Agents & Commands Creator

Comprehensive guide for creating agents and commands in Claude Code plugins. Ensures proper YAML frontmatter, markdown structure, and plugin.json integration.

Critical Difference from Inline Configuration

⚠️ COMMON MISTAKE: Trying to define agents/commands as JSON objects inside plugin.json

❌ WRONG:

{
  "agents": [
    { "name": "my-agent", "description": "..." }
  ]
}

✅ CORRECT:

{
  "agents": [
    "./agents/my-agent.md"
  ]
}

Each agent and command is a separate markdown file, not a JSON object. The plugin.json only lists the file paths in arrays.

---

Part 1: Creating Commands

Commands are CLI-like utilities invoked with /command-name. Each command is a single .md file with YAML frontmatter.

Command File Structure

Location: commands/command-name.md

Basic Template:

---
name: kebab-case-name
description: Brief description of what the command does
---

# /command-name

One-sentence summary of purpose.

## Usage

/command-name [optional-args]

## What This Does

Paragraph explaining functionality.

## Output

What the command produces or returns.

## When to Use

When should a user invoke this command?

## Related Commands

Links to complementary commands.

Required YAML Fields

Field	Type	Required	Example
name	string	✓	"review-code"
description	string	✓	"Get instant feedback on code"

Command Creation Best Practices

1. Keep it focused

• One command = one clear purpose
• Example: /review-code does code review, not code generation

2. Use imperative language

• ✅ "Review code for patterns"
• ❌ "This command reviews your code"

3. Simple bash commands

• Use only if absolutely necessary
• NO complex pipes or chains
• Single operations only

4. Clear output format

• Describe exactly what the command returns
• Include examples if helpful
• List any side effects (file creation, etc.)

Example: /review-starters Command

File: commands/review-starters.md

---
name: review-starters
description: Run full validation on your 20 starter scripts
---

# /review-starters

Comprehensive audit of your 20 starter scripts from your unity-starter repository.

## What This Does

Validates all starter scripts for:
- Code quality and consistency
- Best practices (Singleton patterns, error handling)
- Integration health between scripts
- Memory safety (pooling, cleanup, leaks)
- Modern pattern readiness

## Usage

/review-starters

## Output

Detailed report with:
1. Critical Issues
2. Best Practice Suggestions
3. Integration Warnings
4. Modern Pattern Gaps
5. Ready Status

## When to Use

**Pre-jam** (3-4 days before): Run to validate starter scripts before jam begins.

## Related Commands

- `/jam-status` - Quick project health check
- `/design-feature` - Plan a new feature

Command Documentation Guidelines

• Title - Start with /command-name (matches command name)
• One-liner - First sentence explains purpose
• What This Does - 2-3 bullet points describing functionality
• Usage - Show command syntax with examples
• Output - Describe what command produces
• When to Use - Help users understand when to invoke
• Related Commands - Link to complementary commands

---

Part 2: Creating Agents

Agents are autonomous assistants invoked with @agent-name. Each agent is a single .md file with YAML frontmatter followed by detailed instructions.

Agent File Structure

Location: agents/agent-name.md

Basic Template:

---
name: kebab-case-name
description: Brief description of agent purpose and capabilities
---

# @agent-name

One-sentence summary of role.

## What This Agent Does

The agent-name is your [role description]. Invoke with `@agent-name` to:

- **Capability 1** - Description
- **Capability 2** - Description
- **Capability 3** - Description

## Capabilities

✓ Specific capability 1
✓ Specific capability 2
✓ Specific capability 3

## Usage Examples

@agent-name: [example task 1] → Get: [example output 1]

@agent-name: [example task 2] → Get: [example output 2]

## When to Invoke

- **Situation 1** - Use agent for...
- **Situation 2** - Use agent when...
- **Situation 3** - Use agent for...

## Works With

- `/command-name` command
- `@other-agent` agent
- Specific tools or workflows

Required YAML Fields

Field	Type	Required	Example
name	string	✓	"code-reviewer"
description	string	✓	"Code review focused on quality"

Agent Creation Best Practices

1. Clear role definition

• What is this agent's specialty?
• What problems does it solve?
• Be very specific (not "general assistant")

2. Concrete examples

• Show exactly how to invoke the agent
• Provide 2-3 usage examples
• Include expected outputs

3. Capability list

• Use checkmarks (✓) for readable list
• Group related capabilities
• Be specific, not vague

4. Context and timing

• When should user invoke this agent?
• What prerequisites are needed?
• What's the expected workflow?

5. Integration awareness

• List related commands
• Link to complementary agents
• Show how agent fits in larger workflow

Example: @code-quality Agent

File: agents/code-quality.md

---
name: code-quality
description: Code review focused on jam-time quality. Catches memory leaks, pattern inconsistencies, async issues, pooling problems without blocking workflow.
---

# @code-quality

Code review specialist focused on jam-time quality. Catches memory leaks, pattern inconsistencies, async issues, pooling problems, and architectural conflicts without blocking workflow.

## What This Agent Does

The code-quality agent is your fast-feedback reviewer. Invoke with `@code-quality` to:

- **Review code quality** - Spot issues without slowing development
- **Memory safety** - Catch leaks, pooling problems, cleanup issues
- **Pattern consistency** - Does it match your codebase patterns?
- **Modern practices** - Input System, async/await, DI, pooling done right?
- **Integration validation** - Will it work with existing systems?
- **Performance checks** - Inefficiencies, unnecessary allocations
- **Quick fixes** - Get suggestions to improve in-place

## Capabilities

✓ C# code review
✓ Memory safety analysis
✓ Pattern consistency checking
✓ Performance analysis
✓ Integration validation
✓ Best practice validation
✓ Quick fix suggestions

## Usage Examples

@code-quality: Review this enemy spawner - is it efficient during high-spawn moments? → Get: Performance analysis, pooling assessment, suggestions

@code-quality: Check this UI code for memory leaks → Get: Memory analysis, Manager connection validation, quick fixes

## When to Invoke

- **During development** - Get quick feedback without formal review
- **Before merging** - Validate quality before integration
- **Complex code** - Need expert validation on tricky logic
- **Performance concern** - "Is this efficient?"

## Works With

- `/review-code` command - Quick status checks
- `@feature-developer` agent - Reviews generated code
- Your team - Fast feedback during crunch

Agent Documentation Guidelines

• Title - # @agent-name (matches agent name)
• Role Statement - Describe the agent's role/specialty
• What This Agent Does - List key capabilities
• Capabilities - Checkmark list of specific abilities
• Usage Examples - 2-3 concrete examples with expected outputs
• When to Invoke - Situations where agent is useful
• Works With - Related commands, agents, tools

---

Part 3: Updating plugin.json

plugin.json Structure for Commands

{
  "name": "your-bundle",
  "version": "1.0.0",
  "skills": ["./skills/skill-name"],
  "commands": [
    "./commands/review-starters.md",
    "./commands/design-feature.md",
    "./commands/generate-feature.md"
  ]
}

Key Points:

• commands is an array of file paths
• Each path points to a .md file in the commands/ directory
• File names can be different from command names (use kebab-case for both)

plugin.json Structure for Agents

{
  "name": "your-bundle",
  "version": "1.0.0",
  "skills": ["./skills/skill-name"],
  "agents": [
    "./agents/code-quality.md",
    "./agents/feature-developer.md",
    "./agents/jam-architect.md"
  ]
}

Key Points:

• agents is an array of file paths
• Each path points to a .md file in the agents/ directory
• File names should match agent names (use kebab-case)

plugin.json with Both Commands and Agents

{
  "name": "game-jam-toolkit-bundle",
  "version": "1.0.0",
  "description": "Complete game jam development partner",
  "author": {
    "name": "Your Name",
    "email": "your@email.com"
  },
  "skills": [
    "./skills/validate-starters",
    "./skills/feature-architect"
  ],
  "commands": [
    "./commands/review-starters.md",
    "./commands/design-feature.md",
    "./commands/generate-feature.md"
  ],
  "agents": [
    "./agents/jam-architect.md",
    "./agents/feature-developer.md",
    "./agents/code-quality.md"
  ]
}

---

Common Mistakes & Fixes

Mistake 1: Inline JSON Objects

❌ WRONG - Inline agent definition in plugin.json:

{
  "agents": [
    {
      "name": "my-agent",
      "description": "Does something"
    }
  ]
}

✅ CORRECT - Reference markdown file:

{
  "agents": [
    "./agents/my-agent.md"
  ]
}

Mistake 2: Directory References Instead of Files

❌ WRONG - Reference to directory:

{
  "commands": "./commands",
  "agents": "./agents"
}

✅ CORRECT - Reference individual files:

{
  "commands": [
    "./commands/command1.md",
    "./commands/command2.md"
  ],
  "agents": [
    "./agents/agent1.md",
    "./agents/agent2.md"
  ]
}

Mistake 3: Missing YAML Frontmatter

❌ WRONG - No frontmatter:

# /my-command

This is my command...

✅ CORRECT - YAML frontmatter required:

---
name: my-command
description: This is my command's purpose
---

# /my-command

This is my command...

Mistake 4: Wrong File Extensions

❌ WRONG - Wrong extensions:

commands/
├── review.json    # Should be .md
└── validate.txt   # Should be .md

✅ CORRECT - All .md files:

commands/
├── review.md
└── validate.md

Mistake 5: Incorrect File Paths in plugin.json

❌ WRONG - Wrong paths:

{
  "commands": ["review-starters", "design-feature"]
}

✅ CORRECT - Full relative paths:

{
  "commands": [
    "./commands/review-starters.md",
    "./commands/design-feature.md"
  ]
}

---

Directory Structure

Typical Bundle with Commands and Agents

bundle-name/
├── .claude-plugin/
│   └── plugin.json                    # References all components
├── README.md
├── skills/
│   ├── skill-1/
│   │   └── SKILL.md
│   └── skill-2/
│       └── SKILL.md
├── commands/                          # CLI-like utilities
│   ├── review-starters.md            # YAML frontmatter + docs
│   ├── design-feature.md
│   └── generate-feature.md
└── agents/                            # Autonomous assistants
    ├── code-quality.md               # YAML frontmatter + instructions
    ├── feature-developer.md
    └── jam-architect.md

---

Validation Checklist

Before Committing Commands/Agents

• All .md files are in correct directories (commands/ or agents/)
• All files have .md extension
• All files have YAML frontmatter with name and description
• name field is kebab-case and matches file name
• description field is present and descriptive
• No JSON objects inside plugin.json for agents/commands
• plugin.json has arrays for commands and agents
• All paths in arrays are relative paths with ./ prefix
• All paths point to actual files that exist
• No typos in file paths
• Markdown content is well-formatted and clear

Testing After Creation

# Validate bundle structure
bundles/dev-toolkit-bundle/skills/bundle-maker/scripts/validate-bundle.sh <bundle-path>

# Check that plugin.json is valid JSON
cat <bundle-path>/.claude-plugin/plugin.json | jq .

# Verify all referenced files exist
for file in $(jq -r '.commands[]? // .agents[]? | ltrimstr("./")' <bundle-path>/.claude-plugin/plugin.json); do
  if [ ! -f "<bundle-path>/$file" ]; then
    echo "Missing file: $file"
  fi
done

---

When This Skill Applies

Use this skill when:

• Creating new commands for a Claude Code plugin bundle
• Creating new agents for a Claude Code plugin bundle
• Fixing command/agent configuration errors in existing bundles
• Learning proper YAML frontmatter format
• Understanding how to structure agent/command markdown
• Updating plugin.json to reference new commands/agents
• Troubleshooting "Invalid manifest" or similar errors

Complements the bundle-maker skill for complete plugin development workflow.