid for: Multi-step conditional bash workflows, write operations without confirmation.

File References with @ Prefix (Official Specification)

Include file contents in the prompt:

Review @src/utils/helpers.js for potential improvements.

Multiple files: Compare @src/old.js with @src/new.js

When to use: Simple file review tasks, direct comparison of specific files, quick analysis of known paths.

Avoid for: Large files (token limits), dynamic file discovery where paths aren't known upfront.

Subdirectory Namespacing (Official Specification)

Organize commands in subdirectories:

commands/frontend/component.md → /component (project:frontend)
commands/backend/endpoint.md → /endpoint (project:backend)

Command name comes from filename, subdirectory appears in /help as namespace label.

When to use: Large projects with distinct subsystems, commands specific to tech stack areas, avoiding command name collisions.

Avoid for: Small projects (flat structure is clearer), over-categorization (3-5 commands don't need subdirectories).

Tool Restriction Patterns (Official Specification)

Restrict commands to specific tools via allowed-tools:

---
description: Show git status
allowed-tools: Bash(git status:*)
---

Run `git status` and display the output.

Common patterns:

• Bash(git:*) - Git operations only
• Bash(git status:*) - Specific git command
• Read, Grep, Glob - Read-only file operations
• Omit field entirely to inherit conversation tools

When to use: Security-sensitive operations, single-purpose commands with clear scope, preventing accidental file modifications.

Best Practices (Opinionated Guidance)

When to Delegate vs Handle Directly

Commands can do substantial work. Delegation to sub-agents isn't about complexity—it's about specific benefits:

Delegate when you need:

• Context isolation — Large outputs (test results, logs, file contents) shouldn't clutter the main conversation
• Parallelization — Multiple independent tasks that can run simultaneously
• Reusable workflows — Same logic invoked by multiple commands or contexts

Handle directly when:

• Task is straightforward even if it involves multiple steps
• Output is reasonably sized and useful in main conversation
• No need for the workflow to be reusable elsewhere

Deep dive: Load the box-factory-architecture skill for multi-component patterns. Traverse when: designing workflows that span multiple components. Skip when: single command design.

Generation Pattern

For creating files/code, be specific about structure and requirements:

---
description: Create a new React component with TypeScript
argument-hint: component-name
---

Create a new React component named `$1` in the components directory.

Include:
- TypeScript interface for props
- Basic component structure with proper typing
- Export statement
- Test file in __tests__ directory

Follow project conventions for imports and file structure.

Real-World Examples

Direct command (no delegation needed):

---
description: Show current git branch and status
allowed-tools: Bash(git:*)
---

!git branch --show-current
!git status --short

Current branch and working tree status displayed above.

Direct command with file operations:

---
description: Review package.json dependencies
---

Read package.json and analyze the dependencies. Flag any outdated packages, security concerns, or unused dependencies.

With delegation (context isolation for large output):

---
description: Run full test suite and analyze failures
---

Use the test-runner agent to execute all tests and provide failure analysis.

The test-runner agent is appropriate here because test output can be large and detailed analysis benefits from isolated context.

With delegation (reusable workflow):

---
description: Commit all changes with smart message generation
argument-hint: [commit-message]
---

Use the git-committer agent. Pass arguments: $ARGUMENTS

The git-committer agent encapsulates commit workflow logic reused across multiple entry points.

Component Paths

Deep dive: Load the box-factory-architecture skill for component path resolution. Traverse when: need to determine exact file paths, working in plugin context. Skip when: caller specified path or using default project location (.claude/commands/).

Quality Checklist

Before finalizing a command:

Structure (from official docs):

• Valid YAML frontmatter (if used)
• Proper markdown formatting
• Filename is kebab-case (becomes command name)

Best Practices (opinionated):

• Includes description field for discoverability
• Uses argument-hint if arguments expected
• Action-oriented (not knowledge storage)
• Clear, single-purpose design
• Tool restrictions only if specifically needed for security/scope

Delegation Decision:

• If delegating: justified by context isolation, parallelization, or reusability need
• If not delegating: task fits naturally in main conversation context