Command Writer Skill

Expert guide for creating Claude Code custom slash commands.

Official Docs: https://code.claude.com/docs/en/slash-commands

1. Quick Reference

YAML Frontmatter Template

---
description: Shows in slash command menu
argument-hint: <arg1> [arg2]
project-agnostic: false  # default; set true only if truly reusable across any project
allowed-tools:
  - Read
  - Write
  - Edit
  - Bash
---

Minimal Command Template

---
project-agnostic: false  # Required - default value
---

Describe what this command does and instruct Claude how to execute it.

That's it. Save as

.md

File Locations

.claude/commands/<name>.md

.claude/commands/<name>.md

~/.claude/commands/<name>.md

Frontmatter Fields

description

/

allowed-tools

argument-hint

<issue-number>

model

disable-model-invocation

true

project-agnostic

false

true

Variable Syntax

$ARGUMENTS

/deploy $ARGUMENTS

$1

$2

/pr $1

@<filepath>

@src/config.ts

git status

BANG syntax: Write exclamation mark (!) followed by backtick-wrapped command, e.g., BANG then backtick-command-backtick

WARNING - Documentation Safety: When DOCUMENTING this syntax in skills or command files, NEVER use the literal pattern (exclamation + backtick). Claude Code's parser executes this pattern during file loading, causing errors. Use safe alternatives: "BANG-backtick",

[BANG-backtick: cmd]

, or text descriptions like "exclamation + backtick".

2. Validation Checklist

Before saving any command:

• Location:

  .claude/commands/

  (project) or

  ~/.claude/commands/

  (global)
• Extension: File ends with

  .md

• Description: Present in frontmatter if command is non-trivial
• Variables: Using

  $ARGUMENTS

  not

  {args}

  or

  {{input}}

• File refs: Using

  @path

  not

  {{file:path}}

• Bash output: Using BANG-backtick syntax, not

  $(cmd)

  or

  {{shell:cmd}}

• Safety: Destructive ops have confirmation gates
• Paths: No hardcoded absolute paths (use relative or variables)
• Project-agnostic: REQUIRED - Explicitly set

  project-agnostic: false

  (default) or

  true

  (if truly reusable)

3. Command Patterns

Pattern 1: Minimal (Simple Prompt Expansion)

File:

.claude/commands/explain.md

---
project-agnostic: true  # Reusable across any project
---

Explain the code in the current file. Focus on:
- What it does
- Key design decisions
- Potential improvements

Usage:

/explain

Pattern 2: With Variables

File:

.claude/commands/review-file.md

---
description: Review a specific file for code quality
argument-hint: <filepath>
project-agnostic: true  # Reusable across any project
---

Review this file for code quality issues:

@$ARGUMENTS

Check for:
- Bug risks
- Performance issues
- Readability improvements

Usage:

/review-file src/auth.ts

Pattern 3: Full Frontmatter

File:

.claude/commands/refactor.md

---
description: Refactor code with specific pattern
argument-hint: <pattern> <filepath>
allowed-tools:
  - Read
  - Edit
  - Grep
project-agnostic: true  # Reusable across any project
---

Apply refactoring pattern "$1" to file:

@$2

## Patterns
- `extract-function`: Extract repeated code into functions
- `simplify-conditionals`: Reduce nested if/else
- `add-types`: Add TypeScript types to untyped code

## Rules
- Preserve behavior exactly
- Add tests if missing
- Run existing tests after changes

Usage:

/refactor extract-function src/utils.ts

Pattern 4: Git Operations (Safety-Critical)

File:

.claude/commands/release.md

---
description: Create release branch and tag
argument-hint: <version>
project-agnostic: false  # Expects project-specific version files (package.json, etc.)
allowed-tools:
  - Bash
  - Read
---

# Release Command

Create release for version: $ARGUMENTS

## Pre-Flight Checks

1. **Verify clean state**:
   - Run: git status --porcelain (via BANG-backtick)
   - Must be empty; if dirty: STOP and list uncommitted changes

2. **Verify branch**:
   - Run: git branch --show-current (via BANG-backtick)
   - Must be `main` or `master`
   - If not: STOP with "Switch to main branch first"

3. **Verify version format**:
   - Must match `vX.Y.Z` (semver)
   - If invalid: STOP with format instructions

## Confirmation Gate

**STOP HERE** and show user:

Release Summary:

• Version: $ARGUMENTS
• Branch: [current branch]
• Last commit: [short hash + message]

Proceed? (yes/no)

Wait for explicit "yes" before continuing.

## Execution (only after confirmation)

1. Create release branch:
   ```bash
   git checkout -b release/$ARGUMENTS

2. Update version files:

   • package.json (if exists)
   • pyproject.toml (if exists)
   • VERSION file (if exists)

3. Commit version bump:

   git add -A
   git commit -m "chore: bump version to $ARGUMENTS"
   

4. Create annotated tag:

   git tag -a $ARGUMENTS -m "Release $ARGUMENTS"
   

5. Show next steps (DO NOT auto-push):

   Release prepared locally. To publish:
   git push origin release/$ARGUMENTS
   git push origin $ARGUMENTS
   

Abort Conditions

• Any git command fails: STOP immediately
• User says "no" at confirmation: STOP and revert any changes
• Version already exists: STOP with error

Usage: `/release v2.1.0`

---

## 4. Writing Effective Commands

### Frontmatter Best Practices

```yaml
---
# REQUIRED: Explicit project-agnostic declaration
project-agnostic: false  # default - most commands are project-specific
# OR
project-agnostic: true   # only if truly reusable across any project

# GOOD: Clear, actionable description
description: Generate unit tests for a function

# BAD: Vague or missing
description: Tests
---

• project-agnostic: REQUIRED - Must be explicitly set. Default is

  false

• description: Verb phrase describing outcome
• argument-hint: Show expected format (

  <issue-id>

  ,

  <file> [options]

  )
• allowed-tools: Restrict to minimum needed (security + focus)

Variable Usage

# All arguments as single string
Process this request: $ARGUMENTS

# Positional arguments
Compare $1 (old) with $2 (new)

# File inclusion (contents inline)
Analyze this code:
@src/main.py

# Dynamic context from shell (BANG-backtick syntax)
Current git state:
[use exclamation + backtick-wrapped: git log --oneline -5]

Safety Patterns

Confirmation Gate (for destructive operations):

## Confirmation Required

Before proceeding, show:
- What will be modified/deleted
- Estimated impact

Ask: "Proceed with [action]? Type 'yes' to confirm"

WAIT for explicit "yes". Any other response = abort.

Branch Protection:

## Branch Check

Current branch: [BANG-backtick: git branch --show-current]

If branch is `main`, `master`, or `production`:
- STOP immediately
- Error: "Cannot run on protected branch"

Auth/Permission Check:

## Permission Check

Verify access before proceeding:
[BANG-backtick: aws sts get-caller-identity 2>&1]

If error or wrong account: STOP with instructions.

Section Structure

Recommended sections for complex commands:

1. Context - What this does, when to use
2. Pre-Flight Checks - Validations before action
3. Confirmation Gate - User approval point
4. Execution - Step-by-step actions
5. Rollback - How to undo if needed
6. Next Steps - What user should do after

5. Anti-Patterns

Missing Description

# BAD: No frontmatter, unclear purpose
Do the thing with $ARGUMENTS

# GOOD: Clear intent
---
description: Deploy to specified environment
argument-hint: <env>
project-agnostic: false  # Requires project-specific deployment scripts
---

Deploy application to $ARGUMENTS environment...

Incorrect Variable Syntax

# BAD: Wrong syntax (won't work)
Process {{input}}
Read file: ${filepath}
Run: $(git status)

# GOOD: Correct Claude Code syntax
Process $ARGUMENTS
Read file: @$1
Run: [BANG-backtick: git status]

No Confirmation for Destructive Ops

# BAD: Immediate deletion
Delete all files matching: $ARGUMENTS
[BANG-backtick: rm -rf $ARGUMENTS]

# GOOD: Confirmation required
## Files to Delete
[BANG-backtick: find . -name "$ARGUMENTS" -type f]

**CONFIRM**: Type "yes" to delete these files.
[Wait for confirmation before any rm command]

Hardcoded Paths

# BAD: Won't work on other machines
Read config: @/Users/john/project/config.json

# GOOD: Relative or variable
Read config: @./config.json
Read config: @$1

Unmarked Project-Specific Commands

# BAD: References project structure without declaring
---
description: Run our custom linter
---
Run ./scripts/custom-lint.sh on $ARGUMENTS

# GOOD: Declares project-specific nature
---
description: Run our custom linter
project-agnostic: false
---
# Note: Requires ./scripts/custom-lint.sh from this project
Run ./scripts/custom-lint.sh on $ARGUMENTS

Over-Permissive Tools

# BAD: Full access for simple task
---
allowed-tools:
  - Bash
  - Read
  - Write
  - Edit
  - WebFetch
---
Just format this JSON: $ARGUMENTS

# GOOD: Minimal permissions
---
allowed-tools:
  - Read
---
Just format this JSON: $ARGUMENTS

Missing Error Handling

# BAD: Assumes success
Run migration then deploy.

# GOOD: Handle failures
1. Run migration: [BANG-backtick: npm run migrate]
2. If migration fails: STOP, show error, suggest rollback
3. Only if migration succeeds: proceed to deploy

Unsafe Bash Syntax in Documentation

# BAD: Literal pattern triggers execution during file load
# Example of git status using exclamation-backtick syntax...
# (the literal pattern causes: "Error: Bash command failed")

# GOOD: Use safe alternatives in documentation
# Example of git status using BANG-backtick syntax...
# Or: [BANG-backtick: git status]
# Or: "exclamation mark followed by backtick-wrapped command"

Why This Matters: When skill or command files contain the literal exclamation + backtick pattern (even in examples or comments), Claude Code's parser executes it as bash during file loading. This causes immediate errors and prevents the file from loading properly.

Safe Documentation Patterns:

• Use "BANG-backtick" as a text placeholder
• Use

  [BANG-backtick: cmd]

  format in examples
• Write "exclamation + backtick" in prose
• Never combine the actual symbols in documentation

6. Pre-Share Checklist

Before committing or sharing a command:

• Tested with various inputs (happy path + edge cases)
• Description accurately reflects behavior
• No secrets/tokens hardcoded
• Paths are relative or parameterized
• Destructive operations have confirmation gates
• Protected branches are guarded
• Error cases handled gracefully
• Rollback instructions included (if applicable)
• Follows team naming conventions
• Documentation matches implementation

• project-agnostic

  explicitly set (true for reusable, false for project-specific) - REQUIRED

Quick Examples

/explain

/review $1

/test $ARGUMENTS

/deploy $1

/release $1