Agent Creator

Expert guidance for creating effective Claude Code agents following best practices.

About Agents

Agents are specialized subprocesses that handle complex tasks autonomously with isolated context windows. They extend Claude's capabilities by providing focused expertise, specific tool access, and appropriate model selection for different task types.

Agents vs Skills

Aspect	Agents	Skills
Context	Isolated context window	Shared with main conversation
Purpose	Autonomous task execution	Procedural knowledge & guidance
Model	Can specify different model	Uses current model
Tools	Can restrict/expand tool access	Uses available tools
Use when	Task needs isolation, different capabilities, or parallel execution	Task benefits from guidance staying in current context

Choose agents when:

• Task benefits from isolated context (clean slate)
• Different model is optimal (haiku for speed, opus for reasoning)
• Tool access should be restricted for security
• Tasks can run in parallel

Choose skills when:

• Knowledge should stay in current conversation
• Guidance benefits from seeing full context
• No special tool/model requirements

Agent Architecture

File Structure

Agents are Markdown files with YAML frontmatter stored in:

• Project agents: .claude/agents/ (highest priority)
• User agents: ~/.claude/agents/ (all projects)

Configuration Format

---
name: agent-identifier          # Required: lowercase, hyphens only
description: When and why...    # Required: PRIMARY TRIGGER MECHANISM
tools: Tool1, Tool2, Tool3      # Optional: restrict/specify tools
model: sonnet                   # Optional: sonnet, opus, haiku
skills: skill1, skill2          # Optional: auto-load skills
color: blue                     # Optional: UI indicator
---

[Agent system prompt here - Markdown instructions]

Configuration Fields

Field	Required	Details
name	Yes	Lowercase letters and hyphens only (e.g., code-reviewer)
description	Yes	PRIMARY TRIGGERING MECHANISM - determines when agent is invoked
tools	No	Comma-separated list; omit to inherit all tools
model	No	haiku (fast), sonnet (balanced), opus (reasoning)
skills	No	Skills to auto-load when agent invokes
color	No	UI color indicator

Core Principles

1. Single-Purpose Focus

Create focused, single-responsibility agents:

• One clear job per agent
• Better composability and predictability
• Easier to debug and maintain

2. Description is Everything

The description field is the PRIMARY TRIGGERING MECHANISM. Include:

• WHAT the agent does
• WHEN to use it (trigger scenarios)
• Specific examples in angle brackets

Pattern:

description: [Role/capability]. MUST BE USED for [specific scenarios]. [Additional context]. Examples: "[trigger phrase 1]", "[trigger phrase 2]"

3. Minimal Viable Toolset

Grant only necessary tools:

• Read-only agents: Glob, Grep, Read, LS
• Analysis agents: Add WebFetch, WebSearch
• Planning agents: Add Write, Edit, TodoWrite
• Implementation agents: Add Bash, Task

Performance impact:

• 3-5 tools: ~2-5k tokens
• 7-8 tools: ~7-10k tokens
• 15+ tools: 15-25k tokens (avoid)

4. Appropriate Model Selection

• Haiku: Fast, lightweight tasks (exploration, simple analysis)
• Sonnet: Balanced tasks (orchestration, validation, general work)
• Opus: Complex reasoning (planning, architecture, research)

Agent Creation Process

Step 1: Define Purpose

Before creating, answer:

1. What specific problem does this agent solve?
2. When should Claude automatically invoke this agent?
3. What makes this different from existing agents?
4. Does this need isolation, or would a skill suffice?

Step 2: Design Description

Write a comprehensive description:

1. Start with role/capability statement
2. Add "MUST BE USED" or "use PROACTIVELY" for strong triggers
3. Include specific scenarios and trigger phrases
4. Add examples using XML tags for complex scenarios

Step 3: Select Tools

Determine minimum tools needed:

1. List all operations the agent must perform
2. Map operations to specific tools
3. Remove any tools not strictly necessary
4. Consider security implications

Step 4: Choose Model

Select based on task complexity:

• Simple/fast tasks → haiku
• Standard tasks → sonnet or omit (default)
• Complex reasoning → opus

Step 5: Write System Prompt

Structure the prompt body:

# [Role Title]

[One-sentence mission statement]

## Core Principles
- [Principle 1]: [Brief explanation]
- [Principle 2]: [Brief explanation]

## Responsibilities / Methodology
1. [Step/responsibility 1]
2. [Step/responsibility 2]

## [Workflow/Process] (if multi-step)
### Step 1: [Name]
[Instructions]

## Quality Standards / Output Format
[Checklists, criteria, format requirements]

## Constraints (if critical)
- [What agent MUST NOT do]
- [Boundaries and limitations]

Step 6: Define Boundaries

Add explicit boundaries if agent has limited scope:

## YOUR ROLE ENDS HERE

**CRITICAL BOUNDARY**: You are strictly a [role]. Once you have [completed work]:
- Your work is COMPLETE
- DO NOT [prohibited action 1]
- DO NOT [prohibited action 2]

Prompt Engineering Best Practices

Use Imperative Form

• "Identify issues" not "You should identify"
• "Generate report" not "The report should be generated"

Strong Emphasis Markers

• MUST, NEVER, ALWAYS, CRITICAL for non-negotiables
• Bold or capitalize critical constraints
• Use strategically, not everywhere

Include Examples

• Good vs Bad examples (Use checkmarks/X marks)
• Concrete user scenarios
• Before/after for transformations

Be Specific, Not Vague

• "Test min/max values, zero, negative" not "test edge cases"
• "Check for SQL injection, XSS, path traversal" not "check security"
• Include decision criteria and thresholds

Common Pitfalls to Avoid

Over-Engineering

• Creating 15+ tool agents with 30k token prompts
• One mega-agent instead of composable focused agents
• Adding features "just in case"

Solution: Start lightweight (<3k tokens, 3-5 tools), expand based on need

Vague Descriptions

• "Helps with coding" - too generic
• Missing "when to use" information
• No concrete trigger examples

Solution: Specific triggers, scenarios, and examples

Tool Overload

• Giving all tools "just in case"
• Read-only agents with Write/Bash access
• Security risk and performance hit

Solution: Minimal viable toolset per agent purpose

Missing Boundaries

• Agent scope creeps during execution
• No clear "done" criteria
• Agent attempts tasks outside its role

Solution: Explicit "YOUR ROLE ENDS HERE" section for bounded agents

Ignoring Token Budget

• Verbose explanations Claude already knows
• Repeated information
• Unnecessary examples

Solution: Challenge each paragraph - "Does this justify its token cost?"

Integration with Skills

When an agent needs domain expertise, use skills:

---
name: specialized-agent
skills: domain-skill, helper-skill
---

When to create an accompanying skill:

• Agent needs substantial domain knowledge
• Knowledge is reusable across multiple agents
• Information would bloat agent prompt

Use the skill-creator skill to create accompanying skills:

Invoke skill: skill-creator

Quality Checklist

Before finalizing an agent:

• Name: Lowercase with hyphens, descriptive
• Description: Includes WHAT, WHEN, and examples
• Tools: Minimal viable set for the task
• Model: Appropriate for task complexity
• Prompt: Follows structure (role → principles → methodology → standards)
• Boundaries: Clear scope and "done" criteria (if bounded)
• Token efficiency: <5k words for most agents
• Examples: Concrete scenarios in description or prompt
• No overlap: Doesn't duplicate existing agent functionality

Example: Well-Designed Agent

---
name: code-reviewer
description: Expert code review specialist. MUST BE USED when reviewing pull requests, code changes, or architecture decisions. Focuses on security, performance, maintainability, and best practices. Examples: "Review this PR", "Check this implementation for issues", "Review the changes in feature branch"
tools: Glob, Grep, Read, LS, WebSearch
model: sonnet
color: yellow
---

# Code Review Specialist

You are an expert code reviewer focused on identifying issues that impact security, performance, and maintainability.

## Review Priorities

1. **Security**: Injection vulnerabilities, auth issues, data exposure
2. **Performance**: O(n²) operations, memory leaks, unnecessary computation
3. **Maintainability**: Code clarity, naming, separation of concerns
4. **Best Practices**: Framework conventions, error handling, testing

## Review Process

1. Understand the change context (what problem is being solved)
2. Review for security vulnerabilities first
3. Check performance implications
4. Assess code quality and maintainability
5. Verify test coverage

## Output Format

For each issue found:
- **Severity**: Critical/High/Medium/Low
- **Location**: file:line
- **Issue**: Clear description
- **Recommendation**: Specific fix suggestion

## YOUR ROLE ENDS HERE

You are strictly a reviewer. After completing review:
- Your work is COMPLETE
- DO NOT implement fixes
- DO NOT modify code

Resources

• See references/prompt-patterns.md for advanced prompt engineering patterns
• Use skill-creator skill when agent needs accompanying skills