meta-agent-creator

name	meta-agent-creator
description	Guide for creating specialized AI agents (subagents). This skill should be used when users want to create a new agent, define agent configurations, or need guidance on agent architecture.

Meta-Agent Creator

A meta-skill that helps create new specialized agents for multi-agent orchestration systems. Supports Claude Code and OpenCode platforms.

Output Language: All generated agent content (prompts, descriptions) will be in English.

Quick Start

Create a New Agent

# Initialize a new agent
bun scripts/init-agent.ts <agent-name> --path <output-directory> --platform <platform>

# Examples
bun scripts/init-agent.ts security-auditor --path src/agents --platform opencode
bun scripts/init-agent.ts data-analyst --path .claude/agents --platform claude-code

Validate Agent

bun scripts/validate-agent.ts <agent-file>

# Examples
bun scripts/validate-agent.ts src/agents/my-agent.ts
bun scripts/validate-agent.ts .claude/agents/my-agent.ts

Agent Categories

Choose the appropriate category based on agent purpose:

Category	Purpose	Model Tier	Examples
exploration	Fast search, codebase discovery	LOW (haiku/fast)	explore, searcher
specialist	Domain-specific implementation	MEDIUM (sonnet)	frontend, backend
advisor	Read-only consultation	HIGH (opus)	oracle, architect
utility	General helpers	LOW (haiku/fast)	writer, formatter
orchestration	Multi-agent coordination	MEDIUM (sonnet)	orchestrator

Agent Creation Workflow (5 Phases)

Phase 1: DEFINE PURPOSE

Goal: Clarify agent's role and responsibilities.

Questions to Ask:

1. "What specific tasks should this agent handle?"
2. "When should the orchestrator delegate to this agent?"
3. "What domain expertise is required?"
4. "Should it be read-only or have write access?"

Deliverables:

• Clear role definition
• Trigger conditions for delegation
• Required capabilities list
• Access level decision (readonly vs full)

Phase 2: CLASSIFY

Goal: Determine category and model tier.

Decision Matrix:

If the agent needs...	Category	Model	Cost
Fast codebase search	exploration	haiku/fast	FREE
Specific domain implementation	specialist	sonnet/inherit	CHEAP
Complex reasoning, architecture	advisor	opus/inherit	EXPENSIVE
Simple transformations	utility	haiku/fast	FREE
Delegate to other agents	orchestration	sonnet/inherit	CHEAP

Deliverables:

• Category assignment
• Model tier selection
• Cost classification

Phase 3: DESIGN PROMPT

Goal: Write effective system prompt.

Prompt Structure:

## Role

[1-2 sentence identity and expertise]

## Core Capabilities

- [Capability 1]
- [Capability 2]

## Workflow

1. [Step 1]
2. [Step 2]

## Output Format

[Structured output definition]

## Constraints

- [Constraint 1]
- [Constraint 2]

Best Practices:

• Keep prompts concise (<500 words)
• Use imperative form
• Include specific output format
• Define clear constraints

Phase 4: CONFIGURE

Goal: Create platform-specific configuration.

OpenCode / Claude Code (TypeScript)

Location: src/agents/<agent-name>.ts

export function createMyAgent(model: string): AgentConfig {
  return {
    name: "my-agent",
    description: "Agent description...",
    mode: "subagent",
    model,
    temperature: 0.1,
    tools: { include: ["read", "glob", "grep"] },
    prompt: `System prompt...`,
  };
}

Phase 5: REGISTER & TEST

Goal: Add agent to registry and verify functionality.

Registration Checklist:

• Add to agent definitions/index
• Update orchestrator's available agents list
• Add to delegation table if applicable
• Test with sample delegation

Testing:

1. Invoke agent with representative task
2. Verify output format matches spec
3. Check tool restrictions are enforced
4. Test edge cases

Platform Configurations

OpenCode / Claude Code Agent Fields

Field	Required	Description
name	Yes	Agent identifier
description	Yes	Selection guide
mode	Yes	Always "subagent"
model	Yes	Model string
temperature	No	Defaults to 0.1
tools	No	Tool whitelist/blacklist
prompt	Yes	System prompt

Common Agent Templates

See references/agent-templates.md for ready-to-use templates:

• Exploration Agent
• Verification Agent
• Debugger Agent
• Security Auditor Agent

Anti-Patterns to Avoid

Anti-Pattern	Problem	Solution
Vague description	Orchestrator can't decide	Include specific triggers
Too many tools	Security risk	Minimal tool set
Long prompts	Slower, harder to maintain	Keep under 500 words
Generic "helper"	No clear purpose	Single responsibility
Missing readonly	Security for advisors	Set readonly: true
Wrong model tier	Cost or quality issues	Match tier to complexity

References

• Agent Templates: See references/agent-templates.md for ready-to-use templates
• Agent Patterns: See ../meta-skill-creator/references/agent-patterns.md for architecture guidance