Here's a dump of my current Claude Code agents setup. I primarly use /start-task for large features. It helps to write your prompt to this in a different window.
Make it long (like a feature spec or PRD). Only the first agent will have it in it's context, so you can be verbose.
The key here is that we first open a context-manager agent that manages the task context across all agents (via a file).
The coordinator agent then executes tasks using a specialized agent for each task. Each agent has their own context window, this is the key benefit of this flow.
/start-task "Your task description here"
general-purpose - Research, search, multi-step tasks
frontend-pro - React components, UI, frontend issues
debugger - Fix bugs, errors, test failures
performance-engineer - Optimize bottlenecks, profiling
docs-maintainer - Create/update documentation
codebase-janitor - Clean up orphaned files, organize code
code-reviewer - Review code quality and security
architect-reviewer - Review architectural consistency
forensic-code-analyst - Analyze existing code (read-only)
context-manager - Manage context for multi-agent workflows
javascript-pro - Modern JS, async patterns, Node.js
terraform-specialist - Terraform modules, IaC best practices
Simple Solution Subagent - Quick, direct solutions
/start-task "Find JWT authentication implementation"
/start-task "Create responsive navbar with mobile menu"
/start-task "Fix failing user registration tests"
/start-task "Clean up orphaned files from auth feature"
/start-task "Optimize dashboard loading performance"
You can also run specific agents directly without using /start-task:
Use the @agent-debugger to find out why my tests are failing
You can even define the tasks more deterministically:
Task("debug auth", "Fix the authentication error in login.js", "debugger")
Task("review PR", "Review the changes in pull request #123", "code-reviewer")
Task("analyze performance", "Find performance bottlenecks in the dashboard", "performance-engineer")