Create New Agent

Quickly scaffold a new specialized subagent when current agents don't fit the task.

When to Use

• Need specialist for task not covered by existing 6 agents

• Task requires deep domain expertise

• Creating configuration/meta work agent

• Temporary specialist for one-time complex task

Current Agents (Don't Recreate)

• api-specialist (🟡) - Elysia + TypeBox

• database-specialist (🔵) - Drizzle ORM + PostgreSQL

• frontend-specialist (🟢) - React + Vite + Three.js

• testing-specialist (🔴) - Bun Test + Playwright

• security-specialist (🟠) - Privy Auth + Security

• config-specialist (⚙️) - .claude configuration

Agent Template

name: [agent-name] description: [EMOJI] [CAPS NAME] - [one-line description]. Use [PROACTIVELY/when] for [task types]. [What it handles]. tools: [Read, Write, Edit, Bash, Grep, Glob] model: sonnet

[EMOJI] [Agent Name]

[Detailed description of agent's purpose]

Research-First Protocol ⚠️

CRITICAL: Writing code is your LAST priority

Workflow Order (NEVER skip steps):

1. RESEARCH - Use deepwiki for ANY external libraries/frameworks (Claude's knowledge is outdated)
2. GATHER CONTEXT - Read existing files, Grep patterns, Glob to find code
3. REUSE - Triple check if existing code already does this
4. VERIFY - Ask user for clarification on ANY assumptions
5. SIMPLIFY - Keep it simple, never over-engineer
6. CODE - Only write new code after exhausting steps 1-5

Before Writing ANY Code:

• ✅ Used deepwiki to research latest API/library patterns?
• ✅ Read all relevant existing files?
• ✅ Searched codebase for similar functionality?
• ✅ Asked user to verify approach?
• ✅ Confirmed simplest possible solution?
• ❌ If ANY answer is NO, DO NOT write code yet

Key Principles:

• Reuse > Create - Always prefer editing existing files over creating new ones
• Simple > Complex - Avoid over-engineering
• Ask > Assume - When uncertain, ask the user
• Research > Memory - Use deepwiki, don't trust outdated knowledge

Responsibilities

1. [Primary Responsibility]

   • [Specific task]
   • [Specific task]

2. [Secondary Responsibility]

   • [Specific task]
   • [Specific task]

Domain Expertise

• [Technology/framework 1]
• [Technology/framework 2]
• [Pattern/concept 1]

When to Use This Agent

• [Trigger scenario 1]
• [Trigger scenario 2]
• [Trigger scenario 3]

Workflow

1. [Step 1]
2. [Step 2]
3. [Step 3]

Example Agents to Create

Performance Specialist (⚡)

• Bundle optimization, lazy loading

• React.memo, useMemo, useCallback

• Three.js optimization, LOD, instancing

• Lighthouse scores, Core Web Vitals

Documentation Specialist (📚)

• README updates, API docs

• Code comments, JSDoc

• Architecture diagrams

• Changelog maintenance

Migration Specialist (🔄)

• Express → Elysia migrations

• Database schema changes

• Dependency updates

• Breaking change handling

Creating the Agent

• Choose emoji icon

• Define clear domain boundaries

• List specific technologies/frameworks

• Include Research-First Protocol

• Specify tools needed

• Save to .claude/agents/[name].md

After Creation

The agent is immediately available:

I need to [task] → Use [agent-name] agent

No restart needed - agents are loaded on demand.

Best Practices

• Specific domain - Don't make generic "helper" agents

• Clear triggers - When should this agent be used?

• Tool selection - Only tools needed for the domain

• Temporary OK - Can delete agent after one-time use

• Evolve existing - Update existing agents rather than create new ones when possible