Specialized AI subagents for Claude Code - now 50% more concise, 100% more effective.
Tell Claude to use a specific agent:
"Use the backend-dev subagent to create a JWT auth system"
- senior-architect - System design, technical decisions
- backend-dev - APIs, databases, services
- frontend-dev - UI components, state, performance
- devops-eng - CI/CD, infrastructure, deployment
- qa-eng - Testing strategies, automation
- security-eng - Audits, vulnerabilities, compliance
- debugger - Bug hunting, root cause analysis
- error-analyst - Pattern detection, prevention
- codebase-maintainer - Refactoring, tech debt
- pm-agent - Planning, coordination, delivery
- prod-strategist - Market strategy, positioning
- code-roaster - Code reviews, feedback
Each agent now follows the optimized template:
- Front-loaded critical info
- Visual anchors with emojis
- Compressed to ~250 words (50% reduction)
- Explicit input→process→output flows
- Action-oriented with clear deliverables
Requirements → PM → Architect → Backend/Frontend
↓ ↓
QA/Security ← DevOps ← Debugger
The constraint is the feature. What helps ADHD brains navigate limited working memory helps LLMs navigate limited context windows:
- Less tokens = Cheaper API calls
- Clear structure = Consistent outputs
- Visual markers = Easy scanning
- No ambiguity = Fewer errors
- Explicit state = Better context retention
# Complex task with multiple agents
"Use senior-architect to design OAuth2 flow,
then backend-dev to implement endpoints,
then qa-eng to create test suite"
# Quick fix
"debugger: find why login returns 500"
# Code quality
"code-roaster: review this PR brutally"- Token reduction: 40-60% per agent
- Response time: ~30% faster
- Output consistency: 95%+ adherence
- Context retention: Improved across long sessions
- ❌ Overloading single agent with multiple responsibilities
- ❌ Vague requests without clear outputs
- ❌ Skipping agent specification
- ✅ Use specific agent + clear task + expected output
- One agent, one task - Don't multitask agents
- Chain for complexity - Use multiple agents sequentially
- Explicit outputs - Tell agent what format you need
- Validate assumptions - Agents state their constraints
Built with the ADHD Prompting Framework - because the best interfaces acknowledge cognitive constraints rather than assuming unlimited processing capacity.