Agent Creation Skill

Quick Reference: Agent vs Skill vs Command

Type	Use When	Context	Invocation
Subagent	Task needs isolation, custom tools, or separate context	Own window	Auto-matched or explicit
Skill	Reusable knowledge/patterns across tasks	Shared	Auto-discovered
Slash Command	Quick, repeatable workflow	Shared	Manual /command

5-Step Agent Creation Workflow

Step 1: Define Single Responsibility

Ask: "What ONE thing should this agent do exceptionally well?"

Bad examples:

• "Handle all code tasks" (too broad)
• "Review code and deploy and write docs" (multiple responsibilities)

Good examples:

• "Review TypeScript for security vulnerabilities"
• "Generate database migration scripts"
• "Optimize React component performance"

Step 2: Identify Minimal Tools

Only include what's necessary:

Tool	Purpose
Read	View files
Edit	Modify existing files
Write	Create new files
Bash	Run commands
Glob	Find files by pattern
Grep	Search file contents
Task	Delegate to other agents
WebSearch	Search the web
WebFetch	Fetch web content

Rule: If unsure whether a tool is needed, leave it out. Add later if needed.

Step 3: Choose Model

Model	Use When	Cost	Speed
haiku	Fast, simple tasks (exploration, search)	$	Fast
sonnet	Balanced tasks (most use cases)	$$	Medium
opus	Complex reasoning, critical decisions	$$$	Slow
inherit	Use parent's model	-	-

Step 4: Write System Prompt

Structure:

# Agent Name

You are specialized in [domain]. Your responsibilities:

1. [Primary responsibility]
2. [Secondary responsibility]

## How You Work

[Step-by-step approach]

## Examples

### Example 1: [Scenario]
Input: [What you receive]
Output: [What you produce]

### Example 2: [Edge case]
...

## What You DON'T Do

- [Anti-pattern 1]
- [Anti-pattern 2]

Step 5: Create & Test

# Create agent file
cat > .claude/agents/my-agent.md << 'EOF'
---
name: my-agent
description: [Clear, matchable description]
tools: Read, Edit
model: sonnet
---

[System prompt here]
EOF

# Test by asking Claude to use it
> "Use my-agent to [task]"

Agent File Template

---
name: agent-name-kebab-case
description: |
  Clear description of when to use this agent.
  Include keywords Claude will match against user requests.
  Add "Use PROACTIVELY" if it should auto-trigger.
tools: Read, Edit, Bash
model: sonnet
skills: skill1, skill2
permissionMode: default
---

# Agent Name

You are a specialized agent for [purpose].

## Responsibilities

1. **[Responsibility 1]**: [Description]
2. **[Responsibility 2]**: [Description]

## Workflow

When given a task:
1. [First step]
2. [Second step]
3. [Third step]

## Examples

### Example 1: [Common scenario]

**Input**: [Sample input]

**Output**: [Expected output with reasoning]

### Example 2: [Edge case]

**Input**: [Tricky input]

**Output**: [How to handle it]

## Constraints

- [Limitation 1]
- [Limitation 2]

## What NOT to Do

- Don't [anti-pattern]
- Avoid [mistake]

Skill Creation Template

For reusable knowledge (not isolated tasks):

# .claude/skills/my-skill/SKILL.md

---
name: my-skill
description: |
  Knowledge domain description.
  Keywords for auto-discovery.
allowed-tools: Read, Write
---

# Skill Name

## When This Applies

Use this skill when:
- [Condition 1]
- [Condition 2]

## Core Concepts

### Concept 1
[Explanation]

### Concept 2
[Explanation]

## Patterns

### Pattern 1: [Name]
```code
[Example]

Pattern 2: [Name]

[Example]

Anti-Patterns

• Don't: [Bad practice]
• Instead: [Good practice]

References

See reference.md for detailed specifications.

## Slash Command Template

For quick, repeatable workflows:

```markdown
# .claude/commands/my-command.md

---
description: Brief description shown in /help
---

# My Command

[Full prompt that runs when user types /my-command]

Include:
- Context about what this does
- Specific instructions
- Expected workflow

Common Agent Patterns

Code Review Agent

---
name: code-reviewer
description: Reviews code changes for quality, security, and best practices. Use after making code changes.
tools: Read, Grep, Glob, Bash
model: sonnet
---

# Code Reviewer

Focus on:
- Security vulnerabilities (OWASP Top 10)
- Performance concerns
- Code clarity and maintainability
- Test coverage

Provide feedback as:
- CRITICAL: Must fix before merge
- WARNING: Should address soon
- SUGGESTION: Nice to have

Don't nitpick style (leave that to linters).

Debug Agent

---
name: debugger
description: Diagnoses and fixes errors, test failures, and unexpected behavior. Use PROACTIVELY when errors occur.
tools: Read, Grep, Glob, Bash, Edit
model: sonnet
skills: systematic-debugging
---

# Debugger

## Systematic Approach

1. **Reproduce**: Understand the exact failure
2. **Isolate**: Find the root cause location
3. **Understand**: Why is it failing?
4. **Fix**: Minimal change to resolve
5. **Verify**: Confirm fix works
6. **Prevent**: Add tests if missing

Documentation Agent

---
name: docs-writer
description: Creates and updates documentation. Use when code changes need docs updates.
tools: Read, Write, Edit, Glob
model: haiku
---

# Documentation Writer

Update docs when:
- New features added
- APIs changed
- Important decisions made

Follow existing doc style in the project.
Keep docs close to code (prefer inline JSDoc over separate files).

Testing Your Agent

1. Direct invocation: "Use [agent-name] to [task]"
2. Auto-discovery: Just describe the task, see if Claude picks the right agent
3. Edge cases: Try unusual inputs
4. Tool verification: Confirm it only uses granted tools

Iteration Tips

• Start minimal, add complexity as needed
• Watch for tool permission errors (add missing tools)
• Refine description if auto-discovery fails
• Add examples for common mistakes