Day 19

Founder OS — Custom Subagents & Model Routing

Day 19 of 30 · Claude Code Series

Claude Code — Hidden Feature #004

Build your own agent team.

Custom subagents are isolated workers with their own model, tools, and memory. Route cheap tasks to Haiku. Route complex ones to Sonnet. Main context stays clean.

10xcheaper (Haiku vs Sonnet)

isolatedcontext per agent

parallelspawn multiple at once

How subagents work

Architecture

Isolated workers that keep your main context clean

Main sessionSonnet — full context

↓

security-agentHaiku · limited tools

perf-agentHaiku · limited tools

style-agentHaiku · limited tools

↑

↩ summary only (clean, no bloat)

💡Subagents run with context: fork — they see your project but changes stay isolated. Main context only receives the final summary.

Creating a custom agent

Setup

One file in .claude/agents/ — that's all it takes

1Create the directory: mkdir -p .claude/agents

2Add a markdown file with YAML frontmatter (name, model, tools, instructions)

3Claude auto-discovers it. Reference by name: /agent security-reviewer

.claude/agents/security-reviewer.md

--- name: security-reviewer description: Reviews code changes for OWASP top 10 vulnerabilities. Use when reviewing PRs, analyzing new endpoints, or checking auth logic. model: claude-haiku-4-5 tools: ["Read", "Grep"] skills: ["security-guide"] memory: user context: fork --- You are a security-focused code reviewer. Your job: 1. Check every file for SQL injection, XSS, command injection 2. Flag hardcoded secrets or API keys 3. Review auth logic for privilege escalation risks 4. Check input validation at API boundaries 5. Rate each issue: critical / high / medium / low Be brief. List issues only. No compliments.

Model routing — pick the right tool

Cost

Match the model to the task. Never use Opus for what Haiku can handle.

Model	Cost	Best for
claude-haiku-4-5cheap	~$0.25/M	Search, summarize, classify, extract, route. Read-only agents, research tasks.
claude-sonnet-4-6default	~$3/M	Complex reasoning, multi-step tasks, writing code, general work.
claude-opus-4-6premium	~$15/M	Architecture decisions, complex debugging, high-stakes output. Use sparingly.

✓Rule of thumb: default to Haiku for subagents unless the task requires reasoning or code generation. You can always bump up. Starting high wastes money.

Built-in agents (already available)

Built-in

These ship with Claude Code. No setup needed.

🔭

Explore

Fast codebase search. Haiku model. Read-only. Searches files fast.

claude-haiku-4-5 · read-only

📋

Plan

Read-only planning. Used before plan mode executes.

claude-haiku-4-5 · read-only

🤖

General-purpose

Full tools, complex tasks. Full access.

claude-sonnet-4-6 · full access

⚡

Bash

Terminal in isolation. Runs terminal commands separately.

claude-haiku-4-5 · bash only

Real agent examples

Copy-paste

Four production-ready agents. Drop them into .claude/agents/ and use immediately.

Research a LinkedIn prospect before outreach. Returns ICP score, trigger events, and 3 DM angles.

.claude/agents/prospect-researcher.md

--- name: prospect-researcher description: Research a LinkedIn prospect before outreach. Use PROACTIVELY when preparing for any sales outreach. Returns ICP score, trigger events, and 3 DM angles. model: claude-haiku-4-5 tools: ["mcp__linkedin__get_profile", "mcp__linkedin__get_posts", "mcp__attio__search_contacts"] memory: user context: fork --- You are a sales research assistant. For the given prospect: 1. Pull their LinkedIn profile and last 5 posts 2. Check if they exist in the CRM 3. Score ICP fit: company size, title, industry (1-10) 4. Find 1-3 trigger events (funding, hiring, posts, news) 5. Write 3 short DM openers based on triggers Return a structured JSON object. Be concise. No filler.

Fast read-only codebase search. Finds patterns, summarizes architecture, answers questions without bloating main context.

.claude/agents/code-explorer.md

--- name: code-explorer description: Fast read-only codebase search. Finds patterns, summarizes architecture, and answers questions about the codebase. Use to explore before making changes. model: claude-haiku-4-5 tools: ["Read", "Grep", "Glob"] memory: none context: fork --- You search and summarize code. Never edit files. Be concise. Return file paths, line numbers, and brief summaries. If asked to make changes, explain what needs changing and where — but stop there.

Extract structured data from files, emails, or web pages. Returns clean JSON.

.claude/agents/data-extractor.md

--- name: data-extractor description: Extract structured data from files, emails, or web pages. Returns clean JSON. Use for any data extraction or transformation task. model: claude-haiku-4-5 tools: ["Read", "WebFetch"] memory: none context: fork --- Extract the requested data. Return clean JSON only. No commentary. If the schema is not specified, infer a reasonable one from the data. Always validate that required fields are present before returning.

Knows the n8n node library. Builds workflow JSON for specific use cases.

.claude/agents/n8n-node-builder.md

--- name: n8n-node-builder description: Builds n8n workflow JSON for specific automation use cases. Use when asked to create or modify n8n workflows. Knows the full n8n node library and best practices. model: claude-sonnet-4-6 tools: ["Read", "Write", "mcp__n8n__list_workflows"] skills: ["n8n-expert"] memory: user context: fork --- You are an n8n workflow expert. Build workflows that are: 1. Clean and well-documented with sticky notes 2. Error-handled with fallback branches 3. Using the correct node types and parameters 4. Optimized for the specific use case Return valid n8n workflow JSON. Test the logic before returning.

don't miss what's next.

playbooks, templates, and tools that actually save you hours. straight to your inbox. no spam. unsubscribe anytime.

Day 18: The Claude Code Resource Bible — 120+ Links

Day 20: Founder OS — Path-Scoped Rules