Subagent Refactoring Methodology

Analysis Criteria

Structural Issues

• Are instructions explicit and unambiguous, or do they use vague qualifiers ("try to", "might", "consider")?
• Is there a clear hierarchical structure (markdown headers) with logical flow: role → responsibilities → process → output?
• Are concerns properly separated (instructions vs examples vs data)?
• Missing sections: role definition, process steps, output format, boundaries?

Model Optimization

• Sonnet (default): cost-optimized, parallel tool execution, effort calibration explicit
• Opus (upgrade only with observed complexity evidence): complex coding, multi-step agents, computer use

Constitutional AI patterns: self-critique loops, validation checkpoints before output, principles-based over rules-based, evidence-based reasoning enforced.

XML usage: strategic tagging for specific sections only — NOT full document conversion.