Agent Skill
2/7/2026

meta-agent

Guide for creating and validating specialized AI agents (subagents). Use when users want to create a new agent, make a subagent, define agent configurations, validate agents, or need guidance on agent architecture. Do NOT use for creating skills (use meta-skill) or commands (use meta-command).

P
practical
2GitHub Stars
1Views
npx skills add practical-stack/ai-lab

SKILL.md

Namemeta-agent
DescriptionGuide for creating and validating specialized AI agents (subagents). Use when users want to create a new agent, make a subagent, define agent configurations, validate agents, or need guidance on agent architecture. Do NOT use for creating skills (use meta-skill) or commands (use meta-command).

name: meta-agent description: Guide for creating and validating specialized AI agents (subagents). Use when users want to create a new agent, make a subagent, define agent configurations, validate agents, or need guidance on agent architecture. Do NOT use for creating skills (use meta-skill) or commands (use meta-command).

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.

Workflow Routing

IntentWorkflow
Create a new agent (full process)workflows/create-agent.md
Validate existing agent(s)workflows/validate-agent.md

Core Resources

ResourcePurpose
Agent PatternsArchitecture guidance, interfaces
Agent TemplatesReady-to-use agent examples

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:

CategoryPurposeModel TierExamples
explorationFast search, codebase discoveryLOW (haiku/fast)explore, searcher
specialistDomain-specific implementationMEDIUM (sonnet)frontend, backend
advisorRead-only consultationHIGH (opus)oracle, architect
utilityGeneral helpersLOW (haiku/fast)writer, formatter
orchestrationMulti-agent coordinationMEDIUM (sonnet)orchestrator

Agent Creation Workflow (5 Phases)

For the complete step-by-step guide, see workflows/create-agent.md.

PhaseGoalKey Output
1. DEFINEClarify purposeRole definition
2. CLASSIFYSelect category/modelCategory + tier
3. DESIGNWrite promptSystem prompt
4. CONFIGURECreate configAgent file
5. REGISTERAdd & testWorking agent

Platform Configurations

OpenCode / Claude Code Agent Fields

FieldRequiredDescription
nameYesAgent identifier (kebab-case)
descriptionYesSelection guide for orchestrator
modeYesAlways "subagent"
modelYesModel string
temperatureNoDefaults to 0.1
toolsNoTool whitelist/blacklist
promptYesSystem prompt

Example Configuration

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...`
  };
}

Anti-Patterns to Avoid

Anti-PatternProblemSolution
Vague descriptionOrchestrator can't decideInclude specific triggers
Too many toolsSecurity riskMinimal tool set
Long promptsSlower, harder to maintainKeep under 500 words
Generic "helper"No clear purposeSingle responsibility
Missing readonlySecurity for advisorsSet readonly: true
Wrong model tierCost or quality issuesMatch tier to complexity
Skills Info
Original Name:meta-agentAuthor:practical